Skip to content

Research state

Updated: 2026-11-04. Phase: MIT-licensed research software in a public repository, with dedicated classifier and quantized LLM proofs, not a deployed or fully verified hybrid system. AI-assisted code and research initialization.

SASH and FlashProve comparison — recorded

2026-11-04. SASH and FlashProve compared sets the two LLM provers side by side from recorded runs, with a source pointer for every figure (evidence). make compare-provers re-timed raw cp_verify on the Mac, five calls per proof: 0.63 s for the fresh H200 proof, and 7.88 s at a 4.23 GB peak for the stored 1,000-token Maverick-shaped proof. No paid compute and no new proof. The two backends prove different statements on different models, so the comparison claims no prover speed ratio. Decision 0093 now cites these timings in place of unrecorded ones.

Slurm + Apptainer profiles — merged, proposed, not run

2026-11-03. PR #181 adds Slurm guide templates and CUDA 12.8.1 and 12.9.1 Apptainer profiles. Decision 0092 is still proposed. No cluster run, evidence record or new build identity is recorded. The FlashProve decision was renumbered 0093 to resolve the clash.

FlashProve backend — local verifier and H200 proof accepted

2026-11-03, merged in PR #182. At the owner's request, decision 0093 adds Anthropic's FlashProve (pinned 9e2bdd8, v0.16.0, Apache-2.0, unchanged) as an external backend driven by command-line adapters, as SASH is. make verify-flashprove builds cp_verify on the Mac. It accepts both stored upstream proofs against caller-held values and rejects all 99 of upstream's altered sets, in 97.6 s. make e2e-flashprove built the prover on HF cpu-xl in 1,520 s, then committed to and proved upstream's synthetic small kit on one H200. The proof covers 48 checked tokens plus the next token and is 4,145,980 bytes. Linux and Mac cp_verify accepted it, and six altered copies were rejected. A first prove attempt failed before proving: the worker account lost the driver PATH. It is retained. Computed spend bound USD 0.78, not an invoice (evidence).

FlashProve proves MoE models only, so the TinyLlama LLM path stays SASH. No Maverick or other gated weights, weight-hiding, security-level, qualification or served-model claim. No existing identity, enrollment, model or threshold changes.

Follow-ups after the audit

2026-10-30. #168 re-sends a workbench Stop that the driver ignored during its first 0.2 s and lengthens the installed-package check timeout for CI; #167 adds agent conventions in docs/agents/. No protocol, identity or evidence changes.

Audit fixes — merged and validated

2026-10-29. The owner requested fixes for every finding of the whole-repository audit (#159); PRs #160–#165 and decision 0091 record them. The audit found no wrong acceptance in the Circle path, wire decoder or signatures. Paid LLM ceilings now count the HF startup window once (CLI and composed paths use the workbench's $2.50; worst case $2.27); DELETED ends polling; GPU runs check local verifier prerequisites first; the one-off credit-check exception is gone. Workbench Stop no longer aborts driver cleanup, and recovery, receipt, quit and preflight bugs are fixed. Registry enrollment enforces successor rules; workers may signal only their own sandbox; the verifier time limit applies again; CNN admission rejects possible int32 overflow. Worker, CNN-circuit and Rust verifier rejection tests are in CI. The native LLM suite adds two statement mutations that the compiled verifier itself must reject (40 cases, 38 rejections).

Numeric validation on main 1bc6b72: make e2e 253.87 s with nine fresh proofs and all rejections; the documented conformance command 187.60 s with 22 constraint rejections; native verification of the published LLM proofs 203.59 s with all 38 mutations rejected; all local gates pass. Circle backend identity and enrollment pin are unchanged. The LLM verifier identity is now 39d2ee50…5879ef; signed LLM enrollments need a new owner signature. pytest-cov is removed under a new dependency review.

Single classifier backend — removal of the pairing backend

2026-10-29. At the owner's request, decision 0090 removes the former classifier proof backend, its comparison benchmark, hallmark qualify, the proof-artifacts CI job and the two app workflows that depended on it. Circle STARK (STWO) is the only classifier proof backend; make e2e and make e2e-composed run the Circle drivers. The last commit containing the removed sources is b7b8376b584836f29066708b663006c2cae7328e; history remains in git, and this file no longer describes the removed work.

Circle circuit/compiler sources and engine identity, model exports, thresholds, rule program records (now fixtures/rule-program.json) and the caller enrollment pin are unchanged. Remaining evidence receipts were pruned of the removed backend's entries only; their original bytes are at b7b8376. The native LLM verifier inventory now lists only the 17 files it loads, so its identity changes; statements below that the 55 LLM source pins were unchanged describe their own changes, not this one. Signed LLM enrollments bound to the earlier identity need a new owner signature, and the circle-reference-2026-10-28 release verifies at its own tag. The shared worker sandbox is hallmark._worker. Acceptance is recorded in the decision and its PR. No security, hiding, hardware or qualification claim changes.

Classifier CUDA benchmark — accepted A100 experiment

PR #135's optional CUDA classifier prover passes decision 0089. The report and numeric evidence retain all 54 proofs' measurements, 378 rejected public-input mutations, 54 fresh Linux checks and 54 fresh stock Mac checks against unchanged C1 caller inputs. All proof bytes match C1. CUDA executes 27 classifier AIR calls with zero CPU AIR fallback. This is fixed-case research, not backend qualification.

On one A100-SXM4-80GB / Xeon 8275CL host, with four Rayon threads, median nine-proof command sums are 92.406 s CPU / 22.216 s CUDA (4.16×) over three pairs. Median summed final STWO calls are 44.310 / 7.550 s (5.87×), inside those commands. Peak sampled Rust RSS is 19.456 / 0.624 GB; sampled VRAM peaks at 18.25 GiB. Per nine-case suite, proof JSON remains 18,063,779 bytes. Host witness and interaction generation remain CPU work. Training/preparation and native builds are reused; these are not fresh full-workflow or Mac speedup measurements.

