0086: Publish verifiable Circle reference artifacts and complete stage numbers¶
Date: 2026-10-28. Status: owner requested; local crypto acceptance and artifact publication passed.
The README must distinguish proof-kernel time, complete proving commands,
verification, preparation, checks and total workflow time. Every number needs a
named run, machine, timing boundary, exact artifact bytes and a retrievable source.
Baseline: 9bea024dc7d4604f17d53c915b64bc4a23c408b5.
Hypothesis: the retained GPU, classifier and composed runs can support a public, hash-pinned reference collection with actual proofs, caller inputs, environment records and selected raw logs. The collection must identify separate component runs; no uninterrupted fresh GPU-plus-classifier workflow is inferred. Missing measurements remain explicit. The previous numerical records remain unchanged.
Mutable paths: a transparent local timing recorder, reference packaging/audit and proof-only verification helpers, focused tests, a new numeric record, README, reference/benchmark guides and site, STATE/LEDGER. Preserve the 55-file LLM verifier identity, Circle engine/compiler identity, models, thresholds and all historical experiments. No new paid job, data/model selection or arithmetic change.
Acceptance: one fresh local default classifier run with observed stage walls; all published proof bytes hash-match accepted originals; safe explicit publication inventory excludes credentials, private data, witnesses, checkpoints, caches and executables; fresh proof-only verification of the public LLM/classifier bundle; public GitHub artifact downloads with SHA-256 manifests, Discussion index and machine/job/parameter/source metadata; tables reproduce raw numbers; docs and package/tests plus hosted publication checks pass. A hash check alone is never reported as proof verification.
Local classifier limits remain four prover threads, 30 GiB sampled worker RSS, 4 GiB output, 1,800-second proof campaign and two-hour controller, with 8 GiB free storage during proofs. Native LLM verification keeps its existing bounds. Prefer retained proofs and caches. Use new directories; preserve failures. Publication is explicitly authorized by the owner's request for a public reference Discussion and all necessary proof/log artifacts. This is a research artifact release, not security qualification or policy adoption.
Numeric acceptance, reference guide and Discussion #133 retain the fresh C1 observation, actual proof assets and archive-based verification. All four assets pass anonymous download/hash checks. The first quota-scoping helper failure is preserved. Source/math/model identities are unchanged; final CI/Pages delivery is tracked on the associated PR.