0091: Audit fixes for paid runs, verifier boundaries and lost tests¶
Date: 2026-10-29. Status: owner-requested; implemented and validated on main.
The owner requested fixes for every finding of the 2026-10-29 whole-repository audit (tracking issue #159). The audit found no wrong acceptance in the Circle classifier path, the wire decoder or the signature checks. The fixes, merged as PRs #160–#165 and this record's PR, change three recorded contracts.
Paid LLM ceilings. HF bills a job's Starting window, and the controller
cancels a job only at its timeout plus 600 s. planned_cost now counts that
window once for every caller; the duplicate wrappers are gone. At the recorded
prices the startup-inclusive worst case of make e2e-llm is $2.27, above the
earlier $2.10 CLI ceiling and inside the $2.50 the workbench already used for the
same workflow. The CLI and composed paths now use that $2.50 ceiling. The GPU
benchmark bound ($4.66) is unchanged. DELETED is a terminal job stage; paid
GPU runs check local verifier prerequisites before creating any job; the one-off
credit-check exception flag from 0088/0089 is removed.
Native LLM mutation suite. Changed prompt and changed model root were rejected only by the Python pre-check. Two native-only variants now require the compiled verifier itself to reject them while all 23 proofs still verify. The native suite grows from 38 to 40 cases (38 rejections); the Linux suite is unchanged. Evidence records the first run on the published proofs.
Verifier-side boundaries. Registry enrollment now enforces successor rules
between consecutive snapshots, so a retired entry cannot be revived or dropped.
Proof workers may signal only processes in their own sandbox. The verifier time
budget applies to circle-verify again, and a directory in a work directory no
longer loses the cost record. CNN admission rejects any export whose per-site
integer bound could leave int32; the frozen CNN passes. The worker runner tests
deleted with the pairing backend are restored, and the CNN circuit and Rust
verifier rejection paths now have tests in CI.
Identities. The Circle backend identity c035913b…1ded7d, the caller
enrollment pin, models, thresholds, proof relations and the LLM case are
unchanged. The native LLM verifier identity becomes 39d2ee50…5879ef, because
_worker.py, run_llm_e2e.py, classifier_manifest.py and uv.lock are in its
inventory. Signed LLM enrollments need a new owner signature, as in 0090. The
unused pytest-cov dependency is removed under a new
dependency review. The design-power plan
reproduces unchanged under the new lock
(lock-only refresh).
Acceptance¶
Numeric record,
local Mac, main 1bc6b72 with all six fix PRs merged. Bare make e2e passes in
253.87 s: nine fresh Circle proofs, 63 public-input mutations, 18 substitutions
and three rosters rejected, enrollment pin unchanged. The documented conformance
command passes in 187.60 s: ten boundaries, 22 invalid witnesses rejected by
constraint checks and three maximum-input proofs freshly verified. Native LLM
verification of the published proofs passes in 203.59 s with identity
39d2ee50…: 23 proofs, 22 links, two honest checks and all 38 mutations
rejected, including the two statement mutations. make check (2,251 tests, 45
optional skips), check-train (63), check-demo (60), docs (147 pages) and
package-check pass. One false mypy error came from a stale local cache and
passes with a clean cache. These are single local observations, not timing
distributions. No cloud job, spend, hardware, hiding or qualification claim.