Evidence ledger¶
2026-10-28: PR #135 fresh-clone recovery¶
Resumed research/gpu-repeat at 3d9196c0ca32b0bad1a9f248fb9ce14596893a79;
the initial worktree was clean. Retained decisions 0087/0088, models, evaluator,
classifier/default sources and existing backend identities. Only STATE, this
ledger and numeric recovery evidence
change. Numeric/source/log hashes were saved before prose.
HF CLI access through ignored .env recovers the existing LLM job's timeout
status; full worker output is missing because its bucket is deleted. No new
classifier job is present. Public release downloads match the externally
recorded manifest hash and all 988 member hashes. All nine program/witness pairs
reconstruct against the original cost-record hashes without retraining. Exact
source preparation uses scripts/prepare_classifier_gpu.py; local admission uses
scripts/run_classifier_gpu.py:inputs and retains 104 files under
runs/pr135-recovery/admitted-inputs. The stock arithmetic fixture freshly proves
and verifies; this is CPU evidence, not device acceptance.
uv run --locked pytest -q tests/test_classifier_gpu.py tests/test_llm_gpu.py:
45 pass. make check: 2,612 pass / 50 skip, 156 software conformance matches.
make package-check initially fails the uv version gate; rerunning the same
package script via uv tool run --from uv==0.9.18 uv run --locked python scripts/reproduce_package.py
passes. Failures remain in new local directories. No paid run, new GPU proof,
speedup or completion claim. Current numeric HF credit balance is still needed
before decision 0088 permits the paid campaign; token access reports prepaid and
canPay, neither of which establishes a balance. No credential values are saved
in this public evidence.
Entries distinguish local software results, statistical conclusions, literature reports and proposed work. Never replace a failed entry with a favorable rerun.
2026-10-20: bootstrap software checks and negative classifier result¶
Purpose: exercise the research and integer-execution contracts on harmless amber/violet marker data. Dataset SHA-256: a7b2c3f3b978623273b72fa3853344b4059725a5b147540c5a1e906a11977ae7. Fixed config: fixture. There are 24 training, 12 calibration and 16 test records, balanced within splits.
Local software evidence: make check passes lock validation, Ruff lint/format, strict mypy, 29 tests and local Markdown file-link checks. make package-check builds wheel/sdist, checks required contents and installs the wheel with locked core dependencies in a fresh environment, then runs the installed CLI. Host: macOS arm64, Python 3.12.12, uv 0.9.18. Training used one BLAS/OpenMP thread. CI definitions are present; these local results do not establish hosted CI status.
Repeat experiment commands, each with OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1:
uv run --locked --extra train hallmark run configs/fixture.json --output runs/bootstrap-003
uv run --locked --extra train hallmark run configs/fixture.json --output runs/bootstrap-004
uv run --locked hallmark audit-run runs/bootstrap-003 --dataset fixtures/smoke.jsonl --expected-manifest 4326f9b8987b6e465ab2bc32d15d2a94eef76d3c21c1ff8b79bc6f36a8fe9088
diff -qr runs/bootstrap-003 runs/bootstrap-004
Both complete run directories were byte-identical, with manifest SHA-256 4326f9b8987b6e465ab2bc32d15d2a94eef76d3c21c1ff8b79bc6f36a8fe9088. Replay returned research-replay-matched. Source snapshots identify the exact pre-publication code; the later CLI type annotation changes no arithmetic. Runs remain local and gitignored, not hosted evidence.
Negative result: threshold -273; calibration TP=6, FP=0, TN=6, FN=0; test TP=8, FP=8, TN=0, FN=0. Test FPR is 1.0. The report correctly says insufficient-evidence; this is not a useful classifier result. The fixture includes split names in its text, a visible distribution shift/confound for the weak byte baseline. That is a candidate explanation, not an experimentally isolated cause. We did not tune the threshold or change the fixture after seeing test outcomes. Even a perfect result on eight benign examples would lack power for a 1% FPR claim.
Earlier bootstrap-001/002 were also retained; repeated model/prediction/result files matched. Tamper tests reject changed reports under the original hash and mathematically inconsistent reports even after rehashing a bundle. Those checks exercise software integrity/replay only. No real safety-policy, hardware or proof result is claimed.
Hosted execution: the first private GitHub run could not start any steps because of an account payment/spending-limit block. This is not a hosted pass or a code-test failure. Billing settings were not changed.
The local Python 3.13 core-only check caught a mypy configuration gap: the optional NumPy import was not exempted when NumPy was absent. Adding it to the existing missing-optional-import override fixes that environment without installing training dependencies into core. The train-enabled environment still checks against installed NumPy types. This is a tooling fix, not a model, data or evaluation change.
Final local matrix: Python 3.12.12 with the train extra, make check, 29 passed; fresh Python 3.13.11 core-only environment, make check, 28 passed and the optional train test skipped. Both passed strict typing, lint/format and docs checks. Fresh installed-wheel/CLI checks passed on both interpreter versions. Linux/hosted CI remains unexecuted because of the account block; these are local macOS results, not substituted hosted passes.
2026-10-20: v1 specification corpus and implementation decomposition¶
Hypothesis: the historical hybrid POC can be expressed as one coherent, typed v1 contract with every implementation deliverable mapped to a bounded work item, while preserving the existing research evidence. Baseline: commit 1024ae9a0598ae842e86eaf98c479c5d11ea1f23, no issues or PRs, software foundation only. Mutable paths: root specification/navigation, docs/spec/, docs/rfcs/, docs/roadmap/, decision 0002, state/ledger/evidence and source-distribution inclusion. Runtime, evaluator, dataset, fixture config and old results were not changed. The existing .gitignore edit is unrelated and was preserved.
Raw identities/counts and commands were saved first in spec bootstrap evidence. Local validation: smallest contract/reference/docs fixture, 18 passed; OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 make check, 29 passed; make package-check, passed including isolated installed CLI. The initial formatting gate failed on Python signatures inside a documentation code block; formatting was corrected and the full gate rerun successfully. The structural audit checked 23 index/section/RFC files, all 10 exact RFC templates, local anchors, 40 unique work items and an acyclic dependency graph. Archive inspection found every corpus/work-map path and no harness, runs, data, checkpoints or environment-secret paths.
The corpus preserves genuine hardware, rule/MLP/encoder proofs, B1/B2, complete-input binding, independent verification and private reviewer reproduction. Open questions have owners and resolution checkpoints. Drafting a requirement or filing its issue does not perform the experiment or authorize its resources. No statistical, hardware, proof, performance or security result was produced.
Privacy correction: live repository metadata showed public visibility and disabled issues. Before publishing this corpus, visibility was changed to private and verified; issues were enabled under the explicit specification/backlog request. This does not establish whether older public material was copied. No collaborators or billing settings were changed.
Acceptance requires both the pushed corpus/PR and the complete live issue set with tracker and dependency links. Publication receipts are recorded in the follow-up entry after filing; the local checks alone do not complete that gate.
Private publication and live backlog audit¶
Phase 1 was committed in four logical commits and pushed as spec/bootstrap-2026-10-20; corpus commit ad058ad74a6dc23ae95cc523d7bf468a06d4c4ce is the immutable reference used by the issues. PR #1 remains open for owner review.
Phase 2 used gh label list/create, gh api for milestones, gh issue create --body-file, gh issue edit --body-file and gh issue comment --body-file. Initial gh issue list --state all --limit 200 found no issues. Created 40 implementation issues (#2–#41) and ten RFC trackers (#42–#51), with exactly one type/area/priority/effort label per issue, applicable RFC labels, and milestones v0.2/v0.3/v1.0. The total milestone counts, including trackers, are 16/19/15. No issue was closed or represented as implemented.
The live publication audit was saved before this summary. It fetched all 50 live issues and compared their complete bodies with the work map, checked metadata and pinned references, validated all 102 implementation dependency edges and all 50 dependency comments, and compared each tracker checklist with its exact child set. Missing labels, unmapped items and missing children: zero. The tracker and work map carry all returned issue URLs and owner/date resolution gates.
Final local checks after adding the tracker: OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 make check, 29 passed; make package-check, passed. The source distribution includes the tracker and complete corpus; forbidden harness/run/data/checkpoint/secret paths were absent. Runtime, evaluation code, fixture/config, tests and lock remained unchanged from the baseline. Documentation structure/anchor checks passed.
Hosted spec run 39393331644 on ad058ad completed without starting any job steps. Its check annotation reports an account payment/spending-limit block. This is unavailable hosted execution, not a test failure or pass; final-head delivery/check state is reported on the PR. Billing was not changed.
The deliverable is the proposed corpus plus filed backlog, not a completed v1 system. Next bounded research work is the admission proposal in issue #14. Hardware/proof/data campaigns, independent implementer access, policy adoption and reviewer sharing remain gated. Prior public exposure is unassessed.
2026-10-20: owner-authorized merge and public visibility¶
At the owner's request, PR #1 was merged as f0f117285d2ccf82b122d1155b8a1fb56a2a2217; local and remote main matched. Its tree matches the previously validated PR head 3951d1f49c9e5f91c0ef4c592c973d95dfa0828a. Hosted run 39396000433 started no job steps because of the same account billing/spending-limit block.
The owner then explicitly requested public visibility and instructed agents to keep it public unless explicitly asked otherwise. Executed gh repo edit AbdelStark/hallmark --visibility public --accept-visibility-change-consequences. Verification with gh api repos/AbdelStark/hallmark returned private: false and visibility: public; an unauthenticated HTTP request to the repository returned 200. AGENTS.md records the standing instruction, superseding the earlier private-repository instruction. The unrelated .gitignore edit was preserved. No research execution authority was expanded.
Documentation validation: uv run --locked python scripts/check_docs.py passed; OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 make check passed all gates with 29 tests. Package checks were not repeated for these instruction/state/ledger edits.
2026-10-20: W01 strict v1 primitive encoding¶
Hypothesis: a bounded RFC-0001 codec can give each supported value one canonical
representation while preserving the existing research identities. Baseline:
48f2d7066b62a42f971cd10dbbc2cdbae720b82e; the legacy contract fixture passed
14 tests and the new fixture failed because hallmark.wire was absent. Mutable
paths: new wire module/tests, package smoke, schema/API documentation, decision
0003, changelog and this state/evidence record. No data, evaluation, training or
legacy identity code changed.
Saved source/environment hashes and numeric results
before this entry. The codec accepts only the restricted primitive types and 22
domains, rejects duplicate/unsorted keys, overlong forms, trailing data, tags,
invalid Unicode, integer overflow and resource excess. Fixed vectors, seeded
differential checks against cbor2 and truncation tests passed: 108 codec tests.
An initial mypy gate caught a redundant cast; it was removed. Final make check
passed 137 tests plus lint, format, typing and documentation checks;
make package-check passed including an isolated installed codec smoke.
The separate v1 module preserves both recorded legacy digest fixtures. No artifact schema, signature, attestation, proof or security-review result is claimed. Decision 0003 records limits and the raw codec/schema boundary; owner review before enrollment remains outstanding.
Delivery: PR #52 merged W01 as
516143e0fb7ae696ab76a1a6d8a314496a60f6a3; issue #2 closed. Hosted checks passed
on both the PR head and the merge commit,
including both core interpreter jobs and the train fixture.
2026-10-20: W30 typed errors and CLI privacy¶
Hypothesis: one allowlisted diagnostic boundary can prevent private input/path
and external exception text from reaching CLI output, while preserving success
output and research arithmetic. Baseline: 516143e0fb7ae696ab76a1a6d8a314496a60f6a3.
A missing-file fixture reproduced a path-sentinel leak with exit 2.
Mutable paths: error module, CLI boundary, wire error inheritance, privacy tests,
package smoke, error/API docs, decision 0004 and state/evidence records.
Saved source/environment hashes and raw results
before this entry. All 19 codes are exercised through the CLI, including exception
notes/causes. Additional tests cover argument parsing, malformed JSON/UTF-8,
private paths, timeout/OOM, child-process stderr, failed result serialization,
stage precedence and retained failed-run artifacts. The 52 privacy tests pass;
make check passes 189 total tests plus lint, format, typing and docs.
make package-check passes with an installed invalid-argument diagnostic smoke.
Initial mypy attribute annotations and Ruff zip strictness failures were fixed
and retained in the evidence record.
Input errors retain exit 2; execution/storage/resource failures now exit 3. Unexpected exceptions get random local diagnostic IDs. No error returns success. No arithmetic, dataset, evaluator or historical bundle was changed. Event persistence, process isolation and real provider integrations remain separate gates; this does not establish hardware/proof evidence or sanitize old snapshots.
Delivery: PR #53 merged W30 as
ad664aa16f591c6c9fef08fa6b309faa0b69c1b6; issue #31 closed. Exact-head PR and
merge-commit checks
passed. The repository remained public.
2026-10-20: W36 required training and interpreter matrix¶
Hypothesis: the train CI job can reject missing execution without changing the
core-only environment's legitimate optional skip. Baseline:
ad664aa16f591c6c9fef08fa6b309faa0b69c1b6. A forced-skip fixture run with
pytest -m train returned exit 0 with one skipped test, confirming the false-green
case. Mutable paths: pytest guard/tests, Makefile, workflow, contributor/operations
docs and state/evidence records. Runtime, data, training and evaluation code
were unchanged.
Saved source/config/environment hashes and raw numeric results
before this entry. The guard's 16 subprocess tests cover real execution,
optional core skipping, missing extras, call/setup/teardown skips, partial
execution, xfail/xpass and collection/setup-only runs. make check-train sets
one BLAS/OpenMP thread and requires successful setup/call/teardown for every
selected training test, with at least one training test selected.
Fresh macOS arm64 environments used CPython 3.12.12, 3.13.11 and 3.14.6 with
uv sync --locked --python VERSION. On each: make check passed 204 tests and
skipped the absent optional training dependency; make package-check passed;
then make check-train installed the locked extra and executed one training test
successfully with 204 other tests deselected. These are six local core/train
configurations, not classifier-quality measurements. An initial Ruff E501 test
literal failure was fixed; its output was preserved before fresh attempts.
The workflow runs core/package and training separately on all three versions. Inspection confirms read-only contents permission, disabled checkout credential persistence, no secrets, no self-hosted runners and no artifact uploads. Workflow dispatch runs the same fixtures; it is not authority for a privileged integration job. Real proof/hardware execution still requires its separate protocol, host and budget. Hosted delivery must be checked on this PR's exact commit before integration.
Delivery: PR #54 merged W36 as
6b2a6acda5813ddcdf9d14554a99ada8ad2e6245; issue #37 closed. Exact-head PR
checks and all six merge-commit jobs
passed. The repository remained public.
2026-10-20: W03 immutable model manifests¶
Hypothesis: a strict frozen manifest can bind every execution-relevant field
while rejecting malformed references before explicit file inspection. Baseline:
6b2a6acda5813ddcdf9d14554a99ada8ad2e6245, with 108 passing existing wire tests.
Mutable paths: the new manifests module/tests/vector/schema, installed-package
check, API/decision docs and state/evidence/tracker records. The existing research
hashes, dataset, evaluator and training protocol are unchanged.
Saved source/config/environment hashes and numeric results
before this entry. The selected gate passes locally: 128 manifest tests,
333 total tests under make check, and make package-check, including the
351-byte fixed CBOR vector, typed model identity and bounded rooted-file check
from an isolated installed wheel. Python 3.12.12/macOS arm64 used the locked
environment, one training thread and harmless local files only.
Negative cases cover path traversal/absolute paths, Unicode byte bounds, unsorted/duplicate/oversized lists, integer/boolean confusion, invalid greedy configuration, missing/extra fields, unknown versions, malformed wire values, unsafe unchecked model copies, symbolic/hard links, filesystem aliases, FIFOs, file size/hash mismatches, directory swaps and files changing during reading. Parsing performs no I/O. Every valid file/digest/generation/numerical-mode mutation tested changes identity; unsupported generation changes are rejected.
The initial vector setup found a JSON decoding problem caused by manually converting a file list before Pydantic validated nested strict bytes. The manual conversion was removed; JSON projection round trips and wire byte/text rejection now pass. Two focused lint findings were also fixed. These failures remain in the evidence record rather than being counted as successful checks.
The schema interpretation and path/inspection limits are recorded in decision 0005. Wire-profile enrollment still needs owner/independent review. The file checker describes opened files, not a protected later load, genuine attestation, classifier quality or proof. No real model was loaded or downloaded. Hosted checks must pass on the exact PR commit before merging.
The live #3/#6 dependency correction is now mirrored in the work map and tracker: W02 requires W05 for typed request expectation/context binding. The current graph has 103 edges; the original publication audit remains an immutable 102-edge record. Next implementation order after W03: #5, #6, then #3.
Delivery: PR #55 merged W03 as
f844342c18af85d8cb0b1160521c58938f9d87ea; issue #4 closed. All six exact-head
PR and merge-commit jobs
passed. The repository remained public.
2026-10-20: W04 classifier and arithmetic contracts¶
Hypothesis: immutable classifier/profile identities plus exact parameter-file
binding can reject malformed topology, shapes, masks and integer range
declarations without loading a model or executing a graph. Baseline:
f844342c18af85d8cb0b1160521c58938f9d87ea, with 128 passing model-manifest tests.
Mutable paths: v1 contracts/shared serializer, tests, schema/vector fixtures,
package checks, the mypy TypeForm option and specification/evidence/tracker docs.
The legacy dataset, evaluator, training protocol and research hashes are unchanged.
Saved source/config/environment hashes and numeric results
before this entry. The selected gate passes locally: 133 classifier contract tests,
all 128 model regressions, 466 total tests under make check, and
make package-check with an isolated installed-wheel classifier/profile/parameter
check. The locked Python 3.12.12/macOS arm64 environment used one training thread
and synthetic local metadata. No model, dataset or proof backend was downloaded.
Each allowed operator family has positive and negative shape/range cases. Further cases cover duplicate producers, forward references/cycles, unused declarations, boolean/integer confusion, capacity boundaries, parameter hash/schema mismatches, masking, lookup bounds, overflow, negative ties-to-even, mixed scales, immutable nested records and unvalidated copies. Large tensor dimensions are checked as bounded-rank metadata without allocating their data. The initial identity-mutation test accidentally left unused roots; correcting the test preserved the validator's rejection. Its failed output hash is retained.
The checked parameter union uses Pydantic's TypeForm annotation. Enabling the locked mypy implementation's TypeForm support resolved its typing error without suppressing checks. The shared serializer preserves all prior model golden bytes, identity and schema projection.
Decision 0006 records the exact operator files, scale/rounding rules, graph bounds and separate parsing/parameter validation phases. Static interval bounds are conservative and do not validate actual external tensors, program semantics, classifier quality, backend behavior, hardware, proofs or independent verification. WIRE-REVIEW/INTEGER-GRAPH owner gates remain. Exact-head hosted checks must pass before integration.
The work map now records W04's actual dependency on W03's file references (104 edges total); historical audits are untouched. After W04, #6 policy/request expectations is the next contract gate, then #3 transcript/input binding.
Delivery: PR #56 merged W04 as
f456b4a749d795f2f22f19e5d2ae5fd90b099539; issue #5 closed. All six exact-head
PR and merge-commit jobs
passed. The repository remained public.
2026-10-20: W05 policies and caller request expectations¶
Hypothesis: strict policy/request records plus explicit resolved manifests and
caller expectations can reject classifier substitution, expectation mismatches
and byte-limit violations before later binding or verification. Baseline:
f456b4a749d795f2f22f19e5d2ae5fd90b099539, with 261 passing model/classifier tests.
Mutable paths: the new policy module, tests, schema/vector fixtures, package check
and specification/evidence/tracker docs. Legacy datasets, evaluators, training
protocols, research hashes and the existing model/classifier modules are unchanged.
Pass criterion: all #6 schema/mutation/continuity/byte-limit cases, make check
and isolated distribution checks pass with no fallback to sender-supplied trust.
Saved source/config/environment hashes and numeric results
before this entry. The selected local gate passes: 95 policy/request tests and all
261 manifest regressions, 561 total tests under make check, and
make package-check including installed-wheel decoding and explicit expectation
comparison. Runs used the locked Python 3.12.12/macOS arm64 environment, one training
thread and harmless synthetic fixtures. All commands completed below 60 seconds.
Negative cases include coordinated digest/manifest reordering, duplicate kinds with distinct identities, changed audience/nonce/interaction/model/policy, required nullable continuity, malformed sequences/predecessors, oversize multibyte prompts, invalid Unicode, unchecked copies and sanitized errors. Exact UTF-8 is preserved, and repeated checks perform no I/O or nonce consumption. The fixed vector's nonce and interaction bytes are public test values.
Decision 0007 records the
versioned external expectation, classifier resolution and comparison boundaries.
inj.direct/1 remains a proposal until owner/relying-party admission. These checks
do not authenticate enrollment, prove freshness or history, execute classifiers,
or verify hardware/proofs. No data/model download or policy adoption occurred.
Exact-head hosted checks must pass before integration; W02 / #3 is next.
Delivery: PR #57 merged W05 as
1d344d55b44efae9b0232c46d4553dfe59f49288; issue #6 closed. All six exact-head
PR and merge-commit jobs
passed. #3, #7 and #9 were unblocked after checking their live prerequisites.
The repository remained public.
2026-10-20: W02 exact transcript and input openings¶
Hypothesis: the specified transcript/context/input preimages bind exact roles,
caller expectations and every original input unit, reproducibly across two
language implementations without hardware. Baseline:
1d344d55b44efae9b0232c46d4553dfe59f49288, with 203 passing wire/policy tests.
Mutable paths: a separate opening module, tests, schema/vector fixtures,
JavaScript vector checker, package checks and specification/evidence/tracker docs.
Pass criterion: all field, Unicode, salt, nonce and late-unit mutations change
the corresponding identity or fail, both encoders match fixed vectors, and
make check/make package-check pass. Legacy hashes and experiments stay unchanged.
The first empty-opening fixture matched a literal CBOR preimage. Saved
source/config/environment hashes and numeric results
before this entry. All 69 opening tests and 203 wire/policy regressions pass;
make check passes 630 tests and static/doc checks. make package-check passes
isolated installed-wheel transcript/context/byte-input reconstruction.
Commands completed below 60 seconds in the locked Python 3.12.12/macOS arm64
environment with one training thread; no data/model/backend was downloaded.
Node v24.18.0, already installed, ran the separate JavaScript CBOR/hash checker using standard libraries only. It agrees on seven transcript/context and 13 input vectors, including U64 IDs stored as decimal strings to preserve exactness. Changing the final token without updating its committed preimage correctly causes an assertion failure; its output hash is retained as an expected negative result. This is a separately written implementation by the same contributor, not independent reviewer or full verifier acceptance. Node is optional developer tooling and was not added to package dependencies or mandatory CI.
Decision 0008 specifies raw UTF-8 role preimages, caller-bound context, per-classifier salts, limits, opening privacy and the incompatible v0/v1 formats. Byte input is derived from the whole prompt; token input requires independently computed original units and an explicit vocabulary bound. The helper validates integer ranges and hashes all supplied units, but cannot authenticate a tokenizer or prove text-to-token correspondence. Chunk execution, complete circuit coverage, enrollment, real proofs/hardware and independent verification remain separate gates. Synthetic salts are public test values. Exact-head hosted checks are next, then #7 receipt contracts.
Delivery: PR #58 merged W02 as
fbd6c90f65e5e535f8dfb839012f1bd3bd544097; issue #3 closed. All six exact-head
PR and merge-commit jobs
passed. #12 was unblocked after checking its live prerequisites; #11 still needs
14 policy/data/resource admission. The repository remained public.¶
2026-10-20: W06 receipt, epoch and proof-container contracts¶
Hypothesis: strict immutable containers and explicit declared-binding checks can
reject malformed layouts, classifier substitutions, interval mismatches and field
aliases without pretending to verify their evidence. Baseline:
fbd6c90f65e5e535f8dfb839012f1bd3bd544097, with 164 passing policy/opening tests.
Mutable paths: a separate receipt module, tests, schema/layout/vector fixtures,
package checks and specification/evidence/tracker docs. Pass criterion: #7's
structural, time, identity, packing and size failures are rejected, golden fields
round-trip, and make check/make package-check pass without an acceptance stub.
Legacy identities, evaluators and experiment protocols stay unchanged.
The smallest fixture matched the exact 100-element layout and minimum signed
threshold offset. Saved
source/config/environment hashes and numeric results
before this entry. All 145 receipt-contract tests and 164 policy/opening regressions
pass; make check passes 775 tests and static/doc gates. make package-check
passes installed-wheel record decoding, declared reference/time checks, signature
message reconstruction and signed-threshold unpacking. Runs used locked
Python 3.12.12/macOS arm64, one training thread and commands below 60 seconds.
No model, dataset, hardware host or cryptographic/proof library was downloaded.
Cases cover missing/extra/reordered classifiers, duplicate identities, exact profile chunk counts, every public-input position, negative/overflow/boolean field aliases, signed int32 extrema, external expectation mutations, continuity, one-hour epochs, issuance/future/expiry boundaries, signature/key sizes, opaque payload caps, immutable copies and sanitized failures. Five schema snapshots plus the field-layout file are shipped. Synthetic proof/signature/quote payloads are labeled unverified and there is no verifier or accepted/verified result.
Decision 0009 fixes the record IDs, complete epoch identity versus quote challenge, exact signature message, 16 MiB per-payload caps and explicit caller-clock/reference boundaries. Actual Ed25519 encoding/signature checks, attestation/collateral validity, independently enrolled proof verification, complete circuit coverage and policy/data admission remain required. Structural success does not establish any of them. Exact-head hosted checks are next, then #9 external enrollment; reassess #8 and #12 against their actual dependencies.
2026-10-20: W08 external registry enrollment and atomic checkpoints¶
Hypothesis: mandatory external trust and a retained checkpoint, complete signed
successor links and a transactional compare-and-swap update can reject rollback
and substituted identities without losing the prior trusted state on failure.
Baseline: 774887d0e40b82459d4f7ca8657400d9905577ad, with 240 passing receipt/policy
tests. Mutable paths: enrollment and checkpoint-store modules, signature dependency
pin/lock, focused tests, schema/signed fixtures, package checks and relevant
specification/evidence/tracker docs. Pass criterion: #9's trust, history, freshness,
rollback and write-failure cases reject; offline local and package gates pass.
The smallest signed fixture round-tripped under cryptography 50.0.2 / OpenSSL 4.0.3. The dependency review recorded exact versions, distributions/hashes, published licence metadata and primary security sources before installation; the final lock's new hashes match that inventory. No operational roots were created. Ed25519 verification is necessary for W08's history gate; W09 retains authoring, retained snapshot contents and key transitions.
Saved source/config/environment hashes and numeric results
before this entry. All 88 enrollment tests and 240 receipt/policy regressions pass.
make check passes 863 tests and static/doc gates; make package-check passes
isolated installed-wheel signature verification, external-state provisioning and
committed-state readback. The first full check failed on one overlong package-check
source string; its failed result is preserved and the final gate passed after
splitting that literal. Commands used one training thread and stayed below 60 s.
Tests include equal-sequence different-digest rejection, skipped/reordered history, invalid intermediate signatures, threshold/duplicate/unknown signers, signature malleation and domain substitution, strict clocks, age/expiry boundaries, retirement, external allowlist intersection, history byte/nesting/deadline bounds, stale writers, concurrent locks, rejected writes, deferred-constraint commit failure, silent update suppression and process exit before commit. Stored checkpoints remain unchanged in the tested failure cases. Four schemas and a fixed signed corpus are shipped.
Decision 0010 documents caller-controlled provisioning, trust identity, chain semantics, process-local enrollment, retirement and SQLite/POSIX assumptions. Physical power loss, whole-store rollback, malicious OS-account changes and global equivocation prevention are not established. Process termination/RSS enforcement, operational root custody, receipt/hardware/proof verification and policy/data admission remain separate gates. No model/data retrieval, external inference, GPU or paid compute occurred.
W06's final head 2fb4bf170503bc7bc4bea4b1c155bdc5fd0e51da was merged in
PR #59. All six
merge-commit jobs
passed on this turn's baseline. W08 exact-head hosted checks are next, then #8,
10 and #12 according to actual prerequisites. The work map now mirrors #8's¶
necessary #9 dependency, for 105 edges; the original publication audit is unchanged.
2026-10-20: W07 offline bundle, response and reported-result contracts¶
Hypothesis: strict bounded containers can reject malformed embedded files,
incomplete or inconsistent declarations, private protocol records and contradictory
collection/claim reports without turning parsing into evidence verification.
Baseline: 866d36a69762104847f1ded8253829a4e11dd786, with 233 passing receipt and
enrollment tests. Mutable paths: one bundle-contract module, focused tests,
schemas/vectors, package checks and relevant spec/evidence/tracker docs.
Pass criterion: #8's path/file/proof/privacy and status cases reject, total/caller
limits hold, and make check/make package-check pass. Existing identities,
evaluators and research protocols remain unchanged.
The smallest fixture round-tripped a 3,998-byte structural bundle; its receipt,
hardware and proof bytes remained unverified. Saved
source/config/environment hashes and numeric results
before this entry. All 91 bundle/result tests and 233 receipt/enrollment regressions
pass. make check passes 954 tests and static/doc gates. make package-check
passes decoding/re-encoding all 17 structural/report examples and bounded bundle
parsing from the installed wheel. Four schema snapshots are frozen. Targeted lint
initially found an unused test import and an overlong package-check literal;
both were fixed before the recorded final gates. Commands used one training
thread and stayed below 60 seconds.
Cases include traversal/locators, duplicate or unsorted paths, exact sizes/hashes, missing/mismatched collateral, three ordered proof declarations, receipt/epoch links, caller depth/byte/per-proof caps, total container overhead, missing/extra fields, canonical private records disguised as files, response/opening binding, UTF-8 boundaries, blocked/failed/expired payload omission, collection status/error coherence, claim strengths, residuals/assumptions, pending versus rejected proofs, policy block and external required-claim checks. A parsing test replaces file opening, socket creation and registry verification with failure guards. Malformed copies are revalidated before use; private payloads are omitted from repr/errors.
Decision 0011 records canonical wire limits, separate opening exchange and the boundary between reported claims and an actual verifier. Canonical embedded opening/response records are rejected; arbitrary opaque or differently encoded bytes still need an admitted artifact inventory. No signatures, hardware or proofs become accepted through parsing. Full artifact closure, release persistence/recovery, worker time/RSS enforcement, operational enrollment and independent verification remain downstream gates. No new dependency, data/model retrieval, external inference, hardware host, GPU job or paid compute was used.
W08 was merged in PR #60 at
866d36a69762104847f1ded8253829a4e11dd786, with all six
merge-commit jobs
passing. W07 now explicitly depends on completed W02 for its private opening type;
source and live #8/#3 dependency records agree. The 106-edge work graph is acyclic;
the original 102-edge publication audit is unchanged. W07 exact-head hosted checks
are next, followed by #10 and #12 according to actual prerequisites.
2026-10-20: W09 registry authoring, rotation and offline checking¶
Hypothesis: explicit dual-threshold transitions, immutable complete authoring
history and atomic checkpoint/trust updates can reject one-sided rotation,
retirement reversal and retired-policy reuse while preserving prior state on
write failures. Baseline: 302a861043636094295c4dc06832e4a8c06afe87, with 140
passing enrollment/privacy tests. Mutable paths: registry authoring/history/CLI,
the shared signature checker and checkpoint store, focused tests, schemas/vectors,
package checks and relevant spec/evidence/tracker docs. Pass criterion: #10's
signer, transition, retirement and persistence cases reject; offline local and
installed-package gates pass. Prior protocols/evaluators and results are unchanged.
The smallest fixture passed one-key-per-side rotation at sequence 1 with a
664-byte signed transition. The committed corpus exercises two-key thresholds
with one shared key and separate old/new role messages. Stored
source/config/environment hashes and numeric results
before this entry: 97 new tests, 237 focused tests, 1,051 total tests through
make check, and a passing make package-check. The installed-wheel check retains
old/new snapshots and transition bytes, invokes the actual registry-check CLI,
reads back the committed state and rejects a replay. All commands used one
training thread and stayed below 60 seconds.
Tests cover duplicate/unknown/insufficient signers; old-only/new-only, swapped and wrong-domain signatures; shared-key role reuse; complete trust/activation/reason/ review binding; immutable prior metadata; missing/expired/retired entries; retirement reversal; stale policy reuse; malformed records and constructor bypasses; byte/time limits; unsafe paths/FIFOs; private CLI failures; stale writers, locks, suppressed/failed writes, failed commits and process exit before commit. The two stores preserve their prior records and trust binding in tested failures. Two schemas and one exact signed rotation corpus are committed.
Review found that sequence/digest-only checkpoints could permit later reuse of the old policy. The store now records external trust on first successful enrollment and changes that binding only with a dual-authorized rotation, in the same transaction as checkpoint state. An unrelated policy change requires owner re-enrollment. Early lint/type checks found line lengths, import order and one digest annotation; these were fixed before the recorded final gates. Final migration review added an atomic HMR1-to-HMR2 application-ID upgrade so older readers cannot ignore the trust binding. Executing the actual W08 reader confirmed it accepts the unbound checkpoint and rejects the upgraded checkpoint; the current reader accepts the committed result. Missing trust in HMR2 fails closed. The full local and installed-package gates were rerun on the corrected source.
Decision 0012 fixes signature role tags, activation, irreversible retirement, append-only local history, lazy checkpoint-format extension and the CLI boundary. No operational key was created, enrolled or rotated. The pinned signature implementation and its pre-install review are unchanged; private custody approval remains absent and required before operational roots are enrolled. The local log does not prevent split views or whole-store rollback by the OS account. Physical power loss, independent review, worker RSS/process isolation, genuine hardware/proofs and policy quality remain unverified. No data/model download, external inference, GPU or paid compute occurred.
W07 was merged in PR #61 at this turn's baseline, with all six main jobs passing. W09 exact-head hosted delivery is next. Its completion makes #34's software corpus dependencies available; reassess that gate before #12, then
32/#38/#39. The full MVP still requires admitted data, real classifiers, proof¶
and hardware integration, and independent verification.
2026-10-20 — W33 frozen component conformance¶
Hypothesis: a caller-pinned corpus can compare exact component outcomes and
detect wrong expectations while preserving every unavailable genuine-evidence
obligation. Baseline: 9357288387bfbcdf0ad28ae34f54b37fe70a1333, with 205 parser/
registry tests passing. Mutable paths were the corpus contracts/adapters, builder,
runner, schema snapshots, tests/vectors, package/Make checks and their documentation.
Pass criterion: every executable vector matches, a deliberately wrong expectation
fails, full scope rejects required unavailable evidence, and local/package gates
pass without treating the primary implementation as independent.
Saved source/config/environment hashes and numeric results
before this entry. The frozen identity is
febda826f0d2cb8c195c8637014f8ca1f5f017cfea709030ca5fbe76d5a9093e.
There are 188 cases: 156 component matches, zero mismatches and 32 unavailable
real-proof/hardware/hybrid obligations. All 15 original PRD cheats and all 19
error codes remain mapped. The builder reproduces every frozen byte using the
frozen source snapshots and refuses an existing destination.
The first builder candidate failed on an incorrect fixture key. The second expected a profile's inclusive 32,768-byte limit to reject; the implementation correctly accepted it, producing one mismatch and exit 1. Preserved both failed directories. A new candidate added separate exact-limit, max-plus-one and complete tail cases. Expected values were not copied from runner output.
All 46 harness tests and 205 parser/registry regressions passed. make check
passed 1,097 tests, lint, format, typing, documentation links and the software
conformance self-tests. make package-check passed with all pinned corpus files
in the sdist and both runner scopes against the installed wheel. Software scope
exits 0; full-evidence scope exits 3 with all 32 unavailable cases retained. The
wrong-expectation self-test requires exactly one mismatch/exit 1. Tests reject
wrong pins, changed files, unsafe paths, invalid ancestry, missing contexts and
falsified report counts. Unexpected implementation exceptions abort the run.
Decision 0013 defines the source snapshots, raw request hash, typed corpus identity, component-only accepted flag, bounds and immutable migration. Every executed case reports both claims not accepted. Full graph arithmetic, actual evidence, independent implementation/ review, operational trust and policy quality remain absent. No private transcript, operational key, data/model download, external inference or paid run was introduced.
W09 was merged in PR #62, with all six main jobs passing. W33 exact-head hosted delivery is next. #35/#36 still require the complete verifier and genuine evidence work. After delivery, reassess #12 before #32/#38/#39.
2026-10-20 — W11 exact native integer MLP¶
Hypothesis: a bounded byte-count MLP can bind complete exports, agree with a
separate formula at every layer and reject int32 overflow before rescaling while
preserving complete prompt coverage. Baseline:
89de51b49cfbf10040384967b289a32b9709cb02, with 248 classifier/opening/conformance
tests passing. Mutable paths were the new runtime/export records, fixed vectors,
builder/checker, schemas, tests, package check and associated documentation.
The pass criterion was exact layer agreement plus negative identity, shape,
range, padding, overflow and resource tests, followed by full local/package gates.
Saved source/config/environment hashes and raw numeric results before this entry. All nine fixed vectors matched. Sixty cases with seed 20261020 compare every count, accumulator, rescaled hidden value, ReLU value and score against direct per-byte sums and signed quotient/remainder rounding. The primary runtime uses count-vector dot products and the existing exact Fraction rescaler. This is a separate formula check with shared repository provenance, not W34 independence or a proof of execution.
All 75 runtime tests and 248 regressions passed. Coverage includes both int8 extrema, int32 endpoints, positive/negative halfway ties, threshold equality, empty input, actual NUL versus padding, split UTF-8, a late positive chunk, complete 32,768-byte input, raw accumulator overflow/underflow hidden by rescaling, cancelling-prefix overflow and execution caps. The smallest empty-input case returned one zero-count chunk with native score zero and no evidence claim.
A new negative test reproduced acceptance of a weight path colliding with an operator path. Preserved the failed output and source hash; added a pre-execution rejection. No file was read through that path, but accepting an ambiguous export would have left its file inventory inconsistent. The final full gate passed 1,172 tests, lint, format, typing, documentation links and unchanged W33 software conformance. Isolated installed-wheel validation matched all nine MLP vectors. The three schema snapshots and the deterministic fixture rebuild also passed.
Decision 0014 fixes export identity, raw weight-file hashing, the three-node numerical profile, bias-first int32 accumulation and exact rescaling. It caps weights at 65,536 bytes, execution at 1,024 chunks and 2,097,152 products. Over-budget input is rejected without a partial result; these caps are not worker isolation or measured performance. Diagnostic traces disclose counts/activations on serialization and belong on private authorized channels. Native verdicts establish neither policy quality nor hardware/proof acceptance. No training, calibration, model/data download, external inference, GPU or paid run occurred. Prior identities and the frozen W33 corpus remain unchanged, including its 32 unavailable genuine-evidence obligations.
W33 was merged in PR #63, with all six main jobs passing. W11 exact-head hosted delivery is next; its dependents still need admitted data, qualified proving or the other native/hardware components. Reassess #32, then
38/#39. The full MVP remains incomplete.¶
2026-10-20 — W31 bounded private event attempts¶
Hypothesis: a private bounded event log can retain failures and reject invalid
metadata, duplicate ordinals, concurrent writers and uncertain writes without
promoting a completed stage into a completed attempt. Baseline:
e8c9d0be603d477e3861b6843538477974b97354; all 55 error-privacy/pipeline tests
passed. Mutable paths were the new writer/reader, schemas, event tests, installed
fixture/package checks and their documentation. The pass criterion was the
smallest complete stage fixture plus negative privacy, sequence, concurrency,
resource and durable-write tests, followed by make check and make package-check.
Saved source/config/environment hashes and raw numeric results
before this entry. The first privacy test exposed unsuppressed Python exception
chaining in generator context managers: nine sentinel cases failed. The output
and source hashes were retained; final exceptions suppress external chains and
the tests check both serialized files and formatted exception output. An
intermediate fixture test also failed because scripts was not importable as
a package. It now executes the actual fixture with a bounded subprocess.
All 101 new tests and 55 regressions passed. The complete local gate passed 1,273 tests, lint, formatting, typing, documentation links and the unchanged 156-case W33 software conformance scope. The isolated installed wheel retained one failed attempt and one explicit successful retry across six events. Three schema snapshots matched their runtime definitions. No training, model/data retrieval, external inference, GPU or paid execution occurred.
Failure tests cover all nine sync points, failed/short/zero writes, publication failure, byte/event caps, malformed metadata, duplicate/gap ordinals and mismatched receipt hashes. Four subprocess exits preserve incomplete attempts, including a retry interrupted before its first stage. Separate processes, a thread and a fork cannot reuse a live writer. Unsafe modes, symlinks, hardlinks, FIFOs and unexpected directory entries reject. Durations use a controlled monotonic clock; negative or out-of-U64 values fail. This is correctness testing, not timing data for a full-path performance claim.
Decision 0015 fixes the private storage protocol and bounds. Failures remain local and metadata has the run's access/retention policy; there is no hosted telemetry or automatic cleanup. The reader cannot authenticate history or prove that a supplied UUID was random. Physical power loss, network filesystems and compound storage failures preventing both failure-marker retention and receipt removal remain unqualified. Quarantine ambiguous archives. Legacy manifests/results and the frozen W33 corpus are unchanged.
W11 was merged in PR #64, with all six main jobs passing. W31 exact-head hosted delivery is next. W32 still requires #14 policy/resource admission and #19 proof qualification, in addition to this storage foundation. Reassess #38 packaging and #39 traceability next; the full MVP remains incomplete.
2026-10-20 — W37 reproducible packages and installed inventories¶
Hypothesis: a reviewed source snapshot and fully pinned build environment can
produce identical wheel/sdist bytes, rebuild the same wheel from the sdist and
execute installed core/train contracts while rejecting inventory defects.
Baseline: 5e23f44ca51a58760ff1c139c114b7db52c5c1c5. A fresh baseline build
passed and reported Hatchling 1.32.4, but the backend constraint and transitive
build selections were not all exact. The existing package gate installed core
only and checked selected archive filenames. Mutable paths were packaging
configuration/scripts, CI, tests, inventory/licence evidence and related docs.
The criterion was repeated byte equality under fixed inputs, exact safe archive
inventories, real installed core/train checks, negative fixtures and full gates.
Saved source/environment hashes and raw numeric results
before this prose. All 58 new packaging tests and 16 training-gate regressions
passed; the full gate passed 1,331 tests, lint, formatting, typing, docs and the
unchanged 156-case W33 software conformance scope. make package-check passed
in 5.036 seconds and make package-check-train in 23.968 seconds on the recorded
local environment. These are command validation durations, not system performance
measurements. Each gate compares two builds and a wheel rebuilt from the sdist.
The two gates used source snapshot
73bbb8673857df59505fa944e0108f1505971204b2d1aec301d70c4fb445039b,
with the same wheel SHA-256
72c410bd17cc6075b6732b3f4399f10d0501c47430bef6e43abc9e04ba18ab3b
and sdist SHA-256
cf843cf16470b5baf2f07b9eff939324d87f86e519eb77e1adc0c8de6571bd19.
This is the captured candidate before this evidence/ledger entry; final-commit
CI rebuilds and records its own source/artifact identities. It is not a release
manifest. The six-package build environment and installed 10-package core and
17-package train environments include exact selected archive and notice hashes.
Core confirms scikit-learn is absent; train has no skip path and checks fit,
caller-pinned replay and altered-result rejection from the extracted sdist.
The first stricter archive check failed because Hatch included a parent
.gitignore in the sdist. That directory and its hashes/logs remain retained.
Empty VCS boundaries now stop ancestor exclusion discovery. A separate first
installed-train attempt exceeded 20 seconds and remains retained; an unchanged
diagnostic rerun passed. The final 45-second fixture timeout is within the existing
60-second per-command budget. Negative tests cover malformed/duplicate/traversal
members, links/nonregular files, size caps, source aliases, credentials, wheel
RECORDs, missing/extra files, unequal hashes and stale build/dependency review pins.
The dependency metadata review
matched all 518 locked artifact hashes for 32 packages against version-specific
PyPI metadata. mypy-extensions lacked licence fields; its locked universal wheel
was hash-verified and its MIT notice recorded. This is provenance and licence
inventory, not a legal opinion, vulnerability audit or release approval. Full
binary-wheel and bundled-component notices remain required for redistribution.
Decision 0016 fixes the gate;
PACKAGING gives commands and limits. Each invocation retains
fresh private local outputs, including failures. Recognizable credential patterns
are not a universal confidentiality detector. Build equality is scoped to recorded
environments, not arbitrary OS/Python combinations or an immutable OS image.
The repository stays public; LicenseRef-Proprietary and unpublished package
status remain. Legacy runtimes, fixture identities and the frozen corpus are unchanged.
