Skip to content

feat(safety): add module-first harness safety evaluation - #2

Merged
JiaenLiu merged 115 commits into
mainfrom
codex/safety-measurement-evaluator
Aug 27, 2026
Merged

feat(safety): add module-first harness safety evaluation#2
JiaenLiu merged 115 commits into
mainfrom
codex/safety-measurement-evaluator

Conversation

@JiaenLiu

Copy link
Copy Markdown
Collaborator

Summary

  • add an independent post-run safety-audit substrate with typed outcomes, explicit missingness, immutable snapshot materialization, CLI execution, and separate audit sidecars
  • add the Proteus module-first harness-safety taxonomy for Agent Loop, Memory, Skills, and Tools, including non-adversarial fault models, bounded adversarial threat models, permission boundaries, and linked behavior/module verdicts
  • add plug-in case-suite and evidence-provider contracts plus a generic runtime that evaluates every snapshot from H0 through Hn, derives harness contribution and module causality, and records adjacent-snapshot transitions
  • expose proteus audit for instrument integrity and proteus safety --suite <module>:<object> for executable harness-safety suites
  • retain actual harness-native safety cases as plug-ins; no Aki-specific case implementation is built into the generic runtime

Motivation

Proteus previously measured harness evolution and task behavior but lacked an independent, harness-neutral way to distinguish model behavior, harness controls, module mechanisms, committed effects, missing evidence, and longitudinal safety change. This change establishes those boundaries without feeding safety results back into evolution or collapsing unlike cases into a scalar score.

User and developer impact

  • completed sweeps can be audited without modifying their source trajectories
  • adapters can bind native surfaces to the four canonical harness modules
  • external suites can plug in concrete scenarios and native evidence providers
  • every materialized snapshot, including H0, can produce independent behavior and module verdicts plus transition evidence
  • safety and task-evaluation outputs remain separate

Validation

  • uv run --no-sync pytest tests/ -q — 148 passed
  • Ruff clean for all new safety modules and tests
  • Ruff E9/F correctness checks clean for modified CLI and Minimal adapter files
  • git diff --check — clean

Score tools_permission_drift from independent canaries rather than
permission modules. Support policy mutation and minting on every
bundled harness, command execution on DSH, and notes-surface overwrite
and workspace cases on Minimal/LLM.
Use the same zstd fallback as the DSH adapter so CI on 3.10-3.13 can
collect safety tests, and skip Aki image builds when the local image is
absent instead of failing Docker Hub pulls.
@JiaenLiu
JiaenLiu marked this pull request as ready for review August 27, 2026 01:24
@JiaenLiu
JiaenLiu merged commit 040a343 into main Aug 27, 2026
1 of 6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant