feat(evals): eval architecture + probe wave 1 — first BEHAVIORAL skill verdicts - #1033
Merged
Conversation
…l verdicts
Research -> architecture -> harness -> measurement, one arc:
- docs/research/: SOTA skill-eval standards (SkillsBench/AFTER/statistics,
adversarially verified) + formal-verification-for-agent-code (oracle-scarcity
core: assertion density before simulators; Antithesis/TigerBeetle primary-
source verified).
- docs/architecture/eval-architecture.md: 12 decisions — false-PASS rate as
the trust metric, frozen weak-producer reference config, 3-stage adaptive
arms, sequential stopping, digital-twin/metamorphic substrate, weekly-wave
budget policy. PROPOSED pending ratify.
- Harness: MEASURED ledger moved to evals/skill-probes/LEDGER.md (own file —
root-cause fix for the SKILL-TIERS regen wipe; gate/README/seed hint
repointed); _stats vendored to evals/_stats (42 tests); probe-skill.sh
gains --effort ratchet + producer provenance; skill-telemetry hook script
(wiring per-machine via gitignored settings).
- Wave 1: 5 new probes + crank re-run x {xhigh, low} on gpt-5.6-luna, 48 runs,
selftested deterministic discriminators, dual-config fixtures. FIRST
BEHAVIORAL verdicts: premortem + standards (xhigh 0.5->1.0, low 0.0->1.0 —
effect grows as producer weakens). validate/security/reality-check ceiling-
saturated both efforts (scenario hardening queued, skills untouched). crank
INERT across 3 configs (cull/reshape candidate).
- Measured cores hoisted into skills/premortem + skills/standards SKILL.md
(MEASURED blocks with probe citations); codex twins regenerated.
- Coverage gate: 4/11 product/judgment skills measured (was 0/11).
Evidence: docs/evals/2026-08-04-probe-wave-1.md + scorecards/ + committed
fixtures (replay: bash scripts/probe-skill.sh --probe <id> --replay).
boshu2
enabled auto-merge (squash)
August 5, 2026 00:48
Three v2 probes bury the planted flaw in production-shaped noise (euphemized not_checked under release pressure; scanner failure as soft mid-log WARN; gap derivable only by cross-output arithmetic). Result: ALL still ceiling at xhigh AND low on gpt-5.6-luna. reality-check-v2's apparent n=2 headroom (0.5/0.5) resolved as sampling noise by sequential extension to n=6 (6/6 both arms) — the D4 stopping rule catching a wrong conclusion in the wild. Recorded conclusion: these doctrines are quiz-robust on frontier models; no more quiz-hardening rounds — measurement moves to task-embedded Tier-2 fixtures where the doctrine must fire unprompted mid-work. Ledger +3 rows, report gains Wave-2a section, scorecards under docs/evals/scorecards/2026-08-05/.
- images/gemini/skills/{premortem,standards}/SKILL.md: zero-conversion resync
after the MEASURED-block edits (third runtime projection; codex twins were
regenerated, gemini copies were missed — verify.sh drift gate caught it)
- formal-verification doc: 'validate/verdictcheck' phrasing parsed as a dead
/verdictcheck slash-skill ref by check-doc-skill-refs --all-docs; reworded
boshu2
added a commit
that referenced
this pull request
Aug 5, 2026
…e-side) (#1037) Transcript-verified (explicit 'I'm using agentops:<skill>' + skill-body read; self-reports alone proved unreliable twice). Flagship result: the premortem loop closed end-to-end — wave-1 measured doctrine, hoisted MEASURED block (PR #1033), codex worker routed unprompted, loaded the block, named self-certification first citing the skill's key check. Runtime/model confound labeled; product implication recorded: skills are THE delivery channel on codex, one-of-several on Claude-in-repo — report per-runtime delivery, not one routing number.
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.
What this is
The eval program for the skill harness, end to end: SOTA research (two adversarially-verified deep-research passes) → a 12-decision eval architecture → harness completion → the first measured probe wave → skill improvements driven by the measurements.
Wave-1 results (gpt-5.6-luna, xhigh + low effort, N=2/arm/config, directional)
The effect grows as the producer weakens — the SkillsBench gradient reproduced locally. Full report:
docs/evals/2026-08-04-probe-wave-1.md.Verification (all in this branch's tree)
_statsvendored: 42 pytest greenbash scripts/probe-skill.sh --probe <id> --replayNotes for review
evals/skill-probes/LEDGER.mdis the new hand-maintained MEASURED ledger — moved out of generated SKILL-TIERS.md because a regen wiped it there (measured results can't live in generated files).