W31 was merged in PR #65, with all six main jobs passing. W37 exact-head hosted delivery on the expanded Linux/macOS matrix is next, then
39 spec traceability. Admission, real proof/hardware evidence, independent review¶
and the complete MVP remain outstanding.
W37 hosted portability correction¶
Initial push CI
and PR CI failed
on e36e70f89a7b8827b9795c957eda0b02286b2ba5; the matrix cancelled remaining
jobs after failures. Linux rejected valid source reads because complete stat-record
comparison included access time. This is a failed hosted result, not a passed
platform qualification.
A local regression reproduced the problem: one access-time-only case failed while seven identity/modification cases correctly rejected. Saved the failure/source/log hashes and corrected validation before this entry. The guard now ignores access time and still compares device, inode, mode, links, uid/gid, size, mtime and ctime. All eight cases pass.
The corrected local gate passed 66 packaging tests, 82 focused tests and 1,339
total tests, plus unchanged lint/type/docs/conformance checks. Both package gates
passed again with identical repeated artifacts and sdist wheel rebuilds. The
corrected candidate source is
7a0ed2baa6c932cfd45668459dbd35793f99bd5df64a7a83d9e79246f1121743;
its wheel hash is unchanged and its sdist hash is
c2dce2adb0810d81ab3ab86349cecaad55d35a02a81172cea1778f0d0a07a6d1.
These candidate IDs precede this follow-up record/prose. The original evidence
and failed runs are preserved. Corrected exact-head Linux/macOS checks remain
the next delivery gate; no platform success is inferred from local validation.
2026-10-20 — W38 specification traceability¶
Selected #39 after W37 delivery. Hypothesis: a bounded offline gate can reject missing requirements, broken anchors and orphan issue mappings while preserving the historical packet; an explicit live query can establish current work/issue metadata agreement without giving normal CI network access. Mutable paths were scripts, focused tests, Makefile, the coverage register and maintainer docs. No runtime, fixture, model/data, conformance or old-evidence files changed.
The raw evidence was saved before
this entry. Against base 56b53f42cdc25f7690b6e8d1ed95a4e87b1e0bfb,
make check-spec exited 2 because the target was absent; the old doc checker
accepted a fixture with a missing same-file fragment. The new isolated tests
reject that fragment and missing/orphan requirement rows, issue mappings, cycles,
tracker drift, unknown coverage owners, malformed RFC sections/questions and
historical source changes. Fake API failures retain a failed result without raw
provider output, and a pre-existing result is preserved before any network call.
Focused validation passed 54 tests (53 new); make check passed 1,392 tests,
lock/lint/format/type checks, all local links and the frozen software corpus.
The spec gate covers 51 sections, 18 invariants, 40 work IDs, 10 RFCs and 106 edges.
Both make package-check and make package-check-train passed, including spec
validation from the extracted sdist and the existing installed contracts. The
candidate source ID is a0041cb17d54d01f6062520e372f3ecb65464a1681560da09d5cca5a4b2f5d41;
wheel SHA-256 is da20dbc02441fd0746687ff627669298ae8fb252f0b87302e9ab34b71384b2e1,
sdist SHA-256 is 4721e28e46276f1e8c0733678982bdcd50c933c646270a7f83ad3572a761456d.
The hashes identify the candidate before evidence/prose; final-commit CI rebuilds it.
The live audit, produced by
uv run --locked python scripts/check_spec.py --live --output research/evidence/w38-live-issues-2026-10-20.json,
matched all 50 work/tracker issues: actual titles, scopes, acceptance criteria,
labels, milestones, forward dependency edges and tracker children. It records
observation time, source/response/body hashes and issue states. It is a dated
observation, not a permanent GitHub guarantee. The earlier successful live audit
remains local; no failed scientific or hosted outcome is hidden by these checks.
Decision 0017 defines the limits. The checker supports the documented repository Markdown forms; structural coverage cannot establish that an implementation satisfies the requirement's meaning. Owner questions remain open and all RFCs remain Draft. Public visibility persists; licensing, data/policy admission, hardware/proofs and independent review remain separate.
W37's main CI passed all 12 jobs after PR #66. W38 exact-head hosted delivery is next. Afterward reassess the full remaining backlog; the broader MVP is not complete merely because its local foundation is green.
W38 parser review correction and final local candidate¶
Review found nested repetition in the new Markdown target regex. An isolated 10,014-byte unfinished-link probe exceeded a one-second limit and was terminated; the probe and candidate source hashes remain retained. Removing the nested quantifier makes that input plain text without exponential backtracking. Added a regression; final validation records 55 focused tests (54 new), 1,393 total tests and both package gates passing. The prior candidate evidence above is preserved.
The final candidate source is
f29c9340099fa787e7f8380888f33d0e738b3e54f9af1a392ab524040188a458;
its wheel hash is unchanged and its sdist hash is
0f141689018b583aa8487cfc2b508f61955e1f65285a24432e920c3f614ca46b.
These identities precede this follow-up record/prose. A fresh explicit live audit
matched exactly the original 50-issue metadata inventory and response hash;
the follow-up records its own time and source hashes. Hosted final-head validation
is still required before merge. No historical packet, corpus or prior run changed.
W38 delivered¶
PR #67 merged as
1da52da9c5f45b481d031ab3dd1ec5b4601c7a6d. Push run 39440344222,
PR run 39440355607 and main run 39440839286
each passed all 12 jobs. Issue #39 is closed; the release tracker remains open.
2026-10-20 — W13 admission preparation, approval outstanding¶
Selected the explicitly permitted proposal portion of #14. Hypothesis: strict study terms and a caller-pinned review dossier can reject missing/changed declarations while a separate concrete proposal preserves every unresolved owner decision. Baseline is W38's 1,393 passing tests. Mutable paths are the new admission module, fixtures/schemas/tests, package/check entry points, proposal and related maintainer documentation. No data/model asset, historical experiment or frozen conformance corpus changed.
The raw validation record
was saved before this entry. uv run --locked pytest tests/test_admission.py -q
passed 81 tests; make check passed 1,474 tests plus lock/lint/format/types,
spec traceability, the existing software corpus and the new admission fixture.
The builder reproduced the fixture and three schemas byte for byte in a fresh
directory. Negatives cover absent/wrong external pins, mutated nested terms,
missing role/source/code/config documents, exact file inventories, invalid caps,
ambiguous decimals, unsafe constructed records and the actual unapproved proposal.
The initial make package-check failed: its tracked-only source snapshot omitted
new unstaged files, and extracted-sdist link validation rejected the missing
proposal/decision. Diagnostics remain in dist/package-check-tfkme4t0; the evidence
retains the failed command and output hash. Staging the reviewed new paths fixed
the candidate inventory without changing the gate. Both make package-check and
make package-check-train then passed, including repeated build equality, fresh
installed core admission checks and installed synthetic training/replay negatives.
Their source ID is 9518789db603e79363316833eadcbe686201b65ffd535e65ffa453568dc20392;
wheel hash d8d481d1eacefeb8d6f9ee92eb3134d92ac2e5265874638aae95d2be76876cde,
sdist hash ad5fa5b0da0785892da8df7ebeecca57af040e64120be568074391aab4ac1fe7.
These identify the candidate before this evidence/prose; final-commit CI rebuilds.
The public metadata record retains exact metadata/card/config/notice hashes and advertised asset inventories. No dataset, checkpoint or tokenizer asset was downloaded. Public dataset previews were exposed during page inspection, so the proposed source is exploratory, not blind confirmatory evidence. Licence metadata conflicts, provenance, source overlap and all access approvals remain holds. No upstream performance claim was adopted. The power calculation records source/environment and conditional binomial results, including 299-group zero-event power; it is not an approved sampling or quality plan.
Decision 0018 proposes a pre-training StudyPolicy and acyclic review terms. The check compares exact declarations; it does not authenticate reviewers, verify actual assets, enforce resource limits or establish scientific/legal adequacy. The marker fixture explicitly grants no measurement authority. The actual proposal retains null quality/power/budget and reviewer choices. W13 remains open and blocked for owner/relying-party review; a draft PR is the next delivery step. No prior research result is reused and no full MVP, hardware or proof claim is made.
W13 preparation delivered¶
PR #68 merged as
346f9bcc1940985c90053a719eb2ed4fea4671b6. Push run 39444024829, PR run
39444100293 and main run 39452888007
each passed all 12 jobs. This delivery did not itself adopt the study.
2026-10-20 — W13 owner approval and delegated study design¶
The owner approved the English single-turn study and its quality targets, acting as prototype decision owner; delegated evaluation design and licence-reviewed public source/software selection; approved the local four-thread/16-GiB/20-GiB, two-hour/job, six-hour/$0 envelope; and authorized conditional HF GPU Jobs, preferring the Mac. The owner requested hardware/reviewer options only. Decision 0019 normalizes this authority without publishing a private transcript or credential.
Hypothesis: the current authority and proposal can agree on the approved scope without admitting unresolved sources or an incomplete measurement dossier. Baseline: 1,474 tests at the PR #68 merge. Mutable paths: authority, proposal, agent workflow, study/resource docs, the power calculation and admission tests. No runtime contract, marker fixture, historical experiment, dataset or model changed. Pass criterion: coherent authority/budget/roles, reproducible power calculation, continued refusal of the incomplete protocol, and local/package gates.
The raw validation record precedes
this entry. make check passed 1,476 tests plus lock/lint/format/types, spec,
software conformance and admission gates. The first focused lint check caught
one overlong assertion; formatting fixed it before the passing gate.
The two new tests compare approval/budget/role records and reproduce the fixed
power script with independent exact-tail/confidence-bound cross-checks.
make check-train required and passed both training-marked tests.
make package-check passed repeated build equality, installed-wheel checks and
extracted-sdist validation. Its source ID is
5d6697994c9f7fc4ab1c6ac43d6c88eed6caa54900bf04a14ccfe42d1c38dd26,
wheel SHA-256 d8d481d1eacefeb8d6f9ee92eb3134d92ac2e5265874638aae95d2be76876cde,
sdist SHA-256 77478a534c5dc8dc83916c0b2fb941410931168919b6789de71fb43914e92ea7.
The wheel is unchanged; these identities precede this evidence/ledger addition.
Final-commit CI rebuilds the final source.
The power evidence records code and lock hashes, environment, exact cutoffs and conditional probabilities. With 15 Bonferroni-controlled one-sided bounds and family power at least 0.90 under the specified alternatives, the design requires 7,831 independent benign groups and 1,375 positive groups. Those are planning quantities, not measured model quality. Unknown lineage, label audit, actual asset/split/candidate/evaluator identities and the final enrolled review pin remain concrete measurement dependencies.
The options evidence records local hardware and unauthenticated public provider metadata, including the HF price inventory. Reviewer-site direct HTTP attempts failed; the browser reader supplied the cited public pages. The options brief compares local/HF research, Azure/GCP genuine hardware and independent review. No account pricing/quota/credit state was checked, no paid job launched, and no reviewer contacted. Existing source licence holds are preserved. The old fixture-only authority is superseded; W10 implementation can proceed.
W13 approval delivered¶
PR #69 merged as
7f547356a7f91e8d331a6963d0cf40603dd49b44. Final head
af14a223c1060bdbc8c21d703291478c58e5dccb passed all 12 jobs in each of push
39456548720 and PR 39456559150; main run 39457001526
passed all 12 jobs on the merge. Issue #14 is closed and its approval blocker
removed. The repository remains public.
2026-10-20 — W10 complete-byte literal rule¶
Hypothesis: bounded suffix state can reproduce an independent complete-byte literal oracle at every chunk, while exact manifests and file hashes reject substitution. Baseline: 1,476 tests at the PR #69 merge. Mutable paths: native rule records/runtime, fixture builder/checker, schemas, tests, package checks and related API/decision/research documentation. Existing contracts, historical experiments and the frozen conformance corpus are unchanged.
The raw evidence precedes this
entry. uv run --locked pytest tests/test_rule.py -q passed 96 tests; make check
passed 1,572 tests plus lock/lint/format/types, spec, software conformance and
admission checks. The initial 95-test pass was extended with a regression for
short matches retained inside a longer pattern's suffix and partial UTF-8 byte
patterns; the complete gate was rerun successfully.
The independent oracle enumerates complete-input byte slices and match-ending positions. One thousand fixed-seed generated cases compare every chunk; the maximum-size test places a 4,096-byte pattern at the end of a 32,768-byte prompt, using one-byte chunks. Eleven fixed vectors reproduced byte for byte in a fresh directory. No pattern is inferred from private data or a downloaded source. The fixture's hashed provenance identifies hand-authored harmless markers under the repository licence; it is not an operational injection-policy approval.
Decision 0020 fixes native matching, aggregation and bounds. Provenance and parameters are byte-exact, input coverage is complete, equality blocks and invalid/unsafe constructed records reject. The scalar ReLU graph only projects the match feature. A proof of that graph alone would omit the literal extractor and cross-chunk state, and cannot satisfy W19. No learned quality, external enrollment, hardware or proof claim is made.
Both make package-check and make package-check-train passed repeated builds,
fresh installed rule vectors and the existing core/train checks. Their common
source identity is d55ca53816123a94a81aaf92583e6c3b7831cbabcbd07937bc5b54f060800c7a;
wheel SHA-256 650b58c23b14374caddaa9cf8fa85d8155efd58f1c9ba5243d6f3ee854a4ce5b,
sdist SHA-256 e7b114c57fc297d5518fc0fe6a12df0f9674298ecdc4859cb2a13bfe3da1d75d.
These identify the candidate before this final evidence/prose addition;
final-head hosted CI rebuilds it. The unrelated .gitignore edit is preserved.
Review then caught a JSON schema mismatch: the default bytes schema described
binary strings using byte lengths, but the actual JSON encoding is hex. The
pattern schema now admits exactly even-length hex through 8,192 characters.
A maximum-pattern JSON roundtrip and odd-hex rejection test bring the final
focused count to 97 and total to 1,573. The evidence preserves the earlier passes
and the new full check and both package passes. Final pre-prose source identity:
1f4d4e8913de56fca3c3911c6322574fce372e75165618ad6edfa6110d004013;
wheel c71824e2fc237e9851bbbe299fe30fde5ccb3909123b80f66947ed5ad5fad402,
sdist 78383daf2085efb9cc7d11d9238a871a2ec141f6aac40e2a2fff5f50a8c9a50d.
Runtime/fixture wire identities remain unchanged by this schema correction.
W10 delivered¶
PR #70 merged as
fa310ea2a4d1d484f1ef325eb7bfbf5f86a280ed. Final head
19541ba307b456ce295545bd2740debbee69f6a1 passed all 24 push/PR jobs;
main run 39459269341
passed all 12 jobs. Local, tracking and remote main matched the merge commit.
Issue #11 is closed; #15 and #19 had stale dependency blockers removed.
2026-10-20 — W14 source admission and lineage-first split¶
Hypothesis: externally pinned source/review bytes and complete connected groups
can produce reproducible frozen splits while rejecting substitutions and
cross-split descendants. Baseline: 1,573 tests at the W10 merge above.
Mutable paths: the additive import API, fixtures/checker, schemas/tests,
package inventories and related decision/API/research documentation. The
unrelated .gitignore edit remains unstaged.
The raw evidence precedes
this entry. Initial validation passed 55 focused tests, 1,628 total tests and
both package gates. Final review removed repeated held-out-set construction
from the per-group loop and added UTF-8 byte-boundary and policy-mismatch cases.
The final focused command passed 59 tests; make check passed 1,632 tests plus
lock/lint/format/types, spec, software conformance and admission checks.
All recorded initial/final runs passed. Development lint/type corrections
preceded these retained runs.
Twenty-five fixed-seed parent graphs match an independently computed reachability oracle. Split boundaries/group representatives match independent arithmetic, and changing labels/input ordering leaves assignments unchanged. The 26-row fixture has 21 groups: 11 train, five calibration and five test. Parents, cross-source descendants, shared histories and identical/ASCII-equivalent prompts stay together; contrasting labels remain separate records. Held-out source/family members move their complete component into the separate held-out cohort. The fixture reproduced byte for byte in a fresh directory; all seven schemas match their generated definitions.
Source/licence/review changes, malformed JSON/UTF-8, missing parents, cycles, unknown labels, policy mismatches, stale near-review inventories and rehashed split/lineage mutations reject. Four injected sync failures retain partial output without a final manifest; existing/symlink targets cannot be overwritten. These checks do not establish physical power-loss or hostile-parent-path safety.
Final make package-check and make package-check-train both passed repeated
distributions, sdist rebuild and installed import fixtures. Their common source
identity is 8f8a6b09485619d95ff7ad5de62b98a464c8c441b483566464de6bd67e42f4c2;
wheel SHA-256 f7b93027de33840ce2d0ff3e2a6eb3574053d5557ccd1abf684a4a2b2c7c14e0,
sdist dd608f800f8f426c033c1ae0fd92a1cd9d083a182e21c3b6e6df59816fb1f718.
These identify the candidate before this final evidence/prose addition;
final-head hosted CI rebuilds the complete commit.
Decision 0021 records limits, exact identities, review/source-card gaps and trust boundaries. Every source and review is synthetic; mock public-mode tests bind declarations without authenticating a human. No real dataset, checkpoint, external inference or paid job was used. Lexical candidates cover ASCII case/whitespace only; reviewed semantic lineage and final measurement admission remain necessary. Historical results and the legacy 8 MiB loader are unchanged.
W14 delivered¶
PR #71 merged as
c4586a7107c59fb7ba4665247e73ff7e0146cf7d. Final head
0bc7f3c8cf4c13b11f6633ee6193a832a2f2628b passed all 24 push/PR jobs;
main run 39466963984
passed all 12 jobs. Local, tracking and remote main matched the merge commit.
Issue #15 is closed. The repository remains public.
2026-10-20 — W15 record-level MLP component¶
Hypothesis and boundary are recorded in decision 0027: complete-input record-max fitting can resume exactly and export the existing native integer runtime without exposing calibration/test labels to fitting. Baseline: 1,632 tests at the W14 merge above, synthetic native MLP weights, and the unchanged logistic trainer's multi-chunk rejection. Mutable paths are the additive training/export API, shared fixture assembly, schemas/tests, package checks and documentation. No dataset, evaluator, policy threshold or historical result was retuned.
The machine-readable evidence
was saved before this prose. Raw local artifacts are retained at
runs/w15-mlp-training-1792527541853491000 with the protocol, source snapshots,
environment, command logs, optimizer checkpoints, epoch traces and exports.
Protocol SHA-256:
406c6d18fd70796b17246460acff8986d7e9c4fb8470411f88720211489a1d58;
receipt SHA-256:
5bdeb8c5a3428235fec47fd0c177ff79b890dae59fab6848b49b56de645df5cd.
The pre-run source inventory identity is
fca0528fd4eab344ed62315d7df070f9d76dc542407c695881e805a79546ed1e.
Two new directories reproduced all 27 substantive artifacts byte for byte. Each run fits the frozen 12 authored marker records for seeds 17/29/43, compares fresh fitting with a JSON checkpoint resumed after epoch 31, and executes seven native/float probes per seed. Probes include empty, Unicode, chunk boundaries, last-byte markers and the full 32,768-byte limit. Maximum absolute score errors were 0.0064575824, 0.0081562375 and 0.0161565061, below the predeclared 0.1 engineering tolerance. Final losses were 0.3366060358, 0.4234531054 and 0.3295476184. All three reached epoch 80 without convergence; none is reported as converged or as a successful policy detector.
The two sequential one-thread commands took 0.671 and 0.371 seconds, with measured peak RSS 57,458,688 and 58,474,496 bytes. Source and raw evidence occupied 3,262,607 bytes before the receipt. Cost was $0. NumPy 2.5.4 / CPython 3.12.12 ran on the owner Mac; exact CPU features, extension and trainer hashes are in the environment record. No dataset/checkpoint download, external inference, GPU job, HF token access or private upload occurred. Timings are single fixture observations, not a performance distribution or proof-cost claim.
The 108 focused tests passed, including 75 existing native-MLP cases and 33 new training/export checks. Numerical finite differences check the record-max subgradient away from ties; a separate tie case verifies first-max routing. Pre-import rejection covers nontraining rows, changed identities, invalid profiles, malformed shapes and resource limits. Checkpoint substitution, changed configuration/data/runtime/loss, CLI path disclosure and failed output retention are checked. The historical MLP fixture still reproduces unchanged.
make check passed all 1,665 tests plus lock/lint/format/types, spec, software
conformance and admission. make check-train executed all 17 selected tests.
Both package gates passed, including core-only contract checks and actual
installed-wheel three-seed training. Their shared pre-evidence source identity is
d12a88046b5cc57d36db75c1c04831bc6c87daaf166e92b6753d86f11b6eb6eb;
wheel 1fa14f03212f2154143e3963d1969b2017b890a11034b8025dd88e1db68c4627,
sdist a2c789a76539e27868ddb0887d491387b945a9ee55cbef6f2a4bbd345495cc89.
Gate commands consumed another 42.942 seconds. Final-head hosted CI rebuilds the
complete commit including this evidence and prose.
All recorded runs passed; earlier development checks rejected a fractional native
output. The correction preserves the integral-score contract through explicit
score units instead of relaxing validation. Epoch-limit nonconvergence is retained.
These are shared-implementation software checks, not independent authentication,
policy quality, hardware provenance or a proof. The complete W15 roster, grouped
selection, MLP calibration/evaluation and admitted real-data commands remain open.
W18's staged proof work stays separate in draft PR #72. The unrelated .gitignore
edit remains unstaged, and repository visibility remains public.
2026-10-20 — W15 grouped comparisons and native calibration¶
Hypothesis: finite, train-only grouped selection can preserve split isolation and
every candidate failure, followed by separately pinned native calibration and
evaluation. Baseline: 865691f0fe455045f24543f0f5de333a5a2dae5f in draft PR #73,
with 1,665 tests and MLP fitting/resume but no grouped comparison pipeline.
Mutable paths: additive baseline/selection/calibration APIs, fixture/checker,
schemas/tests, package checks and documentation. The historical logistic trainer,
evaluator, datasets and previous MLP experiment remain unchanged.
Decision 0028 preceded the run.
The numeric evidence was saved
before this prose. Raw snapshots/logs/fit states are retained under
runs/w15-grouped-selection-1792528940558781000. Protocol SHA-256:
7974d63824ad53ee0137f9f163a365e3c8b12baebccd71f5eab2b7f287a09285;
receipt 8240caf7da59550237723e2f6c5ce747cce5e5108fbd4cf3973747df430e16a2;
pre-run source inventory
56f5a904f868dae836b73cf2be8f7a849d550a035e29a725cc14cb41e082c97d.