The audit rehashes 1,612 retained files / 133,042,187 bytes; five temporary buckets are deleted. Rounded running compute totals USD 0.37 under the combined USD 5 ceiling, not an invoice. Preserve the H200 allocation canceled while queued and the first A100 worker's pre-proof executable-permission failure. Corrected local binary staging rechecks bytes before execution; the successful A100 native build is admitted by source-content/native-target and retained binary hashes.

The owner explicitly approved continuing without the inaccessible numeric credit check; balance remains null. Fresh-clone recovery rechecks the public manifest and reconstructs all nine C1 programs/witnesses. Local checks pass 2,613 tests with 50 skips; core packaging and strict docs gates pass. Exact-head hosted delivery is tracked on PR #135. The separate LLM repeat build timed out at 3,600 seconds; no fresh LLM GPU performance result is claimed. Stock classifier/core/verifier and 55-file LLM identities, models, thresholds and proof parameters remain unchanged. No hiding, hardware, independent qualification or useful-policy claim is added.

Public Circle reference collection — local acceptance and publication passed

The owner requested complete Circle headline numbers with actual proof/log and machine resources. Decision 0086 fixes baseline, scope and acceptance. The reference guide and numeric receipt distinguish L1 (H200 matched pair), C1 (fresh Mac classifier controller) and J1 (retained-input composition). They are not one measured fresh end-to-end GPU invocation.

Discussion #133 indexes all stage tables, machine/job details and source-pinned verification commands. The artifact release contains actual 23 LLM proof files plus 46 statements/claims, nine fresh classifier controls, three composed classifier proofs and 988 exact supporting files. All four assets were downloaded without authentication and matched their byte counts and SHA-256. The external manifest pin is recorded in the guide and Discussion.

The fresh C1 controller takes 302.96 s: MLP 1.95 s, CNN 172.94 s, Rust builds 24.82/9.57 s and proof campaign 93.50 s. Nine final STWO proving calls total 24.061 s; complete prover commands total 50.028 s; proof-only verifier processes total 0.807 s; actual proof files total 18,063,779 bytes. Peak sampled worker-tree RSS is 23.45 GB. These are one local observation, not timing distributions. The retained L1 CUDA session timer is 53.94 s, including interaction work; complete commands are 155.89 s and separate CPU grid verification 84.32 s. The raw 23-proof payload is 737,858,641 bytes. Reused/unrecorded stages remain explicit; SASH and classifier internal proving timers use different boundaries.

Fresh archive-based cryptographic verification takes 219.25 s including native compilation: all 23 LLM proofs twice, all 36 LLM mutations rejected, and all 12 classifier proofs checked in fresh proof-only processes with caller reconstruction. The first helper attempt passed LLM checks but incorrectly charged input archives to its work quota, failing before classifier checks. Failure is retained; generated Circle work now has its own 128-MiB budget. The 55-file LLM and Circle backend identities, models and thresholds are unchanged. No new paid job, hiding, hardware, independent qualification or policy-quality claim. Final repo/CI/Pages delivery is tracked on this change's PR.

Circle-first defaults — local acceptance passed

Circle STARK is the owner-selected default for make e2e, make e2e-composed, the web/TUI workbench, README, current guides and both published model cards. Decision 0085 records the scope; numeric acceptance pins exact sources, commands, outputs and failures.

Bare make e2e freshly trains/exports, generates and freshly verifies nine Circle proofs; 63 public-input mutations, 18 substitutions and three invalid rosters reject. It takes 320.68 s including local training/build checks. Default retained LLM composition takes 266.29 s including native verifier compilation: 23 SASH proofs, 22 links, two honest checks, 36 LLM mutations, three new Circle classifier proofs, three fresh outer-join verifications and 16 binding rejections. Verdicts [1,1,0] give integrity pass and policy release false. These are single Mac acceptance observations.

A browser-launched retained-model run makes nine new proofs with the same rejections in 97.94 s. Both interfaces select Circle first; 56 app tests pass, and actual desktop/mobile results were inspected. Core validation passes 2,588 tests with 50 optional/platform skips; strict docs and core distribution gates pass. Exact-head hosted/merge/Pages receipts are tracked with this change. No new cloud jobs or compute spend were needed. Both HF card-only updates leave weights and original release tags unchanged. All 55 LLM source pins and the Circle engine/compiler identity remain unchanged. No hiding, hardware, independent qualification or useful-policy claim is added.

Circle classifiers — verified implementation

Epic #112 has a complete experimental direct-STWO path for the unchanged rule, four-unit integer MLP and contextual CNN. It preserves complete 0–96-byte inputs, salted canonical SHA-256, strict integer decisions and caller-owned context/model/enrollment checks. The completed experiment exposed the nine-proof Circle control workflow; the later default change is tracked above. No core inference arithmetic, model, threshold or historical result is changed. Implementation and commands. The actual Make reproduction freshly builds both locked targets and passes all nine proofs and rejection checks; driver wall time is 97.81 s excluding compilation.

Three additional Circle proofs join the retained TinyLlama request. Fresh Mac verification checks those, 23 SASH proofs and 22 links, and rejects 16 changed bindings/rosters. Host composition takes 11.13 s including hashes and host work. Verdicts [1,1,0] give valid integrity with policy release false. No new LLM proving, recursion, hiding or hardware evidence is claimed. All 55 native LLM verifier source pins and their identity remain unchanged.

The acceptance record retains two nine-proof Mac runs, 63 direct public-input rejections, 18 request swaps and three roster rejections per run. Full-length conformance adds three proofs and 21 real constraint failures plus an omitted-edge program-pin failure. Six arithmetic fixtures cover both strict thresholds below/equal/above the boundary; local make check passes 2,580 tests with 51 optional skips. Final packaging, exact-head CI, merge and deployment receipts are tracked on PR #130.

