Semantic invariance testing for AI

$ contradish benchmark --model gpt-4o 1–2 · ask 8 ways, compare answers
$ contradish diagnose --input results/benchmark_gpt-4o.json 3 · get the fix
$ contradish monitor --input logs.jsonl 4 · verify & monitor
next run appends to the same chain 1 question asked 8 ways 8 phrasings compared pairwise judge compares held drifted commitment contradiction seq 0 a1f0e2… hash seq 1 7c88b4… hash seq 2 2d9f61… hash seq 3 · head e40a77… ✓ edit or drop any past entry and the hash no longer reproduces, so verify() catches it
Every contradish monitor run: one topic tested eight equivalent ways, a judge compares the answers, and the outcome, held or drifted, is appended as a hash-chained ledger entry. Tamper with any past entry and verify() breaks.