The frozen fixture has 100 train rows in 50 groups, plus 20 calibration and 20 test rows in disjoint groups. Five declared candidates and three seeds produce 75 rotations: three folds fit, one selects an internal threshold, and one validates. Only fitting rows supply vocabulary or optimizer inputs. Group and representative assignment are label-independent. Exact rational ranking selects per family and seed, with lexical candidate ties. Controls never fit/calibrate. Every nonconverged or failed candidate remains in the report.
Two fresh one-thread runs reproduced all 31 substantive artifacts byte for byte. Each retains 15 deliberately nonconverged MLP rotations and refits/calibrates/ evaluates nine selected models. All nine primary test reports contain ten group representatives; all-row reports contain 20 descendants. The authored-marker logistic and character models each have TP=5, FP=0, TN=5, FN=0. The deliberately loose-tolerance MLP has TP=0 and FN=5 for every seed; FP counts are 0/1/2 and TN counts 5/4/3. These are unacceptable as task-quality results, not hidden by the software pipeline's pass. Neither MLP fixture configuration is a real study candidate, and no result is an injection-policy estimate.
Recorded command times were 2.282 and 2.165 seconds; macOS time -l peak RSS was
142,442,496 and 141,213,696 bytes. Source/raw files occupied 8,016,420 bytes before
the receipt. Cost was $0; no model/dataset download, private upload, HF token or
paid job was used. These two fixture observations do not establish latency
percentiles or production throughput.
All 44 focused tests pass. They check outer-label rejection before serialization or optional imports, group/representative invariance, fitting-only vocabularies, exact decimal allowances and blocking ties, resource limits, malformed numeric types, preserved long-input floor rejection, explicit liblinear nonconvergence, candidate ties, wrong fit/calibration pins and overlapping groups/prompts. A development check found that numeric literals accepted float-valued seeds/folds; strict integer fields and regression cases corrected that before this protocol. The final fixture/schema snapshots reproduce, and injected controller failure retains an incomplete directory without a final result.
make check passed 1,709 tests plus lock/lint/format/types, spec, software
conformance and admission. make check-train executed all 35 selected tests.
Both package gates passed, including the core-only contract path and actual
installed three-seed grouped comparisons. Their common pre-prose source identity
is d9b7a89726fb7b24624c778438c713197ef0b8b1049e5ad5ea0bfad39bafa2b8;
wheel 5754530a3759c13a35653480bc599d4c63e1be27a5f2434565aac769d3896326,
sdist 5dea06cf58fa65aeff0f8c2be01e88a50c97cd0990e3269d69ed0fc9db501b7c.
These gates consumed another 58.399 seconds; final-head hosted checks rebuild the
complete source including this evidence and prose.
The separate guard source review
retains current public metadata and base-model notices. Under delegated selection,
decision 0029 chooses the pinned
published guard for bounded qualification using its explicit model grants. It
does not approve acquiring unnamed unlicensed upstream training data, endorse
the archived project's metrics, or resolve contamination. The local raw metadata
is in runs/w15-guard-metadata-1792528650785139000; no weights/tokenizer or runtime
packages were acquired. Exact runtime review, asset hashes, complete-input guard
execution and final real-data admission remain the next W15 gates. This PR stays
draft; W18's proof work stays separate. The unrelated .gitignore edit is preserved.
2026-10-20: W15 complete-input CPU guard and optional runtime¶
Hypothesis: the selected pinned guard can score every content-token block of a 32,768-byte prompt on the Mac and reproduce its numeric outputs in a fresh process. Baseline: metadata selection only; the publisher's example truncates. Decision 0030 freezes the complete-input max-margin adaptation before real inference. Mutable paths are the optional adapter, its comparison binding, tiny runtime fixture, packaging, tests and documentation. The previous learned baselines and evaluator are unchanged.
The dependency review matches 684 locked artifacts for 52 packages. Existing versions are unchanged. It records selected-wheel notice hashes and exact upstream Apache notices for the two wheels missing licence files. Acquisition verifies all five advertised model/tokenizer digests, including the 738,563,188-byte ONNX graph. Model/base notices accompany the ignored assets. No remote model code, pickle, HF token or paid job was used.
The CPU evidence retains raw scores,
source/runtime/binary hashes, exact asset/spec/input identities and receipts.
Raw model acquisition is runs/w15-guard-assets-1792529768485391000; execution is
runs/w15-guard-cpu-1792530020726182000. Its protocol hash is
30e2e46c066c526ad4abf9a71278cee6445bfcc1afee486471a517205ca1bce1, receipt
5113ada479ae8fa5267d64d5c0a09f6e6b4864376cc38c2f3b80f84aacd8e809.
Two four-thread CPU processes each score seven fixed harmless probes through 50
chunks and reproduce all 11 substantive JSON files byte for byte. Times are
7.253/7.008 seconds; peak RSS is 4,372,217,856/4,371,480,576 bytes. The maximum
ASCII prompt has 5,463 tokens/11 chunks; the maximum Unicode prompt has 16,384
tokens/33 chunks. Both are exactly 32,768 bytes. The late instruction is scored
in its second chunk. Costs remain $0, with no external inference.
Negative evidence remains explicit: the two maximum-length repetitive benign
probes have margins +8.170272/+6.674934. Zero is not an admitted threshold, and
these seven authored prompts cannot estimate a false-positive rate. The published
tokenizer strips/normalizes text; full token coverage is not byte preservation.
Upstream contamination, human label review, task-quality targets and independent
verification remain unresolved. bind_guard adds the fixed pretrained model to
the separate calibration/evaluation API without local fitting or test selection.
Validation: 32 new guard tests; make check passes 1,741 tests; make check-train
executes all 35 required tests; make check-guard executes the real tiny ONNX
fixture and all 32 tests. Core packaging and a combined train/guard package gate
pass through separate installed environments and reproduce identical distributions.
Their pre-prose source identity is
76fb4bbbc68ec8b7217d494d5800a2cff9135cf1b792068a5d2095b8b30e1cf6;
wheel 2d5a12bb022de94eeeb5f2a8e8498887ebb2f76bffa0c9e9d4f3ffdc657c69ef,
sdist 3f72ce130a180ff8a749e233d49d1250feb949255bad98ed8c8b4ca4025dd52d.
The five passing gates consume 67.287 seconds. Raw logs/receipts are in
runs/w15-guard-gates-1792530308409104000. The earlier failed full gate is kept:
its only failure was the old statistical calculation's lock pin. A separately
retained calculation reproduces every field except the new lock SHA; the original
W13 record is untouched. Runtime tests also reject changed files/pins, links,
external tensors, custom domains, oversized/invalid Unicode inputs, bad outputs,
held-out rows, overlapping groups/prompts and existing output directories.
PR #73 remains draft until admitted real-data commands satisfy the rest of W15.
The compact encoder and full proof/hardware obligations remain separate work.
The repository remains public and the unrelated .gitignore edit is preserved.
2026-10-20 — W15 caller-enrolled comparison command (partial issue #16)¶
Decision 0031 fixes a fresh engineering experiment joining admission, reproduced source import, grouped selection, final fits and calibration before test. Existing evaluator, grouping, study targets and historical results remain unchanged. The command checks an externally supplied review identity, exact source/config/runtime pins and dataset reproduction; it does not authenticate a human approval. Fixture-only mode cannot relabel authored data as real study evidence.
The raw numeric evidence points
to runs/w15-comparison-command-1792531617672623000. Protocol
c912d278451562cc2f15abc1b7191e2bc31dbdc7b9dc7c7ef931fb0c4a1dd6f2
was written before measurement with exact source/input identities, two commands,
55 seconds per command, one thread, 2 GiB memory and 16 MiB new storage. Two fresh
processes took 2.931/2.892 seconds and peaked at 272,400,384/270,286,848 bytes RSS.
All 38 output artifacts match byte for byte. The full 200-row authored source,
invented review dossier and tiny ONNX graph remain local. No downloaded data,
HF token, GPU, paid compute or external inference was used in this experiment.
The runner retains 75 train-internal rotations, including 15 deliberate
nonconvergences, then nine selected final fits plus the fixed marker guard.
Every fit/calibration identity is saved before scoring test. Negative results
remain: each MLP seed misses all 20 positives; each character baseline has 2/20
benign false positives. Logistic/marker-guard fixture results are not scientific
quality evidence. Controls report their full errors; compact encoder is explicitly
unavailable. full_study_complete and quality_claim remain false.
Validation: make check passes 1,772 tests, make check-train executes 35 tests,
make check-guard executes its graph plus 32 tests, make check-comparison
executes the real CLI plus 31 binding/order/failure tests, and make package-check
passes. The five gates total 63.908 seconds. The existing fork-rejection test
emits Python's multithreaded-fork deprecation warning and passes. Raw logs are in
the experiment's gates/ directory. Core package source identity is
49bd4b6098c32ec95579924492abc3b442f40054245c3817147678fb10286ac5, wheel
a8929702d6cb278097b1c62e7cf93b60c1af6ff081af0d18d178da0efa942b3f, sdist
ded0fad4b12738742daa23a9d9c0bd66398299424fa359f55db011b67f3ec733.
These distributions precede this evidence prose. Another local combined-package
environment would exceed the remaining 20 GiB allowance, so it was not run;
Linux/macOS hosted jobs now require the installed combined comparison fixture.
The storage-choice request remains pending and old evidence was not removed.
PR #73 stays draft: the real-data command is implemented, but an actual reviewed
source/label/contamination dossier is still missing. docs/OPERATIONS.md provides
the exact invocation template, required layout, external-pin and resource
obligations. No real-data or final study claim is made by this increment.
The first hosted Linux comparison check on 5027e50 failed one test after 30
passed: ONNX Runtime emitted a native device-discovery warning while the command
correctly returned partial-result exit 3. The test had incorrectly required an
empty diagnostic stream. The correction checks the exact stdout status and
absence of private input paths, retaining exit-code, missing-family, convergence
and sanitized-exception checks. Execution code and measured source/input/output
identities are unchanged. The failed
job
log is retained in the raw experiment directory; its failure is not reported as
a passing package check. All 31 corrected comparison tests pass locally in
10.54 seconds (portability-tests.log); hosted jobs must pass on the corrected
commit.
2026-10-20 — W15 public-source review and human audit handoff¶
Decision 0032 bounds source
preparation under the existing delegation. The
evidence retains exact metadata,
licence notices, acquisition plans/receipts and local raw pointers in
runs/w15-source-review-1792532338623341000. No source is admitted for measurement.
Gandalf revision 04737b65e90a6794ec227012e4a255a7def6344b yields all 1,000 rows
through the public viewer API. Every page reports that exact x-revision, the
expected row count/indices and no truncation. This checks retrieved snapshots,
not local Parquet decoding. Dolly revision
bdd27f4d94b9c1f951818a7da7fd7aeea5dbff1a verifies the 13,085,339-byte JSONL against
2df9083338b4abd6bceb5635764dab5d833b393b55759dffb0959b6fcbf794ec.
The predeclared empty-context filter retains 10,544 exact instructions and excludes
4,467 context-dependent records. Acquisition totals 13,248,844 bytes in 4.555
seconds. Source grants are MIT and CC BY-SA 3.0; copyright, attribution, licence
and change notices accompany the local review adaptation. No raw rows are added
to Git.
The separate Mirza Akhi source at 011a9ae25b31feb48d23f837fdcbc3a6a4ee3b4e
has an explicit CC BY 4.0 grant. Its acquired CSV matches
eb748eb8104f0a736934bf4f3c96335801010b09e17467fd4aadc1c1ae7bbced:
5,441 unique prompts, 5,425 marked English, maximum 223 bytes. The generation
recipe was inspected as text, never executed. Template ancestry and the original
holdout need review; it is a possible supplemental stress set, not real-world
validation. The Qualifire successor is auto-gated and was excluded without token
access. PINT's full corpus remains held for task/provenance/rights review. A direct
Creative Commons text download returned HTTP 403; the official browser-reader
responses and repository notices are retained, without claiming raw HTTP-byte
equivalence for those browser snapshots.
Two fresh review builds reproduce all six files byte for byte. The 64-item packet
SHA-256 is 0ec4a6951b1939474aa70eb89b7b0228890ef71dec93846bbd63d6de51e2a57e.
It hides proposed labels/source class per item and includes complete prompts,
the frozen rubric, attribution and blank human response fields. A separate
editable human-response/labels.json preserves the two original templates.
No human identity, labels or agreement rate were invented. The owner was asked
who will perform the review; no reviewer was contacted.
Preflight finds 114 exact-duplicate groups. A fixed, non-fitted character-hashing screen retains all 563 pairs at cosine at least 0.90 and 172 multirow components; the largest has 33 rows. One two-row component crosses Gandalf's original splits. The tiny equality/unrelated-text fixture passes first. The full screen takes 1.022 seconds in-script (1.80 seconds process wall time), with 216,514,560 bytes peak RSS. Pair/component files, code/dependency pins and the cross-split review example are retained. The screen proposes links; it cannot establish semantic or historical lineage, and no links are represented as human-approved.
The complete study remains unmet: source identity is confounded with provisional class, language/label decisions and parent identities need review, and 1,000 total Gandalf rows cannot supply 1,375 independent positive test groups. These exposed public sources are exploratory. No acquired row was used for classifier fitting, guard inference or test scoring. No paid job or HF token access occurred. Frozen targets and old results are unchanged. PR #73 stays draft pending genuine review and final admission.
Storage accounting was corrected to include the whole checkout, Git and caches;
the earlier three-directory total omitted those bytes and the checkout was about
0.9 MiB above the recorded 20 GiB cap. Only mypy's tagged, rebuildable SQLite cache
was removed, restoring headroom. storage-correction.json records exact before/
after bytes; no research run, failure, environment, asset or review artifact was
removed or archived. Subsequent local validation disables mypy cache writes.
The request for a larger retained-storage envelope remains pending.
Validation with MYPY_CACHE_DIR=/dev/null make check passes all 1,772 existing
tests, strict type/style checks and the spec/admission/conformance gates in
32.696 seconds. The existing multithreaded-fork deprecation warning remains.
Source-card Git blob checks, all acquisition file hashes, canonical candidate
pairs/components and both blank packet builds also pass. Classifier source and
dependency locks are unchanged. Distribution checks run in hosted CI; another
retained local package environment does not fit the remaining storage allowance.
2026-10-21 — Owner-approved synthetic W15 pipeline and storage¶
The owner approved the requested 40 GiB retained-storage envelope, agent review and wholly synthetic data without human review for pipeline development, with the Mac preferred and HF T4/L4 allowed fallbacks. Decision 0033 records this change. Historical natural-source packets, failures and evidence remain unchanged. The W15 work-map/issue criterion now explicitly permits the authorized synthetic pipeline, retaining confirmatory real-data admission as a separate quality gate. No human review identity or agreement rate is fabricated.
Hypothesis: the existing comparison command runs unchanged with the real pinned published guard and the 200 authored controls. Baseline: its tiny-marker guard fixture and the earlier standalone real-guard execution. Mutable paths are the new local dossier/protocol, approval records, documentation and authority checks. No classifier, evaluator, dependency or runtime source changed. Pass criterion: complete W15 execution with every baseline/control retained, exact reproduction, package gates and truthful unavailable/quality status.
Numeric evidence retains
runs/w15-synthetic-pipeline-1792567226626052000: frozen outer protocol, both
controller sources/pins, agent review, complete dossier inventory, caller pin,
commands, results, resource receipts and validation logs. The agent checked all
200 control labels against the recipe (hello=0, ALERT=1). Shared templates
make these unsuitable independent statistical groups or injection-policy labels.
The internal fixture source markers remain mechanical binding inputs; actual
run authority is this owner's decision, not invented fixture approvals.
The externally retained review pin is
bcc2f807bc7bf6849d9cf04c30c7b41b2bacc3ac6fcfd28243b4896fecab079b.
Both actual CLI invocations use --fixture-only, the frozen dossier and existing
pinned guard assets; exact argv records are in the evidence. Each fresh process
exits 0 with all W15 baseline families and fixed controls executed. All 38 files
reproduce byte for byte. Elapsed times are 4.674/4.665 seconds; peak RSS is
4,373,692,416/4,373,921,792 bytes. Preparation itself was not separately RSS-timed.
The outer watchdog limits execution to 300 seconds; resource observations are
not an OS memory/storage sandbox.
Negative results remain visible: the real guard marks 3/20 negative controls, character n-grams 2/20 per seed, and each MLP misses all 20 positive controls. These are fixture-control outcomes, not FPR/TPR estimates for injection policy. The full study, quality and independent-verification flags remain false; compact encoder is explicitly W16-not-executed. This completes W15 software execution under the owner's revised stage, not the entire MVP or confirmatory study.
Validation: 114 focused admission/comparison tests; make check passes 1,772
with the existing fork warning, strict typing/style, 156 software conformance
cases and specification/admission checks. make check-train executes all 35
training tests. Core and combined train/guard package gates both pass; the latter
runs the installed comparison command. Their pre-evidence-update source identity
is 45e564fabc36a7f0256a45fee9213b235d399767e81bdc2b840190bbbebf4d4c.
The live spec audit matches all 40 work items and 106 dependency edges after the
owner-approved W15 acceptance update. Final-head hosted CI remains required
before merge. No HF token, GPU job, paid compute or reviewer contact occurred.
2026-10-21 — W17 frozen group-aware reporting¶
Hypothesis: existing frozen test predictions can support versioned slice and paired root-group reports without training, scoring or threshold selection. Baseline: W15 has separate per-model counts but no shared reporting adapter. Decision 0034 freezes the report semantics; mutable paths are report API/command, fixture/tests, package checks and documentation. Prior evaluator/data/threshold identities remain unchanged.
Raw numeric evidence retains
runs/w17-group-report-1792568480677539000: predeclared implementation/execution
plans, the preparation controller and pin, verified prior artifact inventories,
source/provenance/observation inputs, separately retained plan pin, exact commands,
resource receipts and byte-equal reports. The new synthetic experiment explicitly
records prior exposure to W15 test outputs. It does not reuse them as unseen
confirmatory evidence. All 38 original run artifacts and the entire source dossier
were checked against their existing inventories before preparing report inputs.
The correlated software fixture has 36 authored-score rows in 12 declared groups, three seeds, three paired 10,000-resample comparisons, one positive quantization regression and one benign false positive per seed. Existing calibrator tests check ties and no benign calibration; new checks reject changed labels/thresholds, missing seeds/rows, split overlap, pin changes and adaptive confirmatory requests. Counts for correlated descendants have no binomial intervals. Empty denominators and absent bootstrap denominators remain null. Missing attempts remain explicit.
Two new CLI processes report the retained W15 outputs in 0.292/0.283 seconds, peaking at 41,697,280/41,779,200 bytes RSS. Both report files are byte-identical. Ten available members and nine paired comparisons execute; twelve absent float/encoder members and nine quantization comparisons remain unavailable. Original pass/block control counts are retained in the pinned provenance. The guard's three negative-control positives, character models' two and MLPs' twenty missed positive controls remain unchanged. Shared templates are not independent statistical units or policy labels, so these intervals only check software. No human review, task-quality claim or independent verification occurred.
Validation: 40 new tests, 47 focused checks and all 1,812 tests pass. Both core
and combined train/guard package gates pass, including the installed report
fixture and comparison pipeline. The first package attempt rejected uv 0.13.4
before building; dist/package-check-go3nqolb retains that failure. Restoring
required uv 0.9.18 in a private tool directory (published wheel SHA-256 checked,
MIT/Apache notices retained) resolved it without modifying global tools or source.
The package snapshot before this evidence/documentation update is
e21d8170269b46ab75c2702576326e47705ffbf290af71d7cc0667ce1d4ec51f.
No proof run, HF token, paid compute, model download or private-data upload occurred.
Exact-head hosted checks remain required before merge.
2026-10-22 — W16 contextual byte CNN screen and exact export¶
Decisions 0059 and 0060, issue #17. The owner permits agent-reviewed synthetic pipeline work, local CPU execution and bounded existing-credit fallback. This campaign used only the Mac, with zero new compute spend and no downloaded model or external data.
Hypothesis: a small order-sensitive classifier can distinguish active versus quoted synthetic action clauses where every pair has the same byte histogram. The byte-count baseline necessarily has equal TPR and FPR on these pairs. New candidate scripts reuse the existing integer DAG/encoder and frozen evaluation functions; no evaluator or historical split is modified.
The initial 8-channel candidate fails the all-seed screen: seeds 17/29 classify
20 positives and 20 benign examples correctly; seed 43 has 20 TP, 3 FP, 17 TN.
Two fresh runs reproduce 65 substantive artifacts. Its frozen source is retained
at commit 42ae7ab, with numeric evidence.
The old test is explicitly exploratory training data for the new experiment.
The 32-channel follow-up has 4,161 learned parameters, 156 training groups, 50 calibration groups and 50 fresh test groups. Each group contains opposite labels with the same bytes in a different order. The grammar and vocabulary are shared; labels follow the artificial run/quote rubric. All three seeds use CPU float64, 200 full-batch Adam epochs, exact 100+100 resume, then int8/PTQ export and calibration before any test score. Termination is the fixed budget, not convergence.
Both follow-up runs pass the predeclared TPR >=0.8/FPR <=0.1 structural screen: TP=50 for every seed; FP=0/1/0, TN=50/49/50, FN=0. All 65 substantive files match byte for byte, including full/resumed checkpoints, quantization, graph/weights, calibration and predictions. Native scalar DAG scores match the separate vector integer oracle for all 200 calibration/test records per seed in each run. The maximum float/integer score difference across the reported test roster is 0.2584964230387712 float-logit units; no tested decisions flip under either the integer threshold or separately calibrated floating threshold.
The follow-up numeric evidence
SHA-256 is 7e4d44333797166010e2951552ae6d4ea2fdfc8298178d089d0949d9cf019a29.
Actual run times are 168.01392870908603 and 174.78115574992262 seconds; peak worker
RSS is 159,924,224 and 162,299,904 bytes. Raw artifacts and controller/resource
receipts remain local under runs/w16-contextual-cnn-1792663486688409000/:
run-01/02 preserve the negative first screen and followup-01/02 the new screen.
Every run has an externally retained protocol/manifest pin and before/after source
hashes. See operations for the
exact pinned command. Checkpoints and raw traces are not in git.
Validation: ten focused tests, make check (2,246 passed, 50 optional/platform
skips, 156 software conformance cases) and make package-check-train pass. The
installed-wheel check executes the smallest 32-channel train/export fixture.
Retained failures include the first output-fraction lowering mismatch, a directed
out-of-envelope zero-weight fixture, and an unstaged-file package snapshot; none
is relabeled as a successful run. Full hybrid conformance remains unexecuted.
The complete ten-node map admits only 0–96 bytes, including empty, NUL, Unicode and the maximum boundary; overlength rejects. It preserves masks, signed ranges, rounding and strict threshold equality. Backend/key/setup fields are zero research placeholders, with no operational enrollment. Seed 17 was selected in the second protocol before test for the next proof integration. These results do not establish real-world quality, a ZK proof, independent security review or hardware provenance. Issue #17 stays open pending qualified-graph acceptance.
2026-10-22: reproducible demo documentation and MIT adoption¶
Issue #92. Hypothesis: a reader can reproduce the trained rule/MLP/CNN research
pipeline from a checkout using documented prerequisites and complete commands,
then distinguish its result from operational/hardware qualification. Baseline:
aef20aa9082b5f0a654872f09c61b5d024cd43db; scattered recipes and a chronological
README, with stale replay-only wording. Mutable paths: README, supporting docs,
licence/package metadata and this evidence/state record. Runtime, dependency
lock, models, data, evaluator, thresholds and circuits remain unchanged.
The numeric reproduction record
was retained before this prose. All nine shell blocks in the walkthrough
ran in a fresh isolated worktree (clone/cd omitted because the worktree already
existed), with a newly downloaded, checksum-verified Go archive and separately
installed verifier. Both quantization pins reproduced. The CNN's three seeds
retained TP=50 and FP=0/1/0. Nine real proofs verified under fresh keys, all three
case decisions matched, 18 cross-request substitutions rejected and 405 arithmetic
cases matched their expected outcomes. Source hashes stayed unchanged throughout.
Models and circuits match the original composition study. The proof campaign took
882.44 seconds, retained 2,574,972,675 bytes and peaked at 2,121,187,328 worker-RSS
bytes. Software checks also ran on the host; this is documentation reproduction,
not a controlled performance benchmark or new held-out quality evidence. No cloud
compute was used and no proof worker failed. All raw runs remain local under
runs/readme-demo-2026-10-22/runs/.
Local validation: make check, 2,206 passed with 121 legitimate optional-dependency
skips; make check-train, 76 passed; specification/link and 156 software conformance
cases passed, with 32 full-evidence obligations still unavailable. Core/train
package gates passed. Explicit archive inspection confirmed MIT metadata/text in
six wheels, MIT and all three new guides in four sdists. GitHub Markdown rendering
recognized all four README tables and the Mermaid container, but the live browser
returned “Unable to render rich display.” Replaced the dynamic diagram with a
self-contained SVG carrying a title, description and Markdown alt text; its
local Chrome preview rendered correctly. This is
a focused documentation preview, not a full accessibility audit of GitHub's interface.
After the reproduction, the owner explicitly requested MIT. Decision 0063 records the grant, updates current licence guidance and keeps third-party terms and historical experiment bytes intact. The repository remains public; packages and checkpoints were not uploaded. The documentation gate is complete; backend, hardware, operational verification and real-world quality gates remain open.
2026-10-22: One-command classifier reproduction and proof explanation¶
Issue #94. Hypothesis: a single make e2e can reproduce the frozen manual
walkthrough without changing its models, data, evaluator, circuit or verifier.
Baseline: 7e7eab1c1c514d7e584e44f811af7fe80574a42e and the documentation
reproduction. Mutable paths: Makefile, the new driver and focused tests, public
documentation, state/ledger and numeric evidence. Criterion: both export pins,
nine verified proofs, the exact three decisions, 18 rejected substitutions and
405 expected arithmetic outcomes, with unchanged executable source during the run.
The numeric record was saved before this
summary. Exact command: make e2e, with uv 0.9.18 on PATH. The driver acquired
pinned Go, built the existing circuits, prepared the hash-pinned separate verifier
and invoked the existing MLP, CNN and common-request controllers. Raw evidence:
runs/e2e-entrypoint-2026-10-22/runs/e2e-i8mto7kh; validation logs:
runs/e2e-entrypoint-2026-10-22/runs/validation. All are local and ignored.
Result: passed in 1152.16 seconds; CNN training 171.26 seconds;
proof campaign 955.93 seconds. Nine proofs verified; decisions were
[1,1,1], [1,1,0], [0,1,1]; all 18 cross-request substitutions rejected;
135 arithmetic cases per classifier matched, including valid controls. Both
model hashes and all three circuit hashes match the earlier reproduction. Peak
worker RSS was 2105245696 bytes and retained proof output
2574971771 bytes. New cloud compute cost: zero. Source pins matched.
An initial unit fixture intercepted the platform query incorrectly; it was fixed
and retained in the validation record. A Make dry-run regression then exposed
an unintended expensive default target. After the full reproduction, only the
recipe order changed to preserve make as validation; make e2e and the driver
are unchanged. The focused tests cover default/explicit targets, retained
failures, timeouts and detached-worker cleanup, failed CNN screens, changed
exports, incomplete proof results and unsupported hosts.
The README and proof explainer now explain the exact statement,
public inputs, trusted local enrollment, native verdict recomputation and
separate arithmetic verification. release remains only an all-pass classifier
decision. No training-history, response-generation, real-task quality, hiding,
hardware or qualified backend claim follows. Validation: make check passes 2,292 tests with 50 optional/platform skips,
lock/lint/format/typing/spec/admission and 156 software conformance cases; 32
full-evidence obligations remain unavailable. All 15 focused driver tests and
76 required training tests pass. Core and train package checks pass, with the
new driver, proof explainer and tests present in the source archives. Hosted
exact-head and exact-main checks remain publication gates.
The first hosted core run exposed a test portability issue: GNU Make prepends a
directory banner when the dry-run assertion is itself run under make check.
The assertion now requests --no-print-directory for both dry runs. This changes
only the test's output normalization; the entrypoint and all measured pipeline
bytes remain unchanged. The failed hosted job and correction are retained in the
numeric evidence record.
2026-10-22: Proof-controller cleanup¶
Hypothesis: one shared worker runner can replace four copies without weakening
sandbox isolation, input/output checks or resource limits. Baseline: main
421615b, 79 worker tests. Mutable code: the shared worker module (now
_worker.py), the four proof controllers, two module descriptions and the worker tests. Models, circuits,
fixtures, numerical contracts and evaluator code remain frozen.
Pass criterion: existing checks plus rejection of changed/missing/extra worker artifacts, failed status/process, malformed output and attempts to write through read-only grants; a later worker must not inherit those grants. All 87 worker tests pass. A fresh cubic proof verifies, a changed public input rejects, and three retained classifier proofs pass separate verification through the shared runner. This tests orchestration, not a new classifier or LLM result.
make check passes 2,229 tests with 121 platform/optional-extra skips;
make package-check passes. Required train tests pass 76 cases; focused
controller/export/entrypoint regressions pass 96. No new dependency or cloud spend. The
evidence record pins changed source,
commands, counts and retained failures. Production code shrinks by 99 lines.
No full training/proving campaign or cryptographic audit was repeated.
2026-10-23 — Real LLM build and inference preflight (#19, in progress)¶
Decision 0064 and the raw numeric evidence preserve the expanded owner target: a complete short pretrained response, explicitly quantized, with a fresh verifier and changed-input/output/model rejection. No old classifier result is reused as LLM evidence.
Pinned SASH eb95bfa fails to link on the Mac because stwo_cuda is unavailable.
A separate two-attribute linker candidate also fails on reachable CUDA symbols;
it is retained and not adopted. Both use the pinned nightly and locked crates.
The public Apache-2.0 TinyLlama-1.1B-Chat checkpoint matches its exact SHA-256.
A fresh locked witness environment runs on the Mac. The first wrapper failed on
the chat-template API return type before model loading; explicit tensor extraction
fixes that without changing the frozen prompt, eight calibration texts or gates.
For the one authored prompt, maximum float-logit difference is 0.00009918212890625,
and argmax agreement is 1.0 across all 25 prompt positions. The quantized variant
y-delay-vw12 emits "The capital of France is Paris." then EOS at step eight,
matching the reference's eight token IDs. The 16-step fixed run retains its
post-EOS whitespace. It is a compatibility screen, not a quality benchmark.
The authorized Linux build uses a digest-pinned CUDA image on HF CPU Upgrade,
with the exact source-archive hash, 3,600-second timeout, no automatic retry,
public inputs only and $0.10 total build-attempt ceiling. Live credit and active-job
checks are retained privately. The build completed in 482.898 running seconds. Its aligned synthetic proof
reproduces db6d17a700c25457 with no unbound layer edge or unsound skip: proving
9.464 seconds, verification 0.147 seconds, serialized proof 23,594,585 bytes,
peak process RSS 5,361,160,192 bytes. All three selected upstream integration
suites pass (six test functions). Ten output files totaling 86,269,880 bytes are
retained and rehashed. The conservative compute bound is $0.005, not an invoice.
This synthetic proof does not establish a real LLM proof.
Local commands/logs and their hashes are under the isolated worktree's
runs/preflight/: build-command.json, mac-link-protocol.json,
generation-protocol-v2.json, prepare_model_v2.py, hf-job-plan.json and
hf_build.sh. No changes to Hallmark's classifier arithmetic, models or verifier.
The first complete 22-layer extraction passed numeric bounds but misapplied
--embed to every layer. Source/manifest review caught this before cloud proving;
it is retained as unusable for a full proof plan. The correction re-extracts
layers 1–21 without that flag and retains the same layer-0/head arrays. Model,
source, calibration, arithmetic, prompt and expected answer are unchanged. The
corrected layer pass takes 8.95 seconds, with 10,832,773,120-byte process peak RSS.
The independent byte-level inventory checks 1,856 arrays (1,352,364,544 bytes),
22 residual links, all 32 input IDs and the eight head outputs at positions 24–31.
This includes the final EOS output, which has no following input for a continuity
check. All checks pass. The plan contains 22 layers plus one head, with fixed
21–26-bit table widths; the tightest no-wrap margin is one bit. These are witness
checks, not cryptographic verification. The initial metadata-copy path mistake
and missing _L shape-reader prefix are retained as preparation failures.
Exact extraction commands (from the isolated worktree, using the upstream locked Python environment and the pinned local model/calibration):
python -m harness.witness.extract --tokens 32 --chunk 32 --layers 0 --blocks all \
--spec y-delay-vw12 --ids INPUT_IDS --attn-ranges ATTENTION_RANGES \
--residual-stream --embed --diagonal-block --slim --slim-num --out WITNESS
python -m harness.witness.extract --tokens 32 --chunk 32 --layers 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21 --blocks all \
--spec y-delay-vw12 --ids INPUT_IDS --attn-ranges ATTENTION_RANGES \
--residual-stream --diagonal-block --slim --slim-num --out WITNESS
python -m harness.witness.extract --tokens 32 --chunk 32 --layer 21 --blocks all \
--spec y-delay-vw12 --ids INPUT_IDS --attn-ranges ATTENTION_RANGES \
--residual-stream --head --diagonal-block --slim --slim-num --out HEAD
The corrected run is runs/preflight/full-witness-v2; validation.json records
full file hashes and plan.txt records all 23 segments. The first real-layer
job uses CPU XL (16 vCPU, 124 GB), a 96 GiB address-space limit, one-hour provider
timeout and $1.10 total ceiling. Its model roots must be retained locally before
a caller marker allows proving. The whole-response proof and mutation suite
remain the acceptance gate; no whole-graph qualification is claimed.
The first real-layer job terminated before proving: the root builder rejected
one denominator among 1,024 rows. Its 23.332 running seconds have a conservative
$0.025 compute bound; failed outputs and commands are retained. Local reproduction
identifies head 1, token 16. The exact integer and float64 dot product are
17,866,521; int_gemm_exact returns 17,866,520 after its unconditional float32
cast, exactly matching the recorded maximum. The resulting shift is -1. The
Rust loader rejects it; no mask, denominator or range check was disabled.
Decision 0065 freezes a
new candidate that removes only that final cast. All four helper callers were
inspected. research/probes/sash/check_exact_gemm.py fails on the original helper
and passes four int64-oracle cases after applying exact-gemm.patch. The exact
real-row regression also reproduces twice. The first diagnostic translation
omitted the Rust table's out-of-range behavior and did not reproduce the failure;
its failed assertion is retained separately.
The fresh candidate run reproduces the same calibration hash and all 16 generated IDs, including the eight-token answer through EOS. Generation takes 10.91 process wall seconds; the measured inference/calibration section is 8.766 seconds. Every layer and head was regenerated, with numerator arrays retained for upstream cross-checking. The new inventory validates 1,902 arrays (1,355,379,200 bytes), 22 residual links, all eight output tokens and all 23,552 exact attention rows. The output-head extraction repeats layer 21's attention, so that row count is 23 segments × 32 heads × 32 positions, not 23 distinct transformer layers.
A separately frozen corrected first-layer job is running under the same one-hour and $1.10 ceiling. The failed attempt is terminal, retained and charged separately. The binary is unchanged; the witness patch hash is explicit in the new manifest. Its root result must be locally enrolled before proof execution. Whole-response proofs and changed-claim rejection remain outstanding.
The corrected root builder passes in 15.148 seconds with 15,951,828 KiB peak child
RSS. Its manifest is retained locally with SHA-256
b08d0a7afb4d0485f46cc77283010c6ab36332806b84804edb36a471dae8c638
before the caller marker releases the proof stage. This is enrollment by a
separate controller using the same implementation, not independent security review.
Repository validation: make check passes 2,229 tests with 121 platform/optional
skips, all 156 software conformance cases, lint, formatting, typing, specification
and admission checks. make package-check passes; its source archive includes the
new optional probe, patch and upstream licence/notice. The checker initially
failed import grouping, then the complete gate passed after that formatting fix.
No core runtime, dependency lock or existing classifier relation changed.
The corrected first-layer job subsequently failed during proving, after model-root
enrollment. A 33,554,432-byte allocation failed under its 103,079,215,104-byte
address-space cap; the prover exited with signal 6. Proving ran for 363.917 seconds,
and peak child RSS was 96,638,864 KiB. The provider reports 481.051 total running
seconds; the conservative compute bound is $0.15, not an invoice. All eight output
files (8,442 bytes) are downloaded and hashed in
runs/preflight/layer0-exact-job/result.json; no real-layer proof was produced.
This closes the witness-precision investigation with a passing root build, while
proof feasibility remains open. The next experiment increases the resource
envelope with identical arithmetic, binary, input archive and pre-enrolled roots.
2026-10-23 — Verifier shape reconstruction¶
The larger CPU run used the same binary and witness, reproduced the enrolled
roots, and serialized a 32,039,654-byte proof in 82.493 seconds. Peak RSS was
111,457,673,216 bytes. Overall acceptance remained false: parameter-only shape
reconstruction panicked with v: fold token divisor 2^7 below alpha 9. The
reported STARK subcheck passed, but that does not establish accepted inference.
The job ran for 119.436 seconds, with a conservative $0.07 compute bound, not an
invoice. All 11 outputs are locally retained and hashed. After checking all 33
outputs from the build and three layer attempts, the old temporary bucket was
deleted; no compute job remained live.
Decision 0066 freezes the next candidate. The suspected cause is the parameter-only constructor's token and channel scale placeholders, which must respect the statement's fold split and remainder window. The small constructor regression starts with alpha 9; the actual saved proof must reproduce the original rejection twice before the patch is applied. Source pins, original proof/statement/claims, enrolled roots and all verification checks stay frozen. The new binary must disclose the patch.
The new CPU Upgrade job uses four threads, a 28 GiB address-space cap, one-hour
provider timeout, $0.10 total ceiling and no automatic retry. It receives only
the saved public proof, caller expectations, source patch and regression. Its
protocol and exact commands are retained under runs/shape-repair-2026-10-23/.
The complete 23-segment response proof and caller-side output binding remain open.
The shape diagnostic completed: the unmodified verifier reproduced the rejection
in 1.57 and 1.52 seconds. The constructor regression failed before the patch and
passed afterward; all six tests in the three upstream suites passed. With the
patch, the saved proof verified against the same enrolled roots in 21.406 process
seconds, with 14,332,235,776 bytes peak child RSS. Changed prompt, model root,
calibration, statement root and proof commitment all rejected. The job ran for
959.318 seconds, with a conservative $0.01 compute bound. Every output is retained
and hashed in runs/shape-repair-2026-10-23/retention.json; its bucket is deleted.
This is a constructor result, not complete inference integrity. Review of the
original detailed verdict found rq_scale_open: ["qkv/v"]. Pinned
prover/src/compose.rs explicitly skips the equality between the requantizer's
scale and the delayed-value fold component's scale. The broad binding table still
reports all edges bound. No forged proof is demonstrated; the omitted equality
is sufficient to reject that profile for this task.
2026-10-23 — Non-delayed value projection¶
Decision 0067 selects
the existing y-attnout-w12 profile with value rounding restored and the chain
scale equality implemented. The checkpoint, authored prompt, calibration texts,
Python exact-GEMM fix and Rust binary remain pinned. No delayed-profile proof
or root is reused. Native evidence
records the same eight-token answer through EOS and the disclosed fixed-length
suffix. This is one engineering prompt, not a language-model quality evaluation.
The first runner misinterpreted the extractor's output prefix and stopped after successful layer-0 extraction; the directory and its failure are retained. A new run extracts all 22 layers and the output head in three separate processes. Its first checker wrongly required the head's declared power-of-two dequantization to be chain; the corrected checker requires the exact declared variant for each matmul. It validates 1,902 arrays, 1,355,379,200 bytes, 22 residual links, 23,552 exact attention rows and all eight output tokens. No matmul uses the delayed fold.
The upstream field reader derives actual key bounds before enrollment. Those bounds yield 21–26-bit tables, avoiding the conservative 30-bit table in the initial plan without changing arithmetic. Both plans and archives remain retained. The portable archive verifies all 1,930 entries and changes only three provenance labels, preserving every numeric array byte. Its hash is recorded before upload. The next CPU Performance job has a one-hour provider timeout, 210 GiB address-space cap, 16 threads and $2 total ceiling. It must stop on any failed proof or detailed open binding; the guard rejects the original open-scale verdict and six targeted mutations. The complete response, separate proof-only verification and changed-claim suite remain required.
Repository checks pass: make check reports 2,229 passed and 121 platform/optional
skips, with all 156 software conformance cases passing and 32 unavailable.
make check-spec and make package-check pass after the documentation and probe
additions. The core package check does not execute the train or guard extras.
2026-10-23 — Complete TinyLlama response composition¶
Decision 0068 corrects
the output-head extraction. The first non-delayed run verified all 23 individual
proofs and every token but failed the last residual link. Both --head and
--residual-stream had been supplied; the Rust loader selected the ordinary layer
family before the head's final norm. The head's commitment roots exactly repeated
layer 21's roots. The failed grid and all outputs remain retained; that job's
compute bound was $0.34, not an invoice.
Fresh head extraction omits --residual-stream. Every retained numeric array
matches the earlier extraction; xin_head.u16 equals layer 21's output. A new
head root was locally enrolled before proving. The unchanged 22 transformer
proofs and newly generated head now pass all 23 cryptographic verifications,
23 enrolled shape roots, 22 residual links, 25 prompt positions and seven
generated input positions. The caller also checks all eight output tokens,
including the final EOS. The detailed open-binding fields are closed.
The new head's prove stage took 21.903 seconds. Full-grid verification took 92.097 seconds with peak child RSS 14,341,046,272 bytes on HF CPU Upgrade, four Rayon threads. The original 22 transformer prove stages total 307.004 seconds on the larger CPU worker; these separate timings are not one end-to-end wall-time benchmark. The new complete set has 737,858,641 serialized proof bytes, 761,015,496 bytes including statements and claims. All outputs are retained and hashed; the corrected-head bucket is deleted. Its compute bound was $0.01.
Exact local commands, relative to the isolated research checkout:
runs/preflight/witness-env/bin/python runs/preflight/extract_head_norm.py
runs/preflight/witness-env/bin/python runs/preflight/head-norm-job/package.py
runs/preflight/witness-env/bin/python runs/preflight/head-norm-job/prepare.py
runs/preflight/witness-env/bin/python runs/preflight/head-norm-job/launch.py
runs/preflight/witness-env/bin/python runs/preflight/head-norm-job/enroll.py
runs/preflight/witness-env/bin/python runs/preflight/head-norm-job/retain.py
These retained run scripts are not shipped CLI commands. Their protocols, inventories, commands and raw outcomes are hashed in the composition evidence. The public caller guard checks exact statements against enrolled parameter digests and rejects the old head family. The fresh proof-only worker below closes this verification gate; user-facing LLM reproduction is the next task.
The fresh chain-verify-job-v2 worker received only the 69 proof/statement/claims
files, frozen caller expectations, checker and original binary. It accepted two
honest processes in 83.986 and 79.162 seconds. All 11 caller mutations rejected;
the three cryptographic mutations rejected with concrete sum-check/STARK failures.
The mutated cases cover prompt, answer, model, calibration, coverage, coordinates,
final output, token alias/type, open-scale profile, security parameters, proof
commitment, residual boundary and output claim. Rehashing the control bundle
confirmed that negative-copy mutations did not change its bytes.
The worker ran 421.055 seconds, with 14,347,567,104 bytes peak child RSS and a
conservative $0.01 compute bound. The cumulative campaign bound is $0.62, including
retained failures and builds, not a provider invoice. All outputs are retained and
hashed; the temporary bucket is deleted. Exact commands use the same interpreter
and chain-verify-job-v2/{package,prepare,launch,retain}.py; that directory retains
protocol.json, expected.json, proof-inventory.json, retention.json and raw
native verifier output. The public checker hash equals the frozen worker input.
This is separate-process verification with the same implementation, not an
independent security review.
make check passes: 2,247 tests passed, 121 optional/platform skips; 156 software
conformance matches and 32 unavailable full-evidence obligations. The 18 new tests
cover caller validation, missing coverage, JSON ambiguity and optimized-Python
rejection. No core dependency, classifier relation or admitted backend changed.
make package-check also passes. Its core environment omits optional training;
the train-extra CI jobs remain required. Raw checkpoints, proofs and credentials
remain outside Git.
2026-10-23 — One-command TinyLlama reproduction¶
Decision 0069 packages
fresh native preparation, a fresh Rust build, all 23 proof invocations and the
proof-only verification suite as make e2e-llm. The checkpoint, calibration,
quantized relation, prompt/answer and caller-enrolled roots remain frozen. The
new evidence records source/input
hashes, provider receipts and every verification outcome.
