Documentation

Exhibit A has one rule: a runnable fail-to-pass test, or silence.

A VERIFIED Case shows that an exact generated test failed on the reported state for the expected reason and passed on the fixed state under deterministic execution. It does not prove developer intent, complete program correctness, universal causality, or that a candidate repair should be merged. Research scores and study reports are descriptive. They never override the deterministic flip check.

Two real sealed Case replays: inventory_proven ends VERIFIED, while inventory_silence ends UNCERTAIN.

Produced by python3 -m exhibit_a.cli repro --replay on the checked-in inventory_proven.json and inventory_silence.json Cases. Replay makes no model call and executes no repository code. See the exact commands and capture record.

Read the project overview for the full product pitch, architecture, setup instructions, and Codex usage.

Evidence engine

Open science and EEF

Research instruments