0083: Exact classifier relations and a matched Circle STARK comparison¶
Operational default superseded by 0085: Circle is now the owner-selected default. This document retains its original protocol and results. Superseded by 0090: the pairing backend was removed.
Date: 2026-10-27. Status: frozen protocol executed. Implementation/evidence for epic #112 and #113–#117 are recorded below.
Gate and frozen baseline¶
Hypothesis: a fixed classifier circuit over a pinned Circle STARK engine can prove
the exact existing rule, integer MLP and contextual CNN relations, including their
canonical salted SHA-256 input commitments, and support a complete matched cost
comparison. No speed advantage is assumed. Baseline is main
0b7fcf256e2b1da7f89578014bcc86e6737fc3f5.
The first acceptance gate is #113: executable frozen model/caller/corpus admission, native replay and the complete benchmark protocol. The full epic remains open until all three real STARK paths, nine-proof workflows, common-request LLM composition and repeated measurements pass. An infeasible candidate is a preserved result, not completion.
Mutable paths are new comparison scripts/tests/configuration, the new
research/probes/circle/ backend, its adapter and workflow integration, and the
associated guides/STATE/LEDGER/evidence. Existing proof implementations, native
arithmetic, model exports, datasets, thresholds and historical numeric records
remain unchanged. Backend-specific enrollment does not change model semantics.
Relation and caller boundary¶
The existing classifier statement is authoritative: complete
0–96-byte input, no truncation, exact padding/masks, literal BAD across chunks,
integer rounding and threshold equality blocking. The SHA-256 preimage is
u32be(len(tag)) || tag || CBOR([salt32, profile32, count, byte_values]), with
tag = hallmark/v1/input. It is distinct from the transcript-opening commitment.
The input digest must be constrained inside each proof.
The 100 logical public fields remain context digest, classifier identity, input digest, unit count, chunk count, biased signed threshold and verdict. The caller constructs context/classifier identities and salt derivation, admits exact models and inputs, and checks complete unique classifier coverage. Those host semantics must not be presented as in-circuit context reconstruction.
Classifier manifests include backend/key/setup identities. Their full IDs, policy/context hashes and derived salts can therefore differ across backends. Match the unchanged model/program/profile/arithmetic/threshold and caller opening contents; record each complete backend-specific identity chain. Do not claim byte-identical public statements across distinct backend enrollments. The new proof envelope enrolls the backend program/preprocessing and proof parameters before proving. It must identify the exact compiled relation selected before proving. Public 32-bit values need bounded limbs over M31; reduction modulo M31 is not a transport decoder. SHA words, integer signs, carries, ranges and rounding branches require constrained, unique interpretations. A field equality alone does not establish an integer equality.
Backend and security¶
The engine review selects pinned direct STWO 2.2.0, separately from SASH's Llama application. The experimental candidate passes complete rule/MLP/CNN acceptance; it is not an adopted backend and is not production-qualified. Reuse the Circle FFT/FRI/Merkle/constraint framework, with specialized classifier and exact SHA-256 constraints. There is no general-purpose VM.
Preprocessing wiring, constants, copy multiplicities, model/program identities, column shapes and PCS parameters are caller-enrolled. A fresh verifier accepts proof plus caller-owned expectations/enrollment, without the witness/checkpoint, prover invocation, native-inference fallback or trusted success flag. SHA-256 input binding is separate from the engine's Blake2s transcript/Merkle hashes.
Concrete FRI/lookup parameters and their soundness assumptions must accompany the results. M31/QM31 statistical bounds and implementation assumptions do not establish a concrete security level. The Circle path gains no witness hiding, hardware provenance, an audited setup ceremony or production qualification here.
Measurements and resource bounds¶
The versioned plan freezes the three
cases, host, threads, storage floor and failure handling before candidate
measurement. Preserve failed samples. The matched comparison campaign
(circle-classifiers-matched-v1), its sample schedule, timing overlay, scripts
and report were removed with the pairing backend under
0090; they remain in git history at
b7b8376b584836f29066708b663006c2cae7328e.
The earlier acceptance/conformance trials had protocol files written before execution, but precede publication of this decision commit. They are exploratory, retained under their original identities, and excluded from the formal matched sample set.
Report compilation, setup/preprocessing, model commitments, witness generation, proving, serialization, proof-only verification, caller binding and total wall separately. Measure native-only and native plus binding; include the current caller's native recomputation. Measure proof/key/parameter/disk sizes, encoding, compression, sampled process-tree RSS and exact source/environment/input hashes. Account for reusable cost at a declared amortization denominator. Match host and threads or separate the results; LLM measurements are a different workload.
The initial build/API fixture is capped at 600 seconds, four threads, 16 GiB RSS and 4 GiB new storage. Classifier experiments use the frozen plan's bounds, one foreground command at a time. Review licences/dependencies before execution. Prefer the Mac. A cloud fallback needs a separate frozen plan with live balance, prices, source and public inputs checked under standing authority. No automatic paid retry, top-up, private upload, outreach or scheduler is introduced.
Review and historical evidence¶
This is a new experiment over unchanged semantics. Historical benchmarks are preserved and are not relabelled as matched measurements. If semantics, arithmetic, commitments, datasets or evaluation change, start a new identity and obtain owner review before reusing previous results under that changed protocol. The owner's request authorizes implementation of the full epic; it does not make an untested candidate pass, silently adopt a backend or waive any acceptance criterion.
Measured outcome¶
The matched campaign passed all 144 measured proofs, a fresh Mac postflight and the retained-LLM join, and initially retained Circle experimentally behind the earlier default. In that stateless harness Circle had lower complete-workflow time, with higher memory and larger proof files than the baseline. Concrete security profiles remain unqualified. 0085 made Circle the default and 0090 removed the baseline. No original classifier arithmetic, weights, thresholds or historical measurements are changed.