The frozen MLP always allows in this 96-byte domain: its exact-rational interval score bound is [−5600,−2338] at threshold 0. The README and published card clarification disclose this quality limitation. Weights and original release tags are unchanged. Independent follow-up #131 remains outside this epic.

Terminal and web workbench — complete

make demo-tui and make demo-web launch shared local-first demo apps. One catalog, persistent controller and result adapter expose six existing workflows, observed stages, logs, local process-tree samples, HF jobs, retained artifacts and receipts. The guide includes actual rendered results; decision 0082 records the boundary.

The acceptance record pins the inputs, source and result hashes. The real TUI-launched classifier run verifies 9 proofs, rejects 18 substitutions and checks 405 arithmetic cases. Browser-launched native LLM verification checks 23 proofs and rejects 36 mutations. A fresh complete browser-launched TinyLlama run produces 23 proofs, checks 22 residual links, accepts two honest verifications and rejects all 14 mutations. All three HF jobs are completed, outputs retained and buckets deleted. The $0.31 running-compute estimate is not an invoice; the startup-inclusive admission ceiling was $2.50.

There are 52 passing app tests, including both interfaces, cancellation, request/artifact boundaries and recovery without duplicate work. Desktop/mobile and 120×42/80×24 terminal views were inspected. Core validation passes 2,398 tests with 123 optional skips. Distribution, app build and strict docs checks pass; exact-head hosted/deployment receipts are recorded on PR #129. All 55 native verifier source pins, its identity and the inference lock remain unchanged. No fresh GPU benchmark or composed campaign is claimed for this app work. Monitoring and completion-report checks add no backend qualification, privacy or hardware claim.

Public documentation site and badges

The README also reuses the homepage's proof-flow SVG below its introduction, linked to the detailed LLM statement. No duplicate asset is added.

The site is deployed. A copy-edit pass under the owner's no-ai-slop request replaces homepage slogans and repetitive learning-page framing with specific computations, commands and checks. Existing section anchors remain available. Proof relations, measurements, model files and historical evidence are unchanged; validation and delivery receipts follow on PR.

The owner requested a GitHub Pages documentation site and for-the-badge header badges, including both published Hugging Face models. The candidate site adds an editorial homepage, four educational chapters, a workflow chooser, searchable canonical guides, model cards and a research evidence map. Existing Markdown and numeric evidence stay canonical; staged link/anchor adaptation produces the website without publishing untracked notes or ignored runs.

Decision 0081 records the publication scope. The QA record pins source files and observed desktop/mobile, math, diagram, search, theme and proof-explorer checks; the final clipboard check was interrupted by a browser connection reset. Ten staging/link-boundary tests pass. Strict builds and local link/asset checks pass; exact final counts and hosted delivery are recorded on PR.

The docs toolchain is isolated in website/, with its own lock and dependency review. All 55 native verifier source pins, the inference lockfile and verifier identity remain unchanged. No training, proof benchmark, evaluator or numeric evidence is modified. Pages is enabled for the standard main-only Actions deployment with HTTPS; the repository remains public. Build and maintenance.

Published classifier models — complete

The owner subsequently requested publication of trained models. The exact seed-17 MLP and CNN are public on Hugging Face, under decision 0080, with model cards, original JSON checkpoints, unchanged quantizations, native exports, evaluation and hash manifests. The publication receipt records immutable Hub revisions and anonymous verification of all 41 released files. Full/resumed checkpoints and requantizations match; exported model bytes match the proved models; all 100 retained CNN test scores/decisions replay. Native examples run from anonymous downloads. This adds a hash-pinned inference command, not new training, held-out evaluation or proof qualification. The synthetic task and 96-byte demo scope remain explicit. TinyLlama is upstream, not a Hallmark-trained release. Local/distribution/hosted receipts follow on PR.

README benchmark and onboarding refresh

The README now leads with the recorded generated-kernel result: 53.94 s CUDA proving, 10.15× over same-host SIMD, and 155.89 s complete commands (4.36×). Timing boundaries, memory, proof size, fresh verification checks and raw evidence are adjacent to the headline. Quickstarts distinguish local classifiers, paid CPU inference proofs, paid GPU measurements and retained proof verification. The proof statement/diagram and source-ownership guides distinguish unchanged CPU Rust from generated/patched CUDA. This is documentation of the existing fixed-case evidence; runtime, protocol and numeric records remain unchanged. Local/distribution and delivery receipts are recorded in the ledger and PR.

Generated CUDA constraint kernels — complete fixed-case trial

The owner-requested generated-constraint pipeline passes its complete gate under decision 0079. The numeric evidence pins the measured source, admitted generated build, 77 kernel shapes across 16 families, all source/binary/input hashes and the retained failed attempt.

One matched H200 pair takes 547.75 s SIMD / 53.94 s CUDA summed proving (10.15×), and 679.49 / 155.89 s complete segment commands (4.36×). CUDA executes 1,710 device constraint evaluations, with zero CPU fallback, in 0.47 seconds total constraint evaluation. Both backends prove all 23 segments and pass full grid/token checks. All 69 files match each other and the historical CPU case. Fresh CPU verification accepts two honest checks and rejects all 14 mutations; Mac verification accepts two and rejects all 36. Witness/interaction generation remains on the CPU; hiding, hardware and backend qualification remain unclaimed.

The passing cached-build command takes 2,086.78 seconds. Its running-compute estimate is $1.76; the complete kernel campaign totals $3.37 including the build and telemetry-failure attempt, not invoices. All four jobs are terminal, outputs retained and buckets deleted. The earlier failed run's 23 valid individual proofs stay a failed pipeline record; the complete CSV-framing regression fix precedes the clean rerun. All 51 focused tests and 2,454 local tests pass (50 skipped). Distribution and hosted receipts are recorded on the delivery PR for issue #122. The historical fallback baseline below remains unchanged. One pair establishes no distribution.

