Every test fired at werbos's live 1chat endpoint on SR8, one fresh session each: the 500-row gauntlet plus code, facts, commands, safety and multi-turn continuity. A failure is split by what werbos actually did: said it didn't know, or answered something that didn't match.
Each row is one full run of the same tests, from benchmarks/probe_history.jsonl. Dead-server errors count a crashed or unreachable server, not a wrong answer.
| Run | What changed | Pass | Dead-server errors | Server memory peak |
|---|
| Suite | Pass | Pass · abstain · wrong |
|---|
Every remaining failure in this run, grouped by category. Honest abstain means werbos said it could not ground the answer; answered wrong means it gave an answer that did not match.