The first invocation failed before creating a cloud job: Git discovered the
parent Hallmark checkout and silently skipped the witness patch. The exact-GEMM
oracle caught the unpatched float32 conversion. That run remains unchanged at
runs/llm-e2e-y372z3cr. Isolating Git's parent search fixes the cause; the nested
checkout regression and all four exact integer oracle cases pass on fresh source.
The first repository gate also rejected the changed whole-lock hash. The new
optional llm extra only exposes the already locked Hub SDK; a separate review
confirms all 52 third-party package entries and previous licence evidence are
unchanged. The previous review is preserved.
The corrected invocation, runs/llm-e2e-e8yegx2i in the isolated checkout, exits 0
in 1,875.20 seconds (31 minutes 15 seconds). Native preparation takes 32.22 seconds
and reproduces the calibration hash, answer and all residual arrays. The fresh
build passes the frozen synthetic digest and six upstream control tests. Newly
derived roots match all 23 caller-owned roots before proving. Every proof,
statement and claim file is byte-identical to the prior accepted bundle: 69 files,
737,858,641 proof bytes, 761,015,496 bytes with statements and claims. No old proof
was supplied to the proving job.
The 23 prove stages total 328.459 seconds; the full proving job takes 533.637 seconds, with peak process RSS 116,204,756,992 bytes. A fresh worker with no weights or witness verifies the complete response twice in 80.110 and 78.446 seconds. All 11 caller mutations and three native cryptographic mutations reject; the control files remain unchanged. Peak verifier RSS is 14,345,064,448 bytes.
The build/prove/verify compute bounds are $0.01/$0.29/$0.01, respectively; these are conservative bounds, not invoices. The controller's current-rate timeout bound plus storage reserve was $1.96 against the $2.10 requested ceiling. No retry or top-up occurred. All three output inventories are locally retained and hashed, and all temporary buckets are deleted. The build uses the provider's 32 GB memory limit; the prove and verify workers additionally set process address space limits. A new binary hash is recorded, without a byte-identical build claim.
Commands, with uv 0.9.18 on PATH:
make e2e-llm
.venv/bin/pytest -q tests/test_llm_e2e.py tests/test_sash_response.py
UV_PROJECT_ENVIRONMENT=runs/llm-workflow-2026-10-23/check-env make check
UV_PROJECT_ENVIRONMENT=runs/llm-workflow-2026-10-23/check-env make package-check
The 33 focused tests pass. The full core gate passes 2,261 tests with 122
platform/optional skips, 156 matched software conformance cases and 32 unavailable
full-evidence obligations. The extra-dependent controller case runs in the focused
suite and the new offline LLM CI job; core-only runs legitimately skip it.
Distribution checks pass. Gate logs and the failed first check are retained under
runs/llm-workflow-2026-10-23.
This closes the one-command reproduction gate. The fixed authored prompt is an engineering case; independent security/hiding review, authenticated enrollment, operational backend admission and hardware provenance remain separate obligations.
The first hosted train run at a506dab failed on a second historical whole-lock
pin in the design-power fixture; the failed job
and its local log are retained. A new lock-only fixture preserves all numeric
expectations, evaluator/source hashes and acceptance thresholds. After that
refresh, make check-train executes and passes all 76 selected tests. This changes
no LLM workflow source file or statistical design; the real proof-run hashes
remain unchanged.
2026-10-23 — Native verification with compiled shape enrollment (#19)¶
Hypothesis: a verifier specialized to the frozen TinyLlama parameters can check existing STARK/GKR proofs on the Mac without rebuilding shape trees or linking CUDA. The Linux runtime-derivation verifier remains the baseline. Candidate scope, preprocessing authority and exclusions are in decision 0070.
Command: make verify-llm LLM_RUN=runs/llm-e2e-e8yegx2i, with uv 0.9.18 and Rust
nightly-2025-07-14. Raw numeric evidence and hashes
precede these conclusions. Fresh directories runs/llm-verify-wckqj1l9 and
runs/llm-verify-zh784135 finish in 128.750655 and 128.186243 seconds. Each accepts
23 proofs and 22 links in two processes, rejects the original 14 mutations, and
explicitly rejects unknown shape/calibration in Rust in two additional controls
with caller checks bypassed. Honest verification takes 3.866991/3.275166 and
3.405621/3.215978 seconds. Peak sampled process-group RSS is at most 2,508,308,480
bytes for compilation and 493,993,984 bytes for the complete suite. Cloud cost is
zero; no new model download or proving job occurs.
Both runs use the same native patch, caller/worker code, compiled shape table and proof archive. The confirmation adds a controller-side sandbox fixture and passes fork, network-denial and outside-file-denial checks. Binary hashes differ and are recorded separately; compiler bit reproducibility is not claimed. All original proof files remain unchanged. The verifier's modified-source flag and compiled root authority are explicit.
Failures retained: two direct Mac linker candidates, a sandbox fork denial before cryptographic checking, and the first packaged native unknown-shape control, which reached upstream constraint code before enrollment validation. The final candidate checks the compiled enrollment first and returns a rejection report. The compiled table is a caller trust input, not proof-supplied data. This does not qualify operational enrollment, independent security/hiding review or hardware provenance. #19 remains open.
Local make check passes 2,274 tests with 121 optional/platform skips, 156
matched software conformance cases and 32 unavailable full-evidence obligations.
make package-check passes. Its first attempt omitted new files that were not yet
staged in the reviewed-path snapshot and failed the documentation link check;
staging those files fixes it without a runtime change. Both attempts and the gate
logs are retained. Optional training runs in the required train-extra CI job.
Native linkage inspection reports no CUDA library or undefined CUDA symbol.
2026-10-23 — Native token and proof-shape boundaries (#19)¶
Hypothesis: the optional native verifier should reject field aliases and malformed proof dimensions explicitly, while retaining the frozen TinyLlama proof relation. Baseline: PR #101's native executable, public caller enrollment and original 23-proof bundle. Mutable inputs are fresh per-case copies only. Limits are four threads, 600 seconds, 4 GiB sampled process-group RSS and zero cloud spend. Decision 0071 freezes the acceptance criterion; numeric results and raw hashes precede these conclusions.
The 24-case baseline probe takes 75.522939 seconds and peaks at 1,018,494,976 sampled process-group RSS bytes. Both honest controls pass. Adding the M31 modulus to the final output token ID is accepted with the caller bypassed; the existing caller rejects the out-of-vocabulary value. Empty commitment/decommitment/query lists and an empty preprocessed column-size vector crash the native process. Six aliases in claims/proof fields reject cryptographically. Three altered redundant configuration records and two unknown JSON fields are accepted; the verifier's effective security configuration remains caller-pinned. No forged valid-vocabulary response through the complete workflow is demonstrated.
The native patch now checks token vocabulary ranges, exact tree/column dimensions, preprocessed sizes against the trusted plan, and redundant proof configuration. It returns before Stwo when derived trace sizes disagree. The suite adds 20 native boundary cases and requires their explicit rejection reasons; crashes fail it. Unknown JSON fields retain no proved meaning and are not a canonical encoding claim. The Linux runtime-derivation verifier and proof generation stay unchanged.
Command, with uv 0.9.18 and Rust nightly-2025-07-14:
Fresh directories runs/llm-verify-3m4i8_d2 and runs/llm-verify-80_jq63n pass
all 38 cases: two honest processes and 36 explicit rejections each. Total workflow
times are 214.447364 and 203.981406 seconds. Honest verification takes
3.996549/3.569800 and 3.503289/3.328227 seconds. Peak sampled process-group RSS is
2,474,819,584 bytes for compilation and 1,191,772,160 bytes for the expanded suite.
Both runs pass sandbox fork/network/outside-file controls, preserve all original
proof files and use identical runtime source hashes. Binary hashes differ and are
retained; bit-reproducible compilation is not claimed. No cloud job starts.
This closes the tested native-boundary regression gate. It does not establish all malformed inputs harmless, authenticated operational enrollment, hiding or independent security qualification. Issue #19 remains open.
Local validation: 50 focused tests pass with one optional SDK skip. make check
passes 2,279 tests with 122 optional/platform skips, 156 matched software
conformance cases and 32 unavailable full-evidence obligations. make package-check
passes. Neither candidate binary links CUDA or has an undefined CUDA symbol.
Logs and hashes are retained under runs/boundary-delivery. Required optional
runtime jobs remain separate hosted CI obligations.
2026-10-23 — Signed LLM parameter enrollment (#19)¶
Hypothesis: an externally supplied Ed25519 key can authorize the exact frozen LLM case and verifier source inventory, while invalid or expired records fail before compilation. Baseline: decision 0071's native verifier and retained TinyLlama 23-proof bundle. Mutable candidate paths are the enrollment module and native controller; arithmetic, Rust constraints, quantization and proof files stay frozen. Decision 0072 defines the format and pass criterion. Raw numeric results and hashes precede these conclusions.
The first study passes nine preflight rejections and two real signed proof workflows. Format review then corrects the JSON schema's signature description from 64 characters to the actual 128-character hex encoding. Both preliminary runs remain under their original source identity. A public conformance vector and strict schema snapshots bind the format. A separate Node v24.22.0 fixture check independently constructs the canonical body/message, verifies the signature and rejects a changed message. Shared cryptographic-library risk remains; this is interoperability evidence, not independent security review.
A fresh current-source study rejects ten cases before compilation: wrong key, expired/not-yet-valid enrollment, wrong case, wrong verifier source, tampered signature, malformed record, either missing trust input, and the earlier valid signature over the now-changed source. All have the exact expected errors and retained failure directories. No invalid enrollment falls back to the checkout case.
The real signed command, using a separately supplied disposable public key, is:
LLM_ENROLLMENT="$ENROLLMENT_FILE" \
LLM_TRUST_KEY="$SIGNER_PUBLIC_KEY_HEX" \
make verify-llm LLM_RUN="$RETAINED_LLM_RUN"
Intermediate directories runs/llm-verify-fro254_8 and
runs/llm-verify-nguo9lg8 pass all 38 proof cases each under signed enrollment in
210.808980 and 206.873425 seconds. A separate checkout-case compatibility run,
runs/llm-verify-qrwmb70l, passes the same suite in 205.564529 seconds. Each accepts
two honest verifications and explicitly rejects all 36 mutations. The four signed
honest checks take 3.556955/3.379222 and 3.777087/3.393402 seconds. All original
proof bytes remain unchanged. Source inventories match across the three current
runs. Compilation peaks at 2,480,865,280 sampled process-group RSS bytes and the
suite at 1,187,692,544 bytes. All sandbox controls pass; no cloud job starts.
Version tags are then added to the new authority/result JSON reports. This
changes only the controller's source hash and therefore requires fresh enrollment.
Final-source directories runs/llm-verify-w4bb0vvk and
runs/llm-verify-mw88chgy each pass all 38 proof cases and all three validity
checks. They take 205.468509 and 206.545980 seconds. Honest checks take
3.568990/3.366695 and 3.513864/3.391349 seconds. Peak sampled process-group RSS is
2,447,802,368 bytes for build and 1,152,598,016 bytes for the suite. Ten preflight
rejections pass again, including the previous source's otherwise valid enrollment.
All earlier runs remain unchanged under their own source identities. No constraint,
proof, quantization or calibration changes occur in these format corrections.
The signer authorizes 55 exact source files and case bytes. Both signed runs
check validity before build, before the worker and after verification. The
integration private key is never serialized; its public key is supplied outside
the enrollment. It is a disposable research issuer, not an owner-provisioned
operational root. The legacy path reports checkout-case explicitly. Revocation,
persistent rollback protection, operational custody and independent security
qualification remain separate gates; issue #19 stays open.
Final local gates pass: 2,301 tests with 122 optional/platform skips, 47 typed
source files, 156 matched software conformance cases and 32 unavailable genuine
evidence obligations. make package-check passes. Focused tests pass 72 cases
with one optional SDK skip. Final logs are retained under runs/enrollment-delivery;
all 55 final source hashes still match both signed proof runs.
2026-10-25: owner-scoped LLM integrity MVP completion audit¶
Issue #110; decision 0076. The owner explicitly selected “Finish the LLM integrity MVP without hardware or hiding requirements.” Original private-v1 requirements, failures and evidence remain intact; its 29 open issues are deferred, not closed as implemented.
The numeric audit precedes this interpretation. It rehashes the real checkpoint and all eight model files, all 69 proof/statement/claim files, 207 current runtime files, 55 native verifier files and all 20 files enrolled by the fresh LLM proving pipeline. It checks retained actual verifier outcomes, all five historical signed authorizations, the exact opened request/public-input join, public MIT metadata and all 16 exact-source baseline hosted jobs. The scope is one pinned quantized TinyLlama answer through EOS, with specialized integrity proofs and the existing synthetic classifier pipeline. Current-source signed composition and proof-only handoff already pass; no arithmetic, model, evaluator or proof source changes here.
An initial audit assumed a file named LICENSE; the actual licence is LICENSE.md. That harness failure is retained and the corrected audit passes without changing licence bytes or runtime source. Final documentation/distribution validation and publication receipts follow. No cloud job, new spend, private upload, contact, operational key provisioning or security/hardware/hiding qualification occurs.
Final local validation passes: 2,418 tests, 50 optional/platform skips, 77 required
training tests, 47 typed source files, 156 software conformance matches, valid
Markdown links and reproducible core/installed-training package gates. Their
shared package snapshot is aef678bcbfcdfd84114e44f756fe131459321b60ec07c0c5dc2820fd90f9bebb.
Final metadata follows that package snapshot; publication CI is recorded on the
linked PR. The live audit confirms all 29 original issues remain open with
unchanged bodies and additive post-mvp labels. Runtime sources are unchanged.
2026-10-26: explain proof flows and track Circle STARK classifiers¶
Hypothesis: one command matrix and three explicit flow diagrams resolve the
ambiguity between classifier-only make e2e, SASH LLM proving and host composition.
Baseline: ef8ac66ec4ae132449cac184d8dae1b70370f35d, with a complete
fixed-case SASH LLM proof bundle but no Circle STARK classifier adapter.
Mutable scope is README, current user/architecture/roadmap documentation, STATE
and this ledger. Runtime, models, evaluator and historical evidence are unchanged.
The architecture guide records exact public statements, proof counts, caller/native-rerun checks and upstream versus patched SASH components. The controller, SASH case/adapters, native patch and existing numeric records were inspected; no new scientific measurement is made. Epic #112 has five open native sub-issues (#113–#117), with verified parent links and a four-edge dependency chain. Each has runnable acceptance requirements, failure cases and evidence obligations. The first freezes the protocol; the last requires measurements and a complete composed demo. Filing the epic is not implementation evidence or adoption of a new backend.
Pass criterion: published docs agree with current code/proof boundaries, local
links and distribution gates pass, and the live epic retains actionable open
implementation tasks. Local make check passes: 2,418 tests, 50 optional/platform
skips, 47 typed source files, 156 matched software conformance cases and 32
unavailable future full-evidence obligations. make check-train passes all 77
required training tests. Logs are retained in ignored
runs/proof-architecture-check.log and runs/proof-architecture-train.log.
Distribution and exact-head hosted results belong to the publication PR.
No paid jobs, new proofs, private uploads or qualification claims occur.
2026-10-26: surface recorded headline results in the README¶
Hypothesis: a compact table near the top lets readers assess the demo's scale,
latency and storage without searching detailed experiment reports. Baseline:
1032eb621cb5b4c5f4e8c9c69c0c0ae2d19dc491. Mutable paths are README, STATE and
this ledger. Pass criterion: every headline traces to existing numeric evidence,
units and timing boundaries are explicit, and documentation/distribution checks
pass. No evaluator, runtime or historical evidence changes.
The existing LLM reproduction records 328.4588 seconds summed proving, 1,875.1993 seconds for the full command, 80.1096/78.4465 seconds for the two honest Linux checks and 737,858,641 serialized proof bytes. Its peak prover RSS is 116,204,756,992 bytes and its compute bound is $0.31, not an invoice. The README rounds these with decimal storage units, links native verification and classifier measurements, and includes the 4,010.05-second fresh composed workflow. Compiled shape-root enrollment and different hardware prevent a matched Linux/Mac speedup claim; the smaller classifier workload is separate.
Validation commands: make check and make package-check; exact results and
hosted checks are recorded on the publication PR. No new end-to-end research
proof campaign or paid model job is required. Standard CI retains its fresh
synthetic proof fixture. This changes presentation, not measured results.
2026-10-26: bounded matched GPU prover pipeline¶
Hypothesis: the pinned SASH CUDA proof backend can prove the same complete
TinyLlama relation and produce the same serialized artifacts as SIMD on one
host. Baseline: f4e92789cdea9cd1a093143d89e21b8ba486ae43 and the frozen
y-attnout-w12 case. The admitted public witness archive is
85556911bdf8a492d2d8d31181ec0114ccd8541502ba332753510e59145ee559, with
inventory f8941cf4e3b1c7bec4b35da8b306adca1a9bca2765324295fb8931bf95c80443:
1,917 files totaling 1,355,103,851 bytes. No model, input, quantization,
calibration, root or proof-parameter change is admitted.
Mutable scope: the new benchmark controller/worker/build wrapper, an optional phase-resource argument in the existing HF transport, Make task, rejection tests and documentation. Frozen evaluator: existing caller bundle checks, closed segment verdicts, original SIMD grid verifier and native verifier patch. Decision 0077 defines the host-generation CUDA profile, resources, metric boundaries and acceptance. Issue #120 tracks completion.
The first CPU build job 6adf3532046eb82b1eceddde completed in 696.692 seconds,
producing both binaries from the unchanged pinned upstream source. Its retained
receipt bounds running compute at $0.01, not an invoice, and confirms bucket
cleanup. Before GPU launch, review corrected an exited-process sampler edge
and added startup/cancellation reserves to pricing. The original controller's
source-freeze check then rejected the changed workflow before creating any GPU
job; its protocol, failure and successful build outputs remain immutable.
A new run admits that caller-owned build by case/source/binary/retention hashes. At inspected rates its remaining creation/worker/storage ceiling is $3.78; including the first build's full $0.02 allowance bounds the study at $3.80, under the selected $4 ceiling. The live prepaid balance and rates were checked before execution; credentials and account billing details stay local. No retries, top-ups, private uploads, other-provider rentals or schedulers are authorized by this entry. All 37 focused offline tests pass. Full measurement, proof-only verification and publication results remain pending until their actual gates pass.
Retained failures and native CUDA follow-up¶
The first H200 attempt (6adf3861046eb82b1ecedf3b) rejected the synthetic
fixture's absent shape enrollment after 42.709 seconds. The fixture's actual
STARK proof passed; it has no model shape root. The dedicated fixture check
now retains that distinction, while complete LLM proofs still require enrolled
shape roots. Its running-compute estimate is $0.09, with outputs retained and
bucket deleted. This is a failed attempt, not GPU acceptance evidence.
The next H200 attempt (6adf3a34046eb82b1ecedfe6) ran 1,964.586 seconds and
failed its 1,950-second worker deadline during the output head. The complete
SIMD case passed: 551.268891376 seconds summed proving, 684.794974848 seconds
segment commands, 90.600314894 seconds grid verification and 102,277,988,352
bytes peak sampled process RSS. It produced 737,858,641 serialized proof bytes.
Both synthetic digests matched; their command times were 9.761196 seconds SIMD
and 218.7453799 seconds CUDA. All 22 CUDA layers passed strict segment checks,
but no head proof was produced. Preserve the incomplete result and 66 partial
CUDA artifacts. Its running-compute estimate is $2.75, not an invoice; all
outputs were retained and its bucket deleted. These first attempts total $2.85
in running-compute estimates, excluding invoice reconciliation.
Decision 0078 freezes a
separate follow-up: isolated CUDA target directory, checked native sm_90 code,
level-one archive compression and a 2,550-second worker/2,700-second provider
window. Inputs, evaluator, proof parameters and full-case acceptance remain
unchanged. Live credits/rates were rechecked. The new single-run ceiling is
$5, with an inspected full timeout/startup/storage bound of $4.64. Build job
6adf44a3046eb82b1ecee427 is running; no full GPU result is accepted yet.
The current focused offline gate has 38 passing tests.
Accepted native sm_90 paired measurement¶
The separately frozen follow-up llm-gpu-0x9rgmmh passes. Numeric records were
saved before this interpretation in
evidence/llm-gpu-benchmark-2026-10-26.json.
Build 6adf44a3046eb82b1ecee427 completes in 847.475 seconds and its ELF
inventory confirms native sm_90 code. H200 job 6adf480e046eb82b1ecee5d1
completes in 1,761.889 seconds, and separate CPU job
6adf4f17046eb82b1ecee83a completes in 466.229 seconds. All three output
inventories were retained and hashed before bucket deletion. The successful
command takes 3,372.335784208 seconds with reused public preparation and fresh
SIMD/CUDA builds, both complete proving runs and CPU/Mac verification.
SIMD/CUDA summed proving is 537.197419234 / 583.868902786 seconds; complete segment commands are 669.247454638 / 687.337647136 seconds. Grid verification on the same H200 host takes 88.586419848 / 85.067699138 seconds. All 23 segments, 22 residual links and seven generated-input equalities pass for both backends. All 69 proof/statement/claim files match byte for byte; each proof set has 737,858,641 serialized proof bytes. Host peak sampled RSS is 104,694,411,264 / 11,783,995,392 bytes. CUDA peaks at 68,685 MiB VRAM and 89% sampled utilization. Its synthetic fixture takes 8.009098912 seconds and matches the SIMD digest. The device report identifies an H200 and driver 580.178.04.
This profile is slower by 8.69% in summed proving and 2.70% in complete commands. CUDA records 1,710 CPU constraint-evaluation fallback calls totaling 532.745969551 seconds, versus 23.986044830 seconds total SIMD constraint evaluation. CUDA's reported Merkle/FRI stages are much faster, but the full hybrid path is slower. Generated constraint kernels are a future candidate, requiring a new source inventory and unchanged equivalence gates. Timers may nest; the phase counters must not be added to derive wall time. One paired observation with shared OS caches establishes no latency distribution.
The weight/witness-free CPU worker accepts two honest checks in 81.016620873 / 80.940695357 seconds and rejects all 14 mutations. Its maximum child RSS is 14,346,207,232 bytes. The native Mac verifier accepts two checks in 3.531152291 / 3.385023584 seconds and rejects all 36 mutations. Native compilation takes 105.123901292 seconds; its complete verification worker takes 98.530292708 seconds. Mac compiled shape roots and Linux runtime derivation remain different preprocessing policies, with the same SASH cryptographic implementation. Native error paths are redacted only in the public aggregate; raw local records remain unchanged and hash-pinned.
