Governance, safety & compute-budget controls for the Phase-178 autonomous-research stack (TODOS/phase-178 §178.11). All logic here is real, self-contained rule/stat-based control — no external service required.
Components#
| Module | Summary |
|---|---|
budget.py |
Compute-budget governor: GPU-hours / dollars / wall-clock caps enforced simultaneously per run, researcher and org; a charge over any enclosing cap is atomically rejected and hard-stops that scope; approve_resume clears the stop and optionally raises the cap. (§178.11.1.1) |
guards.py |
Self-amplification guard — flags edits to protected scaffold/budget/governance paths or text requesting self-modification / budget escalation, routing to human approval; distribution-shift detector — Mahalanobis distance + Population-Stability-Index alarms on rapid divergence from validated work. (§178.11.1.2, §178.11.1.3) |
review.py |
Adversarial self-review — the CLAUDE.md stub bright-line as an executable gate: phrase patterns, silent-stub code-patterns (Math.random/NotImplementedError/fabricated-score), eval-contamination (n-gram overlap) and LLM-regurgitation; only stub-free trials accumulate delta credit. (§178.11.2.1) |
dualuse.py |
Dual-use risk classifier (defensive) — bioweapon/CBRN, cyber-offense, child-safety, election-interference lexicons with benign-context discounting; explicit weapon markers force HIGH → halt + escalate. (§178.11.2.2) |
telemetry.py |
Kuan Yin integration — normalises every safety event into one ModerationEvent schema so autoresearch and product safety share a pane of glass. (§178.11.2.3) |
report.py |
Quarterly Accountability Report — trials/accepted-deltas/novel-results/cost/safety-events/reviewer-outcomes as a dict + deterministic Markdown. (§178.11.2.4) |
Note:
review.pyanddualuse.pyintentionally contain stub-scan phrases and risk lexicons as string literals — that is the detection vocabulary, not stubs.
Testing#
bash
python3 -m pytest tests/ -q # 21 tests
python3 -m ruff check .