GPU prover measurements — complete fixed-case trial

The GPU pipeline passes its complete research gate: same TinyLlama relation, reviewed witness archive, 23 SIMD and 23 CUDA segments on one H200, identical 69-file inventories and fresh CPU/Mac verification. Decisions 0077 and 0078 freeze the profile. The numeric evidence retains all four attempts, source/input/binary hashes, profiler counters and mutation results.

One CPU-then-CUDA pair takes 537.20/583.87 seconds summed proving and 669.25/687.34 seconds complete segment commands. CUDA is slower in this initial host-generation profile; its 1,710 CPU constraint fallback calls take 532.75 seconds. Peak host RSS is 104.69/11.78 GB, with CUDA using 67.08 GiB VRAM. Both grids pass; the separate CPU suite accepts two honest checks and rejects 14 mutations, and the native Mac suite accepts two and rejects 36. No hiding, hardware provenance or backend qualification is claimed. At this baseline, generated CUDA constraint kernels were an unimplemented optimization candidate; the later completed trial is recorded above.

The full retained-preparation benchmark takes 3,372.34 seconds. Running-compute estimates are $2.52 for the successful run and $5.37 across all attempts, not invoices. All six jobs are terminal, outputs retained and buckets deleted. Earlier source-freeze, fixture-enrollment and incomplete-head failures remain failed records. The measured snapshot passes 38 focused tests and 2,441 local tests. Four additional controller tests reject unsupported hosts or missing Rust before paid work. At that baseline's delivery, the local gate passes 2,445 tests (50 skipped), with 42 focused tests; publication/distribution gates are recorded on the PR. This is separate from the Circle STARK classifier implementation below; issue #120 tracks the GPU measurement delivery.

Circle classifier tracker and architecture guide

The owner requested clearer proof architecture documentation and a Circle STARK classifier implementation tracker. The architecture guide maps each Make target, the three flows, exact statements and SASH source/patch ownership. Epic #112 tracks five ordered tasks (#113–#117): freeze/replay, rule, MLP, CNN and benchmarks. At that tracker-creation baseline no Circle STARK classifier adapter existed; the implementation update is near the top of this file. This is new integrity research, not a remaining gate for the completed MVP or a substitute for original private/hardware v1 qualification. Runtime, model, evaluation and historical evidence remain unchanged; no proving/cloud run is needed for this documentation/tracker change.

The README now surfaces recorded headline results: model/response size, proving and verification times, serialized proof size and fresh workflow costs. These summarize existing evidence, with hardware and preprocessing caveats; they introduce no new benchmark or backend comparison.

Current MVP — complete

The owner explicitly selected completion of the LLM integrity MVP without hardware or hiding requirements on 2026-10-25. Decision 0076 records that scope; acceptance and limits map it to evidence. The completion audit rehashes all 207 runtime files at that recorded source, 55 native verifier files, 20 fresh-proving pipeline files, all eight model files and all 69 proof/statement/claim files. The complete short TinyLlama response, caller expectations, current signed composition and proof-only verification pass. No runtime or historical evidence is changed.

No further implementation gate is required for this selected research MVP. The original private/hardware v1 roadmap stays open as post-mvp work; its requirements have not been implemented or adopted by this completion decision. Operational custody/revocation, independent security/hiding qualification, general prompts/models and hardware provenance remain limitations/future work. The code retains qualified_backend: false, hiding_claim: false and hardware_claim: "not-accepted". The repository remains public and MIT licensed.

Supporting proof and workflow evidence

The owner's complete short pretrained LLM response now has a passing composed proof: 23 segments, 22 residual links and eight answer tokens through EOS. The model is pinned TinyLlama-1.1B-Chat-v1.0 with the disclosed y-attnout-w12 quantization. The proof explanation states the exact claim; raw evidence retains the failures.

The first non-delayed composition passed every individual proof but rejected the last link. Decision 0068 corrects the head extraction's family selection. Only that head is reproved; the 22 transformer proofs remain byte-identical. The original Rust binary is unchanged. Full verification takes 92.10 seconds on the smaller CPU worker.

A fresh proof-only worker passes: two honest verification processes and all 14 changed-claim rejections, with no model weights or witness mounted. The honest runs take 83.99 and 79.16 seconds. The caller checks pinned shape digests, exact tokens, complete coverage and the final-norm head. make e2e-llm now reproduces this result from fresh preparation, build and proofs. Its recorded run took 31 minutes 15 seconds, with a $0.31 compute bound rather than an invoice. All 69 proof/statement/claim files match the earlier accepted bundle byte for byte. Two fresh verifications took 80.11 and 78.45 seconds; all 14 mutations reject. All outputs are retained and the three temporary buckets are deleted. The first local patching failure remains recorded in decision 0069. make e2e still runs the classifier demo. The remaining operational gates are root provisioning and independent security/hiding qualification; hardware provenance is a separate extension, not part of this LLM demo.

An optional native Mac verifier now specializes the same proof relation to the frozen public case using compiled caller-enrolled shape roots. make verify-llm LLM_RUN=runs/llm-e2e-... rebuilds that verifier and checks retained proofs without weights, witnesses or paid jobs. The native experiment requires two honest checks, all 14 original mutations and two direct native unknown-shape/calibration rejections. Decision 0070 explains the preprocessing change and retained linker/sandbox failures. This variant does not rederive shape trees; operational root qualification remains separate. make e2e-llm retains the original Linux verifier's runtime derivation.

The native suite now includes 20 additional boundary mutations. The expanded probe found an output-token field alias accepted only when the caller was bypassed, plus four malformed shapes that crashed. The native verifier now checks vocabulary ranges, tree/column dimensions and redundant configuration; the complete suite requires two honest checks and 36 explicit rejections. Decision 0071 and numeric evidence preserve the baseline and candidate outcomes. Operational root custody and independent security review remain open.