Running-compute estimates total $2.52 for the passing run and $5.37 across all
four attempts; these exclude invoice reconciliation and retain startup reserves
separately in the frozen $4.64 plan bound under the selected $5 run ceiling.
All six study jobs are terminal and all temporary buckets deleted. No retries,
top-ups, private uploads or schedulers ran. Current runtime checks pass 38 focused
tests and make check passes 2,441 tests (50 skipped), type/lint/spec/conformance
and admission gates. Distribution and exact-head publication checks belong to
the delivery PR. No backend adoption, hiding, hardware or full-GPU-offload claim
is made.
Publication admission checks¶
Review adds controller-only preflight for Apple Silicon and Rust before SDK
initialization or paid work. Four rejection cases pass, bringing focused
checks to 42. The full local gate now passes 2,445 tests (50 skipped), lint,
types, specification, conformance and admission checks. All 26 measured source
hashes were checked against published commit
a81c10a324fe7e849a40cf1f0caa9338e1a0cc3d; this source identity is retained in
public evidence. Prover/worker, evaluator, inputs and numeric records are
unchanged by the preflight addition, so no additional paid campaign is required.
A distribution check caught a published evaluation-heading anchor removed by
the documentation edit. The original anchor is preserved; offline links and
the corrected distribution check pass. Its failure remains in local retained
diagnostics. After the preflight change, rerun make package-check on the
staged final source; exact distribution/hosted receipts are recorded in PR #121.
2026-10-26: generated CUDA constraints, strict no-fallback candidate¶
Owner request: finish the generated constraint path and rerun the fixed model
benchmark. Hypothesis: installing the full generated shape union removes the
scalar fallback bottleneck without changing the proof relation or bytes.
Baseline: bdbf37d8c3e491d2eae2ccc1a39da65d3f3b12bd; previous paired proving
537.197419234 / 583.868902786 seconds, with 1,710 CUDA CPU fallback calls taking
532.745969551 seconds. Preserve all earlier inputs/results/failures unchanged.
Mutable paths: kernel generation/install helper, reviewed no-fallback/resource patch, CUDA build wrapper, GPU controller/worker, rejection tests and docs. Frozen: model/quantization, public witness archive and its inventory, all 23 caller-enrolled statements/roots, proof parameters, original closed/bundle/grid checks and native verifier. Extracted statements are admitted against the existing 69-file inventory; no new model/data selection or private upload occurs.
Decision 0079 and issue #122 define acceptance: first a real synthetic CUDA proof, then complete matched 23-segment runs, zero CPU constraint fallback, positive device constraint execution, all 69 files identical, and fresh CPU/Mac mutation suites. CUDA's installed sources truthfully report dirty provenance; exact generated/modified-source and binary inventories are retained. SIMD stays pristine. The patch only forbids fallback and caps nvcc compile jobs at two.
Initial offline checks: 50 focused tests pass, actual enrolled-statement extraction passes and the guard/resource patch applies to the original hash-pinned source archive. These are preparation checks, not device performance evidence. The selected resource plan is CPU build 3,600s, H200 2,700s, CPU verify 1,800s, plus 600s startup/cancellation per phase, four Rayon threads, a 200-GiB sampled RSS stop and a $5 per-run ceiling ($4.66 inspected timeout/startup/storage bound). Live prepaid balance/rates must be checked before launch. No retries, top-ups, private uploads or schedulers. Real-kernel and publication results remain pending.
Generated build succeeds; telemetry-framing failure retained¶
Frozen source 74bc2db33df5cf4d1e5a5b101b30fecba122960f builds successfully
in HF job 6adf6c9041e90535df7573fe: 1,815.943 seconds, $0.02 running-compute
estimate. It generates 77 kernels across 16 families, 206,910 IR instructions
and 247,496 CUDA lines. All 39 modified-source files are inventoried. Local
application of the pinned patch reproduces the retained guard/resource source
hashes. CUDA binary size is 90,971,384 bytes and native sm_90 is present. All
outputs were retained and hashed, and the build bucket deleted.
H200 job 6adf73cb3fb22a235dd1ebc0 fails after 1,111.887 seconds, with a $1.59
running-compute estimate. All 23 CUDA segment reports individually accept with
zero reference mismatches and zero explicit/automatic constraint fallback; their
summed proving timers are 51.788302807 seconds. All 69 raw CUDA files match the
completed SIMD inventory. However, telemetry parsing fails during the head's
command accounting before the final CUDA grid and fresh CPU/Mac verification,
so this remains a failed/incomplete pipeline run. Preserve all raw artifacts.
The completed telemetry file contains 2,153 well-formed records; the parser read
it concurrently at a partial record boundary. Outputs were fully retained and
the failed job's bucket deleted. These cost figures are not invoices.
The revised sampler admits only complete newline-terminated records within the sample window, drops leading/trailing partial fragments, and still rejects malformed complete records. Regression tests cover both partial boundaries and incomplete-only windows. All 51 focused tests pass. No prover, generated binary, model, witness, proof parameter or cryptographic evaluator changes. A new frozen run reuses the admitted generated build under the same $5 ceiling, with a $4.62 remaining timeout/startup/storage bound and no automatic retry. Recheck live credits/rates before launch. The complete measurement remains pending.
Accepted generated-constraint measurement¶
The corrected frozen source ef09216eabee044c24209f555df39e277ee6dfaa passes
in llm-gpu-oluavcvs, reusing the admitted generated build from the preserved
first attempt. Numeric evidence was saved before this interpretation in
evidence/llm-gpu-generated-constraints-2026-10-26.json.
All 29 runtime hashes match the frozen protocol. The generated-source commit
is 74bc2db33df5cf4d1e5a5b101b30fecba122960f; its 77 kernel shapes across
16 families, 39 modified-source files, patch, native architecture and binaries
are inventoried. Witness and interaction generation remain on the CPU.
The matched SIMD/CUDA summed proving times are 547.748838166 / 53.941293719 seconds: 10.1545 times faster in this pair. Complete segment commands take 679.487932103 / 155.888844427 seconds: 4.3588 times faster. The same-host SIMD grid checks take 85.347811110 / 84.317210196 seconds. All 23 segments pass, including the output head; both grids check 22 residual links and seven generated input equalities. All 69 file entries match each other and the historical CPU case. Each set has 737,858,641 serialized proof bytes. Peak sampled host RSS is 103,747,272,704 / 8,074,182,656 bytes. CUDA peaks at 69,293 MiB VRAM and 100% sampled utilization. Raw profiler counters show 1,710 GPU constraint evaluations, zero explicit/automatic CPU fallback, and 0.473164465 seconds total CUDA constraint evaluation versus 24.294546484 seconds SIMD. Timer categories can nest; do not sum them to reconstruct wall time.
H200 job 6adf7a4141e90535df75798e completes in 1,230.949 seconds. Fresh CPU
job 6adf7f6941e90535df757b86 completes in 392.362 seconds: two honest checks
pass in 77.607382583 / 77.395283685 seconds and all 14 mutations reject, with
14,343,761,920 bytes maximum child RSS and no weights/full witness. Native Mac
verification accepts two in 3.531456250 / 3.401666375 seconds and rejects all
36 mutations. Native compilation takes 215.142094666 seconds and its complete
verification worker 96.980634709 seconds. This is fresh-process verification
using the same SASH cryptographic implementation; Mac shape roots are compiled
enrollments rather than Linux runtime derivation.
The complete passing command takes 2,086.775471416 seconds with preparation and build reused. Its running-compute estimate is $1.76; including the first build and failed telemetry attempt, the generated-kernel campaign totals $3.37, not invoices. All four campaign jobs are terminal, outputs retained/hashed and buckets deleted. Both source identities and the failed attempt remain retained. The telemetry correction changes measurement I/O only; no proof/evaluator/input is weakened. There is one clean paired observation with shared OS caches and a fixture-warmed driver, not a latency distribution or measured full-fresh workflow.
Current gates pass 51 focused tests and make check: 2,454 passed, 50 skipped,
lint/type/spec/conformance/admission checks. Documentation links pass. The final
staged make package-check and exact-head hosted checks are recorded on the
delivery PR. No hiding, hardware, general-model or qualified-backend claim follows.
2026-10-27 — README GPU headlines and proof onboarding¶
Documentation gate from STATE: make the completed fixed-case GPU result legible
at the repository entry point. Hypothesis: adjacent timing boundaries and a
matched CPU/CUDA table prevent the historical CPU-only latency from obscuring
the generated-kernel result. Baseline: main 3fd97ee5191374a439db3cbbfd033322ee298569.
Mutable paths are README, the architecture/LLM/docs entry points, STATE and this
ledger. Acceptance: audited figures, working local links/anchors and Make target
resolution, local/distribution checks; no runtime, protocol or evidence changes.
The README now leads with 53.94 s CUDA proving and 10.15x matched speedup. Its table separates 155.89 s complete commands (4.36x), host/device memory, serialized proof size, verifier preprocessing, cached workflow time and compute estimates. One pair, CPU witness generation, failed attempts, public-witness scope and the separate historical CPU workflow remain explicit. Exact proof/caller boundaries, a composition diagram, prerequisites/costs/results for each quickstart and a source map replace repeated descriptions. CUDA source changes are distinguished from the unchanged upstream CPU binary in the architecture and LLM guides.
A local arithmetic check recomputes both ratios, rounded times, host GB and GPU
GiB from research/evidence/llm-gpu-generated-constraints-2026-10-26.json, verifies
identical serialized sizes and the passing/equality flags. All five Make targets
resolve with make -n; no workflow or paid compute is launched.
python3 scripts/check_docs.py passes paths/anchors; external URLs are not
checked by that command. git diff --check passes. Final local/distribution and
exact-head hosted receipts follow on the delivery PR. Numeric evidence, model,
witnesses, evaluation and runtime files are unchanged; no new benchmark is run.
Local make check completes successfully in the locked core environment; its
pytest totals and final package source identity are recorded on the delivery PR.
Optional training execution is left to the required train-extra CI matrix.
2026-10-27 — Publish the frozen trained classifiers on Hugging Face¶
Owner instruction: publish trained models with model cards and README references.
Decision 0080 records that explicit authority for original synthetic artifacts.
Baseline main: f6b5e07d97ebe62c908bd39d9422356c67bd1bb5. Gate: the exact
preselected seed-17 MLP/CNN used in the proof demo are public, with metadata/cards,
checkpoint/quantization/native exports, provenance/evaluation, immutable revisions
and fully hash-verified anonymous downloads plus runnable native examples. No
training, dataset, evaluator, threshold or prior evidence change is admitted.
Mutable paths: new native inference helper/tests, model cards/release guide,
README/docs entry points, decision, publication receipt, STATE and this ledger.
The retained completed demo is runs/readme-demo-2026-10-22/runs/demo-20261022T124608Z.
Both original full/resumed checkpoints are byte-identical. MLP checkpoint
requantization and CNN parameter requantization exactly reproduce existing pins
f8090161455ee56cdcd73eaab44437efa3b278a786d7c246491da5f1e1431a9c and
8de31ae0215178f5c74aa43e58dd63511bc3b6a881c2045cc96def9e149b8f0b. Regenerated
compact model JSON exactly matches the original proof inputs. All 100 retained
CNN test scores/decisions replay without fitting or selection. Current licence
review follows decision 0063: MIT covers original code/models/authored fixtures;
historical data labels/hashes remain unchanged and a notice accompanies release.
Publication is public under abdelstark/hallmark-byte-mlp at commit
05b2bdef120d3aea9133af042d15ed1a3a621aa9 and
abdelstark/hallmark-contextual-byte-cnn at
51461ea4401188a14997c5003ea2073ffb91e1ed, both tagged v1.0.0. Enumerated
MLP/CNN bundles contain 17/24 files and 185,096/579,781 bytes. HF model-card
metadata validation passes. All 41 released files are downloaded without a token
at pinned revisions and match every manifest SHA-256 and size. Four native
controls per model match from those downloads, including expected CNN context
contrast and block of the ordinary LLM prompt; MLP contrast scores are identical.
The receipt research/evidence/hub-model-publication-2026-10-27.json stores file
pins, metadata, examples and scope. No private paths, tokens, transcripts, setup
keys, upstream checkpoint or unrelated model is uploaded. No paid compute runs.
Native command scripts/predict_published_classifier.py reuses existing integer
exports/graph arithmetic, admits only the frozen quantization hashes, and rejects
modified weights, symlinks and inputs above 96 UTF-8 bytes. It labels output
native-inference-only, making no cryptographic proof claim. Seven offline
admission tests pass. Native MLP runs in core; CNN export requires the train
extra (NumPy), discovered during fresh core validation and documented before
publication. Draft validation failures remain local: an overbroad token-name
scan rejected a public source filename, corrected to token-value matching in a
new preparation directory; a transient docs gate rejected the not-yet-written
publication receipt. Final checks use new logs and a complete source snapshot.
README links, card-source mirrors and the release guide add immutable download
commands and separate the releases from upstream TinyLlama. Local/distribution
and exact-head hosted receipts are recorded on the delivery PR.
2026-10-27 — Public documentation site and model header badges¶
Owner request: for-the-badge styling, both published model badges in the header,
and a beautiful research/educational site deployed through GitHub Pages. Baseline
main is 2e5676cf6b11634c78846b75cdc95345142252b7. Acceptance: accurate existing
claims, safe tracked-file staging, functional navigation/search/educational
components, responsive visual review, strict/link/package/core checks, and a
verified hosted Pages deployment. Mutable paths are documentation, site assets,
new build/link helpers and tests, optional docs project, Make/Pages workflow,
sdist roots and these records; no model, evaluator or proof relation changes.
Zensical 0.0.73 is separately locked under website/. An early draft put it in
the inference lock; inspection identified that lock as part of native verifier
enrollment, so the draft was replaced before publication. Root uv.lock is
byte-identical to baseline. All 55 native verifier input pins match, yielding
unchanged identity 8567c911d686b13b09032b75e1fc041e6bbcc9407cb2755b60746112102b342a.
All 11 isolated docs distributions and their locked artifact hashes match exact
PyPI metadata, with licence/notice hashes in website/dependencies.json. Font
and KaTeX release archives pass registry SHA-512 integrity; local notices and
SHA-256 inventories accompany the self-hosted assets.
The source build selects tracked Markdown, public evidence JSON, documentation images and explicit assets. Tests prove untracked notes/runs, non-asset keys, symlinks and unowned staging directories cannot enter/replace output. Canonical Markdown is not rewritten; staged copies map root README/source links and the GitHub/site repeated-heading-separator difference. The generated site includes a source/hash build manifest. Ten focused tests pass. Initial template and one historical anchor failures were corrected before strict builds passed; logs are retained locally. The GPU guide now presents the generated-constraint result before the retained fallback baseline, preserving numbers and the old anchor.
Browser review covers 1280-pixel desktop and 390x844 mobile layouts, with no
horizontal overflow, plus mobile article tables/navigation, five rendered KaTeX
expressions, a rendered Mermaid flowchart, GPU search results, light/dark styling
and LLM/classifier proof-selector states. Visual testing found literal Markdown
in nested proof panels; it was fixed and checked as rich rendered content.
The browser connection reset during the final clipboard check, which remains
unclaimed in the QA record. Direct Shields SVG auditing returned HTTP 403; five
standard styled URLs and their destinations are checked, with rendered/hosted
receipts on the delivery PR. No external URL success is inferred from local
link checks. The QA record is research/evidence/documentation-site-2026-10-27.json.
GitHub Pages was absent and is now enabled as workflow deployment with HTTPS
at https://abdelstark.github.io/hallmark/. Only main can deploy; PRs build and
validate. Actions are pinned to verified commit SHAs and use ephemeral standard
Pages/OIDC permissions. No personal deploy token, analytics, custom domain,
research scheduler, paid compute or model download is added. Local core/site/
distribution checks and exact-head CI/deployment identities are recorded on PR.
The existing test_make_keeps_validation_as_default caught an initial target-order
regression in the draft Makefile. Moving docs targets after check preserves
plain make; the focused demo/staging suite passes before the final core rerun.
This failure and rerun use separate logs.
2026-10-27 — Documentation copy edit¶
Owner request: apply no-ai-slop to the docsite. Baseline main is
22ccea29c9ee8448c01322f05154071b213da988. Hypothesis: replacing slogans,
metacommentary and compressed protocol jargon with named computations and
checks will make the entry pages easier to read without changing their claims.
Mutable paths: homepage Markdown/template, site metadata, workflow chooser,
learning chapters, research evidence map, the homepage presence assertion and
these records. Proof code, model/data/evaluator files, dependency locks, numeric
evidence and dated research decisions are frozen.
The pass removes stacked fragments from the hero/caption, names the fixed TinyLlama case and classifier workflows, explains caller enrollment in ordinary language, and cuts passages that repeat their own lesson. Benchmark stage boundaries and privacy/provenance limits remain explicit. Renamed sections keep their original anchor IDs. The homepage checker now checks its heading element rather than a slogan. No new copy-specific tests or runtime features are added.
The edited source is reviewed against the skill's eval.md: meaning, technical
detail and document order preserved; generic framing, decorative emphasis and
repeated contrasts cut. Local validation and exact-head hosted/deployment
receipts are recorded on the delivery PR. No new research measurement is run.
2026-10-27 — README proof-flow diagram¶
Owner request: include the docsite homepage diagram in the README. Baseline is
beb8f96e50b2ad5b03a204193b42d0994f3f1151. The supplied SVG matches the tracked
website/assets/proof-path.svg byte for byte (SHA-256
3c6487c4bbe111e2e44ae52af03a595247ed248cd0fdedd967c8922292e62a88).
Reuse that asset below the README introduction and link it to the detailed LLM
statement. Mutable paths are README and these records. Acceptance is a working
image/link on GitHub and in the staged documentation; no asset, proof, model,
measurement or runtime change. Validation and publication receipts follow on PR.
2026-10-27 — Shared terminal and browser workbench¶
Owner goal: full TUI and web demo apps, shared components, extensible proving
workflows, usable monitoring and a clean interface. Baseline:
675c99af66067d129dac30d6acc6612f4e392e68. Hypothesis: one workflow catalog,
result adapter and persistent controller can expose the real existing drivers
without duplicating proof logic or changing enrolled verifier inputs. Mutable
paths and acceptance criteria are in decision 0082. Proof drivers, model/data/
evaluator files and the inference lock remain frozen.
Raw observations precede this account in
evidence/demo-workbench-2026-10-27.json.
All raw runs and failed validation attempts remain ignored and local under
runs/; no credentials, checkpoints, proving keys or financial account records
are published. The app environment is independently locked: Textual 8.2.8,
psutil 7.2.2 and uv 0.9.18; isolated browser QA uses Playwright 1.64.1 and
installed Chrome. Twenty third-party distributions have a licence/hash review.
Real interface acceptance:
- TUI keyboard launch,
runs/e2e-ml7ggpfp: 9 classifier proofs verified, 18 substitutions rejected and 405 arithmetic cases checked. The controller crashed/recovered during proving and observed the same driver PID/creation time without retraining or relaunching it. - Web launch,
runs/llm-verify-q8adsd4t: fresh native verification of a retained public bundle; 23 proofs, 22 residual links, 2 honest checks, 14 baseline, 2 enrollment and 20 boundary mutations rejected. - Web launch,
runs/llm-e2e-2d9zyzcr: fresh TinyLlama preparation and 23 real proofs, 22 residual links, 2 fresh honest CPU checks and 14 rejected mutations. Result SHA-256:02ec4242460a1e8bef06b57825051be036023e63a6a21f3d5c18c2bdba7cd0d7. The complete monitored workflow takes 2,132.35 seconds; running-compute estimate is $0.31, not an invoice. Live startup-inclusive admission bound is $2.27 under a $2.50 ceiling. All three jobs were independently observed completed via the provider API; outputs retained and buckets deleted. A controller-only restart kept the same live paid driver/job rather than launching another.
The original $2.10 CLI estimate covers worker timeouts, but not the complete provider creation windows. The app adapter reserves another 600 seconds per phase before invoking the unchanged driver. No job started under the rejected $2.10 startup-inclusive admission. No automatic retry or top-up was introduced.
The final 52 app tests pass: subprocess success/failure, cancellation and local
escalation, recovery, transient observation/identity errors, nested cloud jobs,
owned storage, split-write redaction, result contracts, artifact identity/hash
checks, Host/Origin/session-token checks, terminal keyboard actions, browser
launch/reload/export/mobile behavior and keyboard tabs. Monitoring errors no
longer cancel a live driver. Early harness failures are preserved: an eval-based
browser wait conflicted with CSP; a process-class replacement broke psutil's
own type checks. Locator assertions and a method-level fault injection fix the
harnesses without weakening production checks. The early metadata/worktree
namespace collision was corrected to runs/workbench/; the controller now
refuses unowned nonempty storage rather than overwriting it.
Rendered desktop/mobile and 120×42/80×24 terminal views were inspected, including
real proof results and policy decisions. make check passes 2,398 tests with
123 optional skips, plus spec/software-conformance/admission checks. Reproducible
package, app wheel and strict documentation checks pass. Optional train/guard
coverage and exact-head hosted/merge/deployment receipts are recorded on PR #129.
All 55 native verifier source files, program identity and inference lock are
unchanged. No fresh GPU or composed campaign is claimed for this app acceptance;
the composed driver retains its 32 GiB free-storage prerequisite. No new backend
qualification, hardware attestation, hiding or model-quality claim is made.
2026-10-27 — Circle classifier implementation: first complete acceptance¶
Selected epic #112 and its ordered #113–#117 gates. Baseline is
0b7fcf256e2b1da7f89578014bcc86e6737fc3f5; hypothesis, mutable paths, exact
relation and frozen repetition/resource protocol are in decision 0083 and
configs/circle-classifiers/. No arithmetic, model, threshold, evaluator,
dataset or historical result is replaced.
The backend uses direct STWO 2.2.0, release revision
289c20de80b7c7f508de9c46151fb81dae404154, Apache-2.0, with
nightly-2025-07-14. Registry engine/framework source files match the pinned
release archive. The lock/review test caught five prover-only dependencies
missing from the initial verifier-feature licence inventory. The amended review
now covers all 109 third-party locked packages, including their selected licence
alternatives/checksums; that amendment is retained in dependencies.json.
No root Python dependency or native LLM verifier source was changed.
Failed engine hypotheses remain in runs/circle-api-fixture-v1 through v3
and the diagnosis directories. A trace smaller than the FRI last layer failed;
the cubic selector/batched-lookup version then exposed a decommitment-domain
mismatch and an OODS constraint mismatch. The accepted AIR adds an explicit
product column and one LogUp entry per interaction column, making every
constraint quadratic. The wiring/constants/fanout are preprocessing, not
prover-controlled multiplicities. No failed version is reported as an accepted
proof. The invalid Mac RLIMIT_DATA diagnostic also remains a failed attempt.
