Skip to content

fix: prevent self-audit 'Violates Known Physics' example bleedthrough #585

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
neoneye wants to merge 5 commits into
base: main
Choose a base branch
from
Open

fix: prevent self-audit 'Violates Known Physics' example bleedthrough #585

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
2b29b17
fix: prevent self-audit "Violates Known Physics" example bleedthrough
neoneye Apr 17, 2026
67ec1c6
strip domain-specific carve-outs from physics instruction
neoneye Apr 17, 2026
c67c006
rewrite physics instruction with positive-only framing
neoneye Apr 17, 2026
6fbd0cd
add SI-units gate to physics check
neoneye Apr 17, 2026
3b77f73
simplify physics check to default-LOW gated on hardware presence
neoneye Apr 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion worker_plan/worker_plan_internal/self_audit/self_audit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,7 +105,7 @@ class ChecklistAnswerCleaned(BaseModel):
"index": 1,
"title": "Violates Known Physics",
"subtitle": "Does the project require a major, unpredictable discovery in fundamental science to succeed?",
"instruction": "Scope: fundamental physics ONLY (e.g., perpetual motion, faster-than-light travel, reactionless/anti-gravity propulsion, time travel). ‘Laws’ here means ONLY the laws of physics (thermodynamics, conservation of energy, relativity, etc.) — NOT legal statutes, treaty law, constitutional law, regulations, or policy. A plan involving legislation, treaties, currency adoption, governance reform, or any non-physics ‘laws’ does NOT violate physics — rate LOW. HIGH only if success literally requires breaking a named law of physics; MEDIUM only if a physics-consistent but unproven physical effect at required scale is mandatory with no conventional fallback; otherwise LOW. Economics/crypto/tokenization/governance/AI/regulation/policy/finance/engineering-scale are out of scope—rate LOW. If you cannot name a specific law of physics (e.g., second law of thermodynamics, speed of light) that is violated, rate LOW. If LOW: Mitigation=None. If ≥ MEDIUM: name the specific physical law violated, ≤30-word justification + mitigation with Owner/Deliverable/Date.",
"instruction": "This check applies only to plans describing physical devices or material processes. Default rating: LOW. Rate HIGH only if the plan requires a physical device or material process that cannot exist under known physics. If LOW: Mitigation=None. If ≥MEDIUM: ≤30-word justification + mitigation with Owner/Deliverable/Date.",
"comment": "If the initial prompt is vague/scifi/aggressive or asks for something that is physically impossible, then the generated plan usually end up with some fantasy parts, making the plan unrealistic. Known false-positive: LLMs confuse ‘laws’ (legal/regulatory) with ‘laws of physics’ for plans about policy, currency adoption, governance, etc."
},
{
Expand Down