The native verifier now supports detached Ed25519 authorization of the exact case and verifier source inventory, using a public key supplied separately by the caller. It checks authorization before build, before the proof worker and after verification, and reports whether it used signed enrollment or the checkout case. Decision 0072 defines the format; evidence retains wrong-key, tampering, expiry and changed-source rejections. The integration issuer is a disposable research key. This implements parameter authentication; operational key custody, revocation/rollback policy and independent security qualification remain separate obligations.

Earlier failures are documented in decisions 0065, 0066 and 0067. The delayed-value profile remains ineligible because its activation-scale binding is open; its constructor repair is not used in the complete-response result.

The owner-requested code cleanup consolidates four proof controllers onto one checked worker runner and replaces global sandbox overrides with per-call read paths. Validation covers retained failure costs, artifact tampering, sandbox isolation and real proof/verification execution. This maintenance gate changes no model, proof relation or qualification status.

The owner-requested usability gate now has a one-command entrypoint: make e2e. The fresh reproduction completed pinned preparation, both training/export runs, nine proofs, all 18 cross-request rejections and all 405 expected arithmetic outcomes. Both model identities and all three circuit identities match the earlier walkthrough; the executable pipeline stayed frozen. The proof explainer distinguishes in-circuit input hashing/arithmetic from caller-owned context/program/key binding, and a valid proof from a passing classifier verdict. Native verdict recomputation remains part of the caller helper. This usability gate does not close the qualification gates below.

Reader entry points: end-to-end demo, results and limits, documentation index. The walkthrough now has fresh reproduction evidence: both exports match, nine real proofs verify, all three decisions match, all 18 cross-request substitutions reject and all 405 arithmetic cases match their expected outcomes. The owner explicitly adopted MIT; historical experiment records remain byte-identical. This closes the documentation reproduction gate, not the qualification gates below.

The second synthetic screen and failed first screen retain their original evidence boundaries; the reused test pair is an engineering case.

For the broader v1 roadmap, the remaining qualification gates are explicit: authenticated compiler/program and key enrollment, setup custody, independent security/hiding review and an admitted B1 verifier adapter. B2 and the full hybrid path additionally need genuine hardware provenance and protected inference/output evidence. Issues #19/#29 remain open; research proofs do not satisfy those operational dependencies. Mac first; approved Modal/Vast/HF credits remain bounded fallbacks, and GCP is not a classifier-proof dependency. Historical entries below preserve prior results.

Implemented

Strict dataset/config/integer-model/run contracts; exact group/transcript split checks; canonical typed research hashes and salted byte-exact transcript commitment; optional synthetic-data logistic training/PTQ; full-coverage integer inference; separate calibration; metrics and uncertainty; fresh run bundles; caller-pinned replay; offline conformance/tamper tests; locked distribution and CI definitions. See ledger for actual commands and results.

W01 / issue #2 adds a separate bounded v1 primitive codec and 22 typed hash domains. The evidence records 108 codec tests, 137 total local tests and an isolated installed-wheel codec check. Legacy fixed hashes are unchanged. This validates value encoding only; artifact schemas and trust remain separate gates.

W30 / issue #31 adds typed errors and sanitized CLI failures, including argparse, filesystem/Unicode and child-process errors. Its evidence records the reproduced path leak, 52 privacy tests, 189 total local tests and an installed CLI error check. Historical run artifacts remain unchanged; new versioned events are described below.

W36 / issue #37 adds make check-train and a required-execution pytest guard. The evidence records 16 guard tests and fresh CPython 3.12/3.13/3.14 environments: each core gate passed 204 tests with one legitimate optional-training skip, each package gate passed, and each guarded train gate executed and passed its training test. CI now covers all three versions for both core/package and training, with no privileged integration job.

W03 / issue #4 adds frozen model/file/generation contracts, a canonical model identity, a JSON schema snapshot and explicit rooted file inspection. Its evidence records 128 manifest tests, 333 total local tests and an installed-wheel vector/file check. Filesystem inspection is bounded and rejects links, aliases, nonregular files and detected mismatches or concurrent changes; it does not load or attest to a model.

W04 / issue #5 adds frozen input/classifier/arithmetic contracts and typed operator parameter files. Its evidence records 133 classifier tests, all 128 model regressions, 466 total local tests and installed-wheel profile/parameter validation. Checks cover topology and, against exact hashed parameters, operation shapes, masks, scales and conservative integer range envelopes. These are static contracts, not an executed or qualified graph.

W05 / issue #6 adds frozen policy, request and caller-expectation contracts. Its evidence records 95 new tests, all 261 model/classifier regressions, 561 total local tests and installed-wheel validation. Checks resolve exactly rule/MLP/encoder in order, compare independently supplied expectations and enforce continuity shape and UTF-8 byte limits. They do not admit policy/data, authenticate enrollment or consume nonces.

W02 / issue #3 adds exact v1 transcript openings, caller-bound context and salted complete-input hashes. Its evidence records 69 new tests, 630 total tests, installed-wheel reconstruction and separate Python/ JavaScript agreement on seven transcript/context and 13 input vectors. The JavaScript checker rejects a late-token mutation. Bytes are reconstructed from the full prompt; encoder tokens still require an independently pinned local tokenizer. This is hash reconstruction, not a proof or independent verifier.

W06 / issue #7 adds strict receipt, epoch, public-input and proof-container records. Its evidence records 145 new tests, 775 total tests and installed-wheel declared-binding/packing checks. It checks ordered identities, exact profile count relations, caller expectations, intervals, opaque payload caps and the 100-element public-input layout. Synthetic signature, quote and proof bytes remain unverified; these contracts expose no acceptance result.

