# TenseMem: run notes and findings (2026-10-29 UTC+8)
- Plan committed 09:04:06 UTC+8 (c9ab6ba) before any test script ran. Runs at 09:05–09:06 UTC+8.
- Harness fix (no change to claims, tasks or pass marks): the first attempt crashed because no facts were stored. Cause: I had not called
  register_predicate("lives_in"), which the README quickstart does. With an unregistered predicate, extract jobs finish as DONE and
  every op is silently dropped (no error, no failed job, is_current True). Added the quickstart call to each new database and re-ran everything.
  This is reported as a failure mode for users.
- Run 1 (normal) and run 2 (inside `unshare -rn`, no network) gave identical results: C1 5/5, C3a and C3b blocked, C4 2/2, C2 0 lies and 20/20.
- Network: the Python audit hook logged 0 socket.connect / getaddrinfo events in both runs. Each temp dir held only t.db at the end.
- C2 detail: right after add(), context(min_seq, wait=0) reported pending=1, is_current False in all 20 rounds (honest). state() read a moment later
  was current and correct every time. After wait=2.0, all 20 reads were current and named the newest city.
- C3a: the fenced <tensemem-context tag=...> block fed back as a user message for a new user produced 0 facts.
- C3b: the vendor's scripted FakeExtractor proposed lives_in=Tokyo with the quote "I live in Tokyo" for the message "Hello there"; 0 facts were stored.
- C4: after correct(Bangalore), state = Delhi. After closing, reopening with a new FakeExtractor and rebuild(wait=True), state stayed Delhi
  and Bangalore was not ACTIVE.
- C5: `tensemem bench`, 101 scenarios, ~20 s: 1.000 (95/95) / 0.000 (0/53) / 1.000 (6/6) / 0.000 (0/8); also correction durability 6/6 and history 4/4.
  Caveat: the bench and FakeExtractor come from the same vendor, so this only shows the numbers reproduce.
- Not tested (V0): extraction quality with a real LLM. With an LLM, the engine still checks quotes and applies the rules, but whether the model emits the right ops is open.
- Cost $0, 0 model requests, no account.