PG
RESEARCH — SEALED EVALUATION
Designed panel only — not a screenshot. A sealed-program test range for agent evaluation.
The goal
Run task batches against agents under a sealed-program standard, so verification happens blind and results are compared only after unsealing.
What it does — the task
- Seal the program first
- Agents verify blind
- Unseal and compare results
- Disagreements go to a third party or the founder
- Batch 001 is complete and mirrored in Roman_Commons
- The home repo is currently ruleset-locked, which is stated here rather than hidden
Why it exists
Agent evaluation needs a standard that keeps verification blind and comparison honest. This is a research program — no savings are claimed here.
Status: Research program — sealed-program agent evaluation; Batch 001 complete.