The exact SHA-256 circuit matches hashlib across block boundaries. Its 32-bit words use limbs; variable-length classifier inputs use constrained canonical CBOR, masks, zero padding and SHA length padding. The rule AIR maximum-input fixture passes fresh proof verification, with placeholder context/classifier values explicitly excluded from a composition claim.
The acceptance evidence then records two complete caller-bound runs:
runs/circle-classifier-acceptance-v1: nine real proofs and nine fresh proof-only verifications; 63 direct public-input mutations, 18 cross-request substitutions and three invalid rosters reject. All MLP intermediates and CNN node outputs match native execution. Complete acceptance wall time is 97.88 s.runs/circle-classifier-acceptance-v2: the same gate passes after adding stage timing records and explicit free-space checks, and sizing twiddles to the common commitment domain. Complete acceptance wall time is 96.34 s. Twelve preprocessing/proving workers emit component timing records.
Both runs retain their raw numbers, source/binary/program/proof hashes, caller pins written before proving, witnesses, proof files, logs and failed controls. Their matching source/binary snapshots are retained separately. They are acceptance trials. Backend identities differ when their source enrollment changes; native model identities, weights and thresholds remain unchanged. The caller still runs native inference.
The first focused Python gate passes 72 tests; the Rust gate passes three tests, including an actual arithmetic proof roundtrip and trace constraint assertions. The new lock/licence admission gate passes seven tests after the inventory amendment. These counts precede final core/distribution/hosted checks.
Storage preparation leaves the busy uv cache unchanged. Thirty-six byte-identical circuit/key copies in the owned acceptance/preparation directories are replaced with verified APFS clones, preserving regular independent inodes, complete bytes, paths and timestamps. The retained local receipt records every source/target hash; observed free space increases by 3,140,980,736 bytes. No model, proof, failed run, Cargo registry or source is deleted. No paid compute is used.
Remaining acceptance work is the explicit invalid-witness/maximum-length gate, same-request LLM composition and final local/distribution/hosted delivery. The epic and all child issues remain open until their complete criteria have evidence; no hardware, hiding, safety quality or backend qualification is claimed.
Circle adversarial gate¶
runs/circle-constraint-conformance-v1/result.json records 210.246452 s,
three new full-96-byte proofs, three fresh verifications, 21 invalid witnesses
rejected by real STWO constraints and one omitted dense edge rejected by the
independent program pin. Ten byte/CBOR/chunk/UTF-8 boundary cases retain exact
native parity or reject overflow. This resolves the invalid-witness gate above.
Public hashes and counts are in the classifier acceptance evidence.
The final small gate-AIR run (runs/circle-api-final-v1) freshly verifies one
165,306-byte proof and rejects all six API mutations, including noncanonical M31.
This adds no classifier-support claim to that small fixture.
The acceptance review finds a model-domain limitation: the frozen MLP is
always-allow for all inputs up to 96 bytes.
Store research/evidence/mlp-domain-bound-2026-10-27.json before prose. Independent
interval relaxation with exact rational ties-to-even rounding gives score
[−5600,−2338] at threshold 0; the larger byte-count relaxation includes every
valid UTF-8 input. Four analytical tests pass, including exhaustive two-byte
small-domain combinations. This is not a tight bound, held-out evaluation or
cryptographic proof. Full MLP arithmetic remains constrained; actual workflow
block controls come from rule/CNN. Issue #131 tracks a separate useful-model
experiment; it is not a dependency or change to #112's frozen comparison.
README, guide and local model card disclose the limitation. Original Hub tags
and artifact bytes remain unchanged; a documentation-only model-card clarification
will be published separately with new manifest/revision receipts.
Model-card clarification published as Hub revision 15c8c4a1a5c86bf14cfef23e57c72c47e1de8a2f.
Only README and its manifest file entry change; every other Git blob and all
weight bytes are unchanged. Anonymous immutable downloads match both new file
hashes; the original v1.0.0 tag still resolves to
05b2bdef120d3aea9133af042d15ed1a3a621aa9. The original publication receipt and
tags remain immutable. New revision/file/authority hashes are in
research/evidence/mlp-model-card-clarification-2026-10-27.json.
The MLP card clarification additionally matches an anonymous download of
model.json byte hash 463860eeaed927cba398a8b05fd3059e9f8a4bbda103a331180e0bf755ed91b8
to the exact exported model used in the bound calculation. Current local
make check passes 2,569 tests/51 optional skips; package and strict docs checks
pass.
2026-10-27 — Circle validation and reproduction commands¶
Exact commit d02e32d4461d35ad1c097cba785d88d36e1d6eaf passes all 18 hosted
research checks
and documentation checks.
The strict local docs build passes 162 pages and 12,405 local links/assets.
PR #130
records the receipt.
The guide adds the already-implemented adversarial conformance commands, their saved caller-pin requirements and exact passing
counts. Both CLI help checks pass; the updated strict docs build passes
162 pages and 12,421 local links/assets (runs/circle-docs-v8.log).
No proof relation or model changes.
The architecture guide additionally maps all three proof paths and the Circle
classifier flow, including caller enrollment and the 100-to-103 public transport
layout. Its strict build passes 162 pages and 12,425 links/assets
(runs/circle-docs-v9.log). Rechecking the 55 native LLM source pins yields the
unchanged identity
8567c911d686b13b09032b75e1fc041e6bbcc9407cb2755b60746112102b342a.
Local free storage fell to about 4.9 GiB. A second targeted cleanup preserves each executable by hash and removes only
12 completed native-verifier target caches through Cargo. Completed result and
worker-binary hashes match before/after; open-file checks pass. The retained
receipt is runs/completed-native-cache-cleanup-v2/receipt.json: free bytes
5,245,190,144 → 9,975,779,328. Failed builds, source, proofs, models and registries
are retained. This changes local retention headroom only.
Six focused gate-equation fixtures check thresholds 0 and −3733 at score
offsets −1/0/+1. Equality blocks; mutations of either public threshold limb
or the verdict reject. All 56 integer tests pass in 0.06 s
(runs/circle-threshold-tests-v1.log). These are arithmetic conformance checks,
not new proofs or reachable-score claims for the frozen models. The strict
verdict implementation and backend identity are unchanged.
The subsequent locked make check passes 2,580 tests with 51 optional skips
in 106.44 s, plus spec/software-conformance/admission checks
(runs/circle-make-check-v5.log). Strict docs pass 162 pages and 12,425 local
links/assets (runs/circle-docs-v10.log). Optional train/guard coverage remains
a hosted gate; this adds no proof timing or backend qualification.
Exact-head hosted research and docs checks pass at 42d6139e792968f1365c30a5bfb05968f90dc877
(research,
docs).
2026-10-27 — Circle retained-LLM join and Make reproduction¶
runs/circle-postflight-v1/composed/result.json freshly verifies the three Circle
proofs plus 23 retained SASH proofs and 22 residual links, rejects all 16 altered
binding/coverage cases and finishes in 11.133048 s including hashes/host work.
Verdicts [1,1,0] establish integrity with policy release false. Host joining is
not recursion, and no new LLM proving occurs. All 55 native verifier sources
retain identity 8567c911d686b13b09032b75e1fc041e6bbcc9407cb2755b60746112102b342a.
After disk headroom recovers to 13 GiB, the actual make e2e
entrypoint runs with the admitted original classifier preparation and saved
enrollment pin. It builds both locked Rust targets, creates
runs/circle-e2e-ui9qpl46 and passes nine generated proofs, nine fresh checks,
63 direct mutations, 18 request swaps, three roster rejections and full
intermediate parity. Driver wall time is 97.808485 s excluding compilation;
maximum worker RSS is 20,231,782,400 bytes. Caller Circle root is again
56c29c74a4efa0443bff84cc516349b0369360e2feaf347b871b860b06cd635e.
The separate command evidence is
research/evidence/circle-make-e2e-2026-10-27.json, SHA-256
6c1a9adbf0540ce066c821c432fc397a12932521575ec7ea6c57dfd3e6b531c0.
This confirms the named Make workflow. All 55 native LLM source pins remain
unchanged.
2026-10-28: Circle STARK default¶
Owner request: use Circle as the default and base case for docs, headlines and
demos. Decision 0085 freezes baseline
f613c611cefa9037320082b73777f26c9ec45990 and limits mutable orchestration/app/docs
paths. Numeric evidence stores source,
model, enrollment, binary and artifact hashes before this prose. Original model
weights, arithmetic, thresholds, data and evaluators remain unchanged; both
backend/compiler and 55-file native LLM identities match baseline.
Actual bare make e2e freshly trains/exports and verifies nine Circle proofs,
rejecting 63 changed public statements, 18 request substitutions and three
invalid rosters, with native trace parity. Wall time is 320.68 s. Default retained
make e2e-composed freshly verifies 23 SASH proofs and 22 links, passes both honest
checks and rejects all 36 LLM mutations, creates three new Circle classifier
proofs and verifies them again at the outer host join. All 16 binding checks
reject; [1,1,0] means integrity passes and the synthetic policy blocks. Wall
time is 266.29 s including native compilation. A real browser launch of the
primary Circle workflow reuses only frozen model exports and completes nine new
proofs/rejection checks in 97.94 s; desktop/mobile results inspected, no page
errors. These are local acceptance observations, not repeated performance samples.
Both clients select Circle first. Both public HF model cards now describe Circle; only README/hash entries changed, all weight files and original release tags remain unchanged. New cloud jobs/spend: zero. Public assets are original MIT synthetic fixtures/models; no new data source or private material is used.
Local gates: 2,588 core tests pass with 50 optional/platform skips, strict typing, lint/format, software/spec/admission checks; all 56 app tests pass; strict docs and core wheel/sdist/installed checks pass. Initial scratch harness, app ordering and three stale-anchor failures are retained in the acceptance record. Exact-head CI, merge and Pages delivery are recorded on the change's PR. Default selection adds no independent security qualification, hiding, hardware, recursive proof or policy-quality claim; the MLP remains always-allow in this fixed input domain.
2026-10-28: public Circle reference proofs and stage accounting¶
Owner request: replace incomplete headline accounting with actual proofs,
all stages, verification/size/resource details and a public Discussion.
Decision 0086 fixes baseline
9bea024dc7d4604f17d53c915b64bc4a23c408b5, frozen semantics, mutable observer/
publication paths and local resource limits. The numeric receipt
precedes prose and retains all derived rows, source identities, artifact hashes,
public asset digests, anonymous downloads and verification receipts.
C1 is a predeclared first fresh local timing observation, not a selected best run: 302.96 s observed controller, including frozen training and initially empty Rust target builds, excluding uv bootstrap. Nine proofs total 18,063,779 JSON bytes; final STWO calls 24.061 s, complete prover processes 50.028 s, fresh proof-only verifier processes 0.807 s. Each internal stage and all per-proof rows are retained. The proof campaign rejects 63 changed public statements, 18 substitutions and three invalid rosters. Source/core/compiler identities and both selected integer exports match baseline. No fitting, threshold or evaluation-selection change.
L1 is the retained 2026-10-26 H200 pair. Raw per-segment logs reproduce 53.94 s CUDA session timing, 155.89 s complete commands and 84.32 s separate CPU grid verification on the H200 host. Source audit establishes that interaction work is nested inside the SASH session timer; classifier final-call timing uses another boundary. Fresh L1 preparation and isolated serialization/transfer totals were not recorded, so remain null. The original build-job running interval is separate. J1 is retained LLM composition, not new GPU proving. Never pool the machines or component runs into one end-to-end latency.
Discussion #133 and the artifact release publish the actual 23 LLM proofs, three composed classifier proofs and nine classifier controls, caller statements/enrollment and 988 exact supporting files. Known credential values and credential/private-path patterns were checked without printing values. Checkpoints, witness archives, executables, caches and private business material are excluded. Original Hallmark metadata/fixtures/exports are MIT; no new data source or checkpoint upload. Anonymous downloads of all four assets match byte counts and SHA-256. Proof sizes and gzip/evidence-bundle sizes are reported separately.
Fresh verification of the exact archive set passes in 219.25 s including native
build, two complete honest LLM checks, all 36 LLM mutations and 12 fresh Circle
classifier checks. First helper attempt passed LLM checks then failed before
classifier launch because input archives consumed its generated-output quota;
that failure is preserved and the output quota now covers only verification work.
The source-pinned utility is published at 5fe56cdb459366850e83162769f8373451131b7d.
New paid jobs: zero. No hardware, hiding, independent qualification, useful policy
or equal-security performance claim. Final tests/package/docs and exact-head
CI/merge/Pages receipts are recorded on the delivery PR.
2026-10-28: fresh GPU proving repeat and classifier CUDA preparation¶
Decisions 0087 and 0088 freeze the owner-requested LLM repeat and separate classifier GPU candidate. Historical L1/C1/J1 numbers and default Mac classifier sources remain unchanged. The LLM build is active; no new device performance result is accepted in this preparation milestone.
Local evidence: runs/gpu-repeat-make-check-v1.log, 2,611 passed / 50 skipped;
runs/classifier-gpu-source-check-v2.log, locked CUDA-feature Rust check passed;
runs/port22-rule.log, a real 1,935,860-byte CPU proof matches C1 byte for byte;
runs/classifier-gpu-input-reconstruction-v2.log, all nine native cases match.
Caller/native and upstream pins are in decision 0088 and ignored source/input
manifests. CPU backend identity remains c035913b3196640797273bb0c22c2e69b09dee2a790d128be51287a6351ded7d;
LLM verifier remains 8567c911d686b13b09032b75e1fc041e6bbcc9407cb2755b60746112102b342a.
Preserved preliminary failures: a relative runpy launcher path fails the LLM
controller before protocol creation or any job; the absolute-path launcher fixes
that without changing the controller. Nitroo's direct older dependency fails
API/proof-format compilation, so the candidate uses an additive 2.2 overlay.
The first local classifier reconstruction lacks optional NumPy and stops before
any upload; the locked train-extra environment passes. Overlay compiler/lint
iterations are retained. GPU correctness, performance, distribution and backend
qualification do not follow from these local checks.
2026-10-28: accepted classifier CUDA benchmark on A100¶
Decision 0089 records hardware fallback from 0088 after the H200 allocation stayed queued. The owner's explicit continuation approval permits a null numeric credit check under the unchanged USD 5 combined ceiling. Live prices and no-active-job checks precede each attempt. Mutable paths are optional GPU transport/controller, native target, worker machine/RSS guard, one regression, reports and evidence. Classifier AIR, stock core/verifier, models, thresholds, corpus and proof parameters are frozen.
The numeric record was saved before prose. All 54 actual proofs pass fresh stock Linux and Mac verification; all 378 public mutations reject. Both positive/invalid arithmetic fixtures gate the run. All 27 CUDA classifier AIR calls execute with zero CPU AIR fallback; all proof bytes match the nine published C1 controls. Three same-host pairs on A100 give median nine-command sums 92.406 / 22.216 seconds (4.16x), final STWO calls 44.310 / 7.550 seconds (5.87x). Stage boundaries and all samples precede interpretation in the JSON; the report labels reused training/preparation, host work, memory and uncertainty.
Commands: scripts/prepare_classifier_gpu.py --out runs/pr135-a100-source-staging
and scripts/run_classifier_gpu.py --prepared runs/pr135-recovery/reference/classifiers/controls --native-preparation-pin 3eab49c2a4a97a3e7d4daaefbbdc4c68a09ae7ba2d9a0d438c803cc7efcb13cd --source runs/pr135-a100-source-staging --gpu-flavor a100-large --owner-approved-credit-check-exception --build-run runs/classifier-gpu-399bl9ec,
each via locked uv run with llm/train extras as applicable. Raw run:
runs/classifier-gpu-sm11ii3z. The independent byte/statistics audit
uv run --locked --extra llm python runs/pr135-recovery/report-final.py checks
all 1,612 retained file hashes / 133,042,187 bytes and writes the public numeric
record, including its own source hash and all 54 raw measurement rows.
Preserved failures: separate LLM repeat timeout; classifier H200 allocation canceled before start; first A100 worker PermissionError on bucket executable. Corrected staging copies/checks binary bytes and sets local executable mode. Its regression passes; retained-build reuse requires source-content inventory, native target, completed build receipt and every binary hash to match. All five classifier buckets are deleted. Rounded running-compute bound USD 0.37, not an invoice; no new LLM measurement or backend qualification is inferred.
Validation after staging: make check, 2,613 passed / 50 optional-platform skips;
focused GPU/controller suite, 46 passed. Final make check passes 2,613 tests / 50 skips in 153.61 seconds; core
packaging and strict docs gates pass (170 pages, 13,442 local links/assets).
Exact-head hosted receipts are tracked on PR #135. Failed package gates remain local.
2026-10-28: generated documentation figures and performance guide¶
Documentation only; no measurement, proof relation, model, threshold or evidence
record changes. scripts/render_figures.py (stdlib, no network) renders 14 SVGs
into docs/assets/figures/ from circle-reference-2026-10-28.json,
classifier-gpu-benchmark-2026-10-28.json, llm-e2e-reproduction-2026-10-23.json and the fixed response-case.json.
tests/test_figures.py fails on any stale figure. Chart palettes pass the
categorical colour validator (lightness band, chroma floor, CVD separation)
against both the light and dark surfaces.
The new performance guide names each run's machine and
timer boundary and lists unmeasured quantities. Two derived values are labelled
as such: L1's per-segment remainder (complete command minus trace, session and
inline verification: 22.28 s SIMD / 35.41 s CUDA) and a sensitivity check
showing that segments L0–L2 take 278.78 s of the 547.75 s SIMD session total
(6.39× summed ratio without them, against the 10.15× headline). The record does
not isolate why those SIMD segments are slower. docs/learn/one-request.md no
longer describes the former backend as the default. The unused, outdated
docs/assets/classifier-pipeline.svg is removed.
Validation: make check passes 2,616 tests with 50 optional skips; strict docs
build passes 171 pages / 13,804 local links and assets; make package-check
passes with uv 0.9.18 after the system uv 0.13.6 stops the first attempt.
2026-10-29: Circle STARK as the only classifier backend¶
Owner request: remove the former classifier proof backend and every workflow that
depended on it. Decision 0090
records the scope. Baseline b7b8376b584836f29066708b663006c2cae7328e is the last
commit containing the removed sources; history remains in git.
Removed: its Go circuits and separate Python verifier, the matched comparison
benchmark/report/archive scripts and plan, hallmark qualify and its protocol
schema, the backend-selection Make dispatch, the proof-artifacts CI job, the two
dependent app workflows, their tests and the backend-only decisions, proposals
and evidence. Ledger entries that only described that work are removed; mixed
entries keep their Circle and LLM content.
Kept unchanged: Circle circuit/compiler sources and Rust engine (backend identity
c035913b…1ded7d), model exports, arithmetic, thresholds, the rule program
records (moved to fixtures/rule-program.json, same classifier and program
identities), caller enrollment pin, request contracts and the remaining numeric
evidence receipts.
Changed identity: the native LLM verifier inventory lists only the 17 files the
verifier loads; its identity becomes 5ba4d1f4…6bda3e09. Signed LLM enrollments
bound to the earlier identity need a new owner signature, and the
circle-reference-2026-10-28 release verifies from a checkout of b7b8376,
whose LLM verifier identity matches its manifest. The shared worker
sandbox is hallmark._worker, with cost records hallmark.worker-cost/1.
Pass criteria: make check and train/app/docs gates pass, and bare make e2e
produces nine freshly verified Circle proofs with the same caller enrollment pin.
Acceptance results are recorded in the decision and its PR. No security, hiding,
hardware or qualification claim changes.
2026-10-29: audit fixes¶
Owner request: fix every finding of the whole-repository audit, with dedicated
issues (#138–#158, tracking #159) and strict validation on main. Six PRs
(#160–#165) were split by file ownership; this record's PR adds
decision 0091, the dependency removal and
the numeric validation.
Hypothesis: every finding is fixed without changing the Circle backend identity,
enrollment pin, models, thresholds or proof relations. Pass criterion: all local
gates, a fresh make e2e, the documented conformance command, native
verification of the published LLM proofs, and hosted CI.
Result on main 1bc6b72: make e2e 253.87 s, nine fresh proofs and all 84
rejections; conformance 187.60 s, ten boundaries and 22 constraint rejections;
native LLM verification 203.59 s, all 38 mutations rejected with identity
39d2ee50…5879ef. Circle backend identity c035913b…1ded7d and pin 56c29c74…
are unchanged. Retained failure: one stale local mypy cache error, cleared.
The CLI LLM ceiling rises from $2.10 to the workbench's existing $2.50 because the
startup-inclusive worst case is $2.27. No cloud job or spend.
2026-11-03: FlashProve verifier and H200 proof of the synthetic small kit¶
Hypothesis: upstream FlashProve at 9e2bdd8, unchanged, verifies on the Mac and
proves its synthetic small kit on an HF H200. Pass criterion: both stored proofs
are accepted against caller-held values; all 99 upstream altered sets are
rejected; a fresh H200 proof whose configuration and request match the pinned
kit is accepted by Linux and Mac cp_verify; six altered copies are rejected.
Result: passed (decision 0093,
evidence). Mac suite 97.6 s. Build 1,520.5 s
on cpu-xl. H200 example 22.3 s wall; proof 4,145,980 bytes; Mac check 0.97 s.
Retained failure: the first prove job exited 127 (nvidia-smi not on the worker
account's PATH) before any commit or proof. Spend bound USD 0.78, not an invoice.
Synthetic weights only; no hiding, security-level or served-model claim.
2026-11-04: SASH and FlashProve comparison¶
Hypothesis: the recorded SASH L1 and FlashProve H200 runs can be set side by side
with every figure traced to an evidence path. Raw cp_verify re-timing on the Mac
also settles decision 0093's unrecorded verifier timings. Pass criterion: every
cp_verify call returns YES, every recorded figure resolves in checked-in
evidence, and no like-for-like prover claim is made.
Result: passed (comparison,
evidence, run
runs/prover-comparison-tdnrr1rm). Medians of five calls on an Apple M4 Max:
small stored proof 0.62 s, fresh H200 proof 0.63 s, full-size proof 7.88 s
(peak 4.23 GB). Decision 0093 had said 1.0 s, 8.8 s and 4.2 GB; it now cites the
check_proof.py timings (1.45 s, 7.99 s) and this re-timing. The FlashProve
decision is renumbered 0092 → 0093 because #181 merged its Slurm decision as
0092 first. No spend.