test(bakeoff): resync protector fixture CVSS + add flip-rate mode#249
Merged
Conversation
Two fidelity fixes prompted by a live false-`exploitable` on protector's own pod (the judge hallucinated [reachability: loaded-at-runtime] on CVEs the evidence tags [reachability: not-observed]): - The `protector_notobserved_cves_broad_rbac` fixture had drifted from live: CVE-2023-45853 was [cvss: 5.3], the live feed now shows [cvss: 9.8]. Resynced so the bench tests what the model actually sees (CVSS shouldn't move a not-observed verdict, but for a small model "critical + 9.8" pulls harder than "5.3", so the fixture must be faithful). - `--flip` mode: run ONE case N times and tally the verdicts. The bench runs each case once, which hides a rare temp-0 tail-flip — a model that refutes 19/20 and flips once reads as a clean pass. HTTP-only (no ollama subprocess management) so it runs against a REMOTE ollama over a port-forward, measuring the DEPLOYED judge on the DEPLOYED prompt. Measured on the deployed qwen3:1.7b against the faithful full-scale prompt (num_ctx=16384): 20/20 refuted (0% flip). The live exploitable is a rare temp-0 tail event, NOT a regression — consistent with ADR-0029 (accepted in shadow; JEF-445 stops it freezing in the verdict cache). No model/prompt change warranted. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VtjoJttCvBY4dzCoE4f9vP
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
A live false-
exploitablelanded on protector's own pod: the judge reported[reachability: loaded-at-runtime]on CVEs the evidence actually tags[reachability: not-observed]— a fabricated tag (its own reason gave it away: "despite not being observed running"). Correct verdict isrefuted. The bakeoff already has this exact case (protector_notobserved_cves_broad_rbac) andqwen3:1.7b"passed" it 12/12 — but two things made that misleading.Changes
CVE-2023-45853was[cvss: 5.3]; the live feed now shows[cvss: 9.8]. CVSS shouldn't move a not-observed verdict, but for a small model "critical + 9.8" pulls harder than "5.3" — so the fixture must match what the model actually sees.--flipmode. Runs one case N times and tallies verdicts. The bench runs each case once, which hides a rare temp-0 tail-flip (refute 19/20, flip once → reads as a clean pass). HTTP-only (noollamasubprocess management) so it runs against a remote ollama over a port-forward — measuring the deployed judge on the deployed prompt.Result
Measured on the deployed
qwen3:1.7b, faithful full-scale prompt,num_ctx=16384:So the live exploitable is a rare temp-0 tail event, not a regression. No model swap, no prompt change — that would be over-fitting a sub-1-in-22 tail the model already handles. Consistent with ADR-0029 (accepted in shadow; JEF-445 stops it freezing in the verdict cache, so it self-heals on the next pass).
🤖 Generated with Claude Code