W08 / issue #9 adds external trust, real Ed25519 threshold verification, bounded complete successor history and an atomic SQLite checkpoint store. Its evidence records 88 new tests, 863 total local tests and installed-wheel enrollment/persistence. Rejections cover rollback, equivocation, skipped or incorrectly signed history, unknown/duplicate signers, freshness, retirement, stale writers and unsafe state files. Update/commit failures and process exit before commit preserve the prior checkpoint in the tested cases. The dependency review preceded installation. Fixture keys are public; operational root custody remains unenrolled.

W07 / issue #8 adds bounded offline bundle, private response, collection-state and reported-result contracts. Its evidence records 91 new tests, 954 total local tests, four schemas and 17 structural/report vectors. It checks exact embedded files, declared evidence links, three ordered proofs, caller limits and status/claim consistency. Openings remain separate, including rejection of canonical private records embedded as files. These are container/report checks; a decoded accepted flag is not verified evidence.

W09 / issue #10 adds caller-owned snapshot signing, immutable local history, irreversible retirement, dual-authorized key transitions and registry-check. Its evidence records 97 new tests, 1,051 total local tests, two schemas and an installed-wheel rotation/CLI fixture. The checkpoint now retains the external trust digest atomically, rejecting old policy reuse after rotation. Failure and process-exit tests preserve prior state and history. All keys are public fixtures; operational custody remains unapproved.

W33 / issue #34 adds a frozen corpus with 156 executable component cases and 32 unavailable genuine-evidence obligations. Its evidence records 46 harness tests, 1,097 total local tests, byte-identical corpus reproduction and installed-wheel checks. Software scope matches every executable outcome and catches a deliberately wrong expectation. Default full-evidence scope exits 3; neither hardware/proofs nor independent verification passed. All original PRD cheats remain in the required map.

W11 / issue #12 adds bound byte-count affine/ReLU/affine execution, with exact ties-to-even rescaling, int32 partial-sum checks and full prompt coverage. Its evidence records 75 native-runtime tests, nine fixed vectors, 60 seeded intermediate-layer comparisons, 1,172 total local tests and installed-wheel validation. Weight/profile identities and file paths are checked before execution; resource limits reject rather than truncate. The exported weights are synthetic markers. Private traces are not proof evidence, and no policy training, quality evaluation or backend qualification occurred.

W31 / issue #32 adds bounded private POSIX event storage, allowlisted schemas, random run IDs, exclusive writers and immutable attempt/failure records. Its evidence records 101 event tests, 156 focused regressions, 1,273 total tests and installed-wheel persistence. Nine sync-failure points and four process-exit cases never produced an accepted complete attempt. Failed attempts survive an explicit permitted retry; unknown metadata, duplicate ordinals and unsafe files reject. Metadata remains private; this does not authenticate execution, qualify physical power-loss behavior or implement W32 full-path cost measurement.

W10 / issue #11 adds a bounded literal byte-rule program and complete-input native execution with suffix state across chunk boundaries. The evidence records 97 new tests, 1,000 seeded oracle comparisons, 11 reproduced fixed vectors and 1,573 total local tests, with both installed-package gates passing. Invalid bindings, parameter/provenance changes, padding-dependent matches, resource excess and inconsistent results reject. The scalar arithmetic DAG projects the Boolean match feature; it does not encode the matcher. The full extractor/state must be constrained by a future W19 proof. Patterns are authored harmless markers, not a measured injection detector. PR #70 merged as fa310ea2a4d1d484f1ef325eb7bfbf5f86a280ed; all 12 main jobs passed. Issue #11 is closed.

W14 / issue #15 adds externally pinned source import, exact raw-row lineage, complete parent/history/prompt grouping and the frozen study split recipe. The evidence records 59 new tests, 25 independent graph-oracle cases, seven reproduced schema snapshots and 1,632 passing local tests, with both installed-package gates passing. The 26-row synthetic fixture retains all contrastive examples in 21 groups. Changed source/review bytes, missing provenance, cycles, policy mismatches, rehashed cross-split descendants and resource excess reject. Four injected sync failures preserve partial directories without publishing the final manifest. Automatic near-candidate discovery covers ASCII case/whitespace equivalence only; broader semantic lineage, actual rights and human label review remain external obligations. See decision 0021. No real source was downloaded/admitted or task quality measured. PR #71 merged as c4586a7107c59fb7ba4665247e73ff7e0146cf7d; all 12 main jobs passed. Issue #15 is closed.

W15 / issue #16 now has the record-level MLP component: train-only complete-input max-logit fitting, pinned same-environment optimizer resume and explicit native integer export. The evidence records 33 new tests, 1,665 total tests, 17 required train tests and both installed package gates. Two fresh three-seed fixture runs produced 27 identical artifacts each, including checkpoint/history/export results. All seeds hit the 80-epoch limit without convergence; maximum observed score error was 0.016157 against the fixed 0.1 engineering bound. These are authored markers, not policy-quality data. The historical logistic baseline and native fixture identities remain unchanged. See decision 0027. The follow-up comparison evidence adds exact-text character n-grams, fixed controls, five grouped train-internal rotations and separately pinned native calibration/evaluation. It records 44 more tests, 1,709 total tests, 35 required train tests and both installed-package gates. Two new runs reproduce all 31 artifacts: 75 rotations, 15 deliberate nonconvergence cases, and nine final evaluations each. The toy MLP detects none of the five positive representatives; its false-positive counts are 0/1/2 across the seeds. These failures remain visible; the pipeline pass is not a task-quality pass. Decision 0028 fixes the selection rule, source bounds and remaining admission obligations.

