Skip to content

Instantly share code, notes, and snippets.

@pulkit6732
pulkit6732 / integrity.py
Created May 31, 2026 16:43
Minimal stdlib reproducer: an autoresearch agent can lower val_bpb without training — plus a ~30-line integrity receipt that catches it. Run: python repro_gaming.py
"""
integrity.py — minimal experiment integrity receipt (Python stdlib only).
Writes ONE tamper-evident line per run that binds the reported val_bpb to the
exact code, data, and model that produced it. This is *detection*, not
prevention: enough to make a gamed run stand out in the overnight log and to
make any kept result reproducible.
Each receipt records:
- val_bpb the reported metric