Merged PR #73 contains this work. The guard source review and decision 0029 select the pinned published guard for bounded asset/runtime qualification under the reviewed model grants. The CPU evidence now verifies all five acquired asset digests and executes the real ONNX graph locally. Two fresh processes reproduce all 11 output artifacts: seven authored probes, 50 chunks per run, including 32,768-byte ASCII and Unicode prompts. Elapsed times are 7.253/7.008 seconds, with 4.372/4.371 GB peak RSS. The optional adapter joins the pinned calibration/evaluation API; its tiny installed-runtime fixture needs no research assets. The tokenizer remains lossy, and both maximum-length harmless repetitive probes have positive uncalibrated margins. These are software results, not measured task quality. Contamination/label review and admitted real-data commands remain required for confirmatory task-quality claims; decision 0033 permits completion of the synthetic software stage. The compact encoder belongs to W16. See decision 0030.

The comparison command evidence adds one caller-enrolled execution path from reproduced source import through train-only selection, final fitting and calibration to test scoring. Every fit and threshold is retained before any test-scoring call. Two fresh authored-fixture processes reproduce all 38 output artifacts, each in about 2.9 seconds and below 273 MB peak RSS. All 75 rotations and 15 nonconverged attempts remain visible. The toy MLP misses all 20 positive test representatives in every seed; the character baseline has 2/20 benign false positives. This is plumbing evidence, not a detector-quality result. Validation passes 31 new tests, 1,772 total tests, 35 required training tests, 32 guard tests and the core installed-package gate. The larger local combined-package check is deferred under the nearly exhausted 20 GiB storage limit; Linux/macOS CI requires that installed comparison fixture. The command template is implemented, but no genuine reviewed real-data dossier has been enrolled. PR #73 remained draft at that point. See decision 0031.

The source review now retains 1,000 Gandalf viewer rows at an exact reported revision and Dolly's SHA-256-verified 15,011-row JSONL. The fixed empty-context filter leaves 10,544 Dolly instruction candidates and excludes 4,467 rows. Two fresh builds reproduce a 64-item blank human label packet. These are review candidates, not admitted labels or data. The 114 exact-duplicate groups and a separate 563-pair lexical screen are retained; one candidate component crosses Gandalf's original splits. Missing historical parents and semantic lineage remain unresolved. A separately acquired 5,441-row public synthetic source cannot replace natural traffic or the study's power requirements. No detector was fitted/scored or HF token used. See the human-review handoff and decision 0032.

The authorized synthetic pipeline evidence now joins the complete comparison command to the real pinned guard weights. Two fresh Mac processes reproduce all 38 artifacts in 4.674/4.665 seconds, with 4.374 GB peak RSS. The agent checked all 200 authored control labels; no human review occurred. The guard produces 3/20 negative-control false positives, the character model 2/20, and each MLP misses all 20 positive controls. These are software-control outcomes, not injection-policy quality estimates. W15's execution criterion passes; full-study/quality/independence flags remain false and the compact encoder stays explicitly unavailable. The owner's decision supersedes the earlier human-review/storage merge blockers for this pipeline stage. All 1,772 local tests, 35 required training tests, core/combined package checks and the live specification audit pass.

W37 / issue #38 adds a fully pinned build environment, reviewed source snapshots, byte-identical repeated distributions and exact archive/dependency inventories. Its initial evidence and access-time correction record 66 new tests, 82 focused tests and 1,339 total local tests. The correction permits access-time updates while retaining identity/modification-change rejection. Both package gates build twice, rebuild the same wheel from the sdist and exercise fresh installed core contracts; the train gate additionally executes synthetic fit/replay and tamper rejection. The dependency review matches 518 locked artifact hashes for 32 packages and records actual licence metadata/notices. Linux/macOS CI now covers core/package and required-train/package on CPython 3.12–3.14. This is engineering reproducibility within a recorded environment, not package publication, an open-source licence, scientific validity or a release packet.

W38 / issue #39 adds the offline make check-spec gate and an explicit read-only live issue audit. Its initial evidence and final validation record 54 new tests, 55 focused tests and 1,393 total local tests; both package gates also execute the spec check from the extracted sdist. The live observation matches 40 work issues, 10 trackers, 106 dependency edges, labels, milestones, scope and acceptance. Local coverage names 51 sections and 18 invariants. Missing requirements, anchors, orphan mappings, cycles, unowned questions and metadata drift reject. Historical source/corpus/legacy results are unchanged. Structural agreement and issue state are not semantic completeness, approval or scientific/security evidence.

W13 / issue #14 now has a concrete admission proposal and strict study/protocol/review records with caller-pinned document checks. The evidence records 81 new tests, 1,474 total local tests, byte-identical marker/schema reproduction and both installed-package gates. The first package attempt rejected unstaged files missing from its tracked-only snapshot; the complete staged candidate passed. At that stage, source/model/host holds and unapproved quality/power/cost terms remained explicit. PR #68 merged as 346f9bcc1940985c90053a719eb2ed4fea4671b6; all 12 main jobs passed.

The owner subsequently approved the prototype study, targets, delegated evaluation design/public source selection, bounded local budget and conditional HF Jobs. Decision 0019 and AUTHORITY record that authority. The frozen design fills quality/power/cost and retention terms, including an exact conditional power plan. The approval-update evidence records 1,476 passing local tests. Existing source holds remain; actual file/split, candidate/evaluator and independently enrolled final review identities are still missing, so the proposal correctly rejects as a measurement admission. The source-selection and engineering work is authorized now. No dataset/model asset, paid job, private upload, hardware evidence or independent review occurred. PR #69 delivered this update as 7f547356a7f91e8d331a6963d0cf40603dd49b44; all 12 main jobs passed. Issue #14 is closed as admission preparation, not completed measurement.

W17 / issue #18 now adds caller-pinned versioned slice and paired group reports. The evidence records 40 new tests, 1,812 total tests and both installed-package gates. The authored correlated fixture keeps 36 records in 12 resampling groups, all three seeds and deliberate float/integer regressions. The report never fits, scores or calibrates a model. Two fresh report commands over retained W15 controls reproduce identical bytes in 0.292/0.283 seconds, with less than 42 MB peak RSS. They retain ten available models, nine 10,000-resample comparisons, twelve unavailable members and nine unavailable quantization pairs. Original fixed controls remain in the pinned provenance. Shared marker templates are not independent policy samples; this is software evidence. Synthetic/exploratory reports do not evaluate a confirmatory decision, and confirmatory requests remain inconclusive. The initial package check rejected global uv 0.13.4; an isolated required uv 0.9.18 made both gates pass without changing the global tool. See decision 0034.

Broader v1 work not implemented or established

TEE evidence, operational registry enrollment/authoring/rotation, proofs for enrolled operational Hallmark classifiers, full bundle assembly and combined verification, separate verifier for complete hybrid bundles, real safety-policy dataset/evaluation, encoder/teacher/student training, full proof/compiler qualification, production service, deployment, partner approval or any safety guarantee. CI definitions alone are not hosted green checks. CODEOWNERS and agent rules are not enforced sandboxing or branch protection.

Implementation history

The proposed v1 corpus defines the full hybrid POC, with 10 draft RFCs and explicit owner-review gates. Merged PR #1 contains the corpus and implementation tracker: 40 work issues, 10 RFC trackers and 102 implementation dependency edges. The live audit checked all issue bodies, labels, milestones, comments and child checklists. This is specification/backlog delivery, not implementation or new scientific evidence; see decision 0002. The owner explicitly authorized public repository visibility on 2026-10-20; the standing visibility instruction is in AGENTS.md.

Local Python 3.12 train-enabled and Python 3.13 core-only baseline gates pass, including isolated wheel checks. Earlier private hosted runs could not start because of an account billing/spending-limit block. After the owner made the repository public, hosted run 39398412605 passed on 48f2d7066b62a42f971cd10dbbc2cdbae720b82e. Each implementation PR still needs checks on its own commit. Billing settings were not changed.

The owner requested autonomous implementation through .agents/harness/goal.md, sequentially with one issue per PR. #2 is merged in PR #52 and #31 in PR #53.

37 is merged in PR #54, with

merge-commit CI passing on 6b2a6acda5813ddcdf9d14554a99ada8ad2e6245. All three merge-commit CI runs passed. W03 / #4 is merged in PR #55; merge-commit CI passed on f844342c18af85d8cb0b1160521c58938f9d87ea. W04 / #5 is merged in PR #56, with all six merge-commit jobs passing on f456b4a749d795f2f22f19e5d2ae5fd90b099539. W05 / #6 is merged in PR #57, with all six merge-commit jobs passing on 1d344d55b44efae9b0232c46d4553dfe59f49288. W02 / #3 is merged in PR #58, with all six merge-commit jobs passing on fbd6c90f65e5e535f8dfb839012f1bd3bd544097. W06 / #7 is merged in PR #59, with all six merge-commit jobs passing on 774887d0e40b82459d4f7ca8657400d9905577ad. W08 / #9 is merged in PR #60, with all six merge-commit jobs passing on 866d36a69762104847f1ded8253829a4e11dd786. W07 / #8 is merged in PR #61, with all six merge-commit jobs passing on 302a861043636094295c4dc06832e4a8c06afe87. W09 / #10 is merged in PR #62, with all six merge-commit jobs passing on 9357288387bfbcdf0ad28ae34f54b37fe70a1333. W33 / #34 is merged in PR #63, with all six merge-commit jobs passing on 89de51b49cfbf10040384967b289a32b9709cb02. W11 / #12 is merged in PR #64, with all six merge-commit jobs passing on e8c9d0be603d477e3861b6843538477974b97354. W31 / #32 is merged in PR #65, with all six merge-commit jobs passing on 5e23f44ca51a58760ff1c139c114b7db52c5c1c5. W37 / #38 is merged in PR #66, with all 12 main jobs passing on 56b53f42cdc25f7690b6e8d1ed95a4e87b1e0bfb. W38 / #39 is merged in PR #67, with all 12 main jobs passing on 1da52da9c5f45b481d031ab3dd1ec5b4601c7a6d. The current work map has 106 dependency edges, including W04's reuse of W03 file references and W07's opening/signed-registry dependencies on W02/W08; the historical 102-edge publication audit is unchanged. Re-evaluate dependencies before #39 spec traceability and the remaining admission/integration work. Research, real proofs, genuine hardware and independent verification remain required for the full MVP.

W14 / #15 is merged in PR #71. W15 / #16 is merged in PR #73 as de29271d21196202fc951ebe5f5fc3348fbf9a63; all 28 candidate jobs and all 14 merge-commit jobs passed. The owner's synthetic-stage approval permits agent review without human labels for pipeline development, 40 GiB retained storage, Mac-first execution and HF T4/L4 fallback. The earlier label packet and all failures remain retained. Confirmatory task quality is separate.

W17 / #18 is merged in PR #74 as 5312460d8a48a4771d502592a05fb3ba362bdc03; all 28 candidate jobs and all 14 main jobs passed.

Under AUTHORITY, review public source rights and complete exact inventories; exclude ambiguous/gated assets. Freeze actual data, candidate and evaluator identities before measurement. Use PROGRAM, ML-PLAN and EVALUATION. The hardware/reviewer options are prepared; procurement and contact remain owner choices when those gates are ready. Neither blocks ordinary local implementation.

Keep further proof work sequential within each recorded bounded campaign. The tiny graphs already have real proof evidence; they do not qualify a larger encoder. Preserve the hybrid hardware/ZK split and the B1/B2 trust distinction in SPEC.

Resume

Inspect git state and actual installed environment, read the latest ledger entry, then take one falsifiable gate. Run the smallest relevant negative test. Update this state and the ledger with evidence, limits and a concrete next step. Use the current approved research authority; repository visibility follows AGENTS.md.