README.md

Live-Qwen run proof

live_run.db is a fully live society run — every agent turn and mediator ruling came from real DashScope calls (qwen3.7-plus role agents, qwen3.7-max mediator), not the offline FakeQwen. Reproduce the verification:

tarmac verify-log docs/proof/live_run.db

live_run.verify.txt is the captured output: 324 entries, all invariants PASS, including I5.replay: replay reproduces the recorded manifest byte-for-byte — i.e. the live run is cryptographically replayable, identical to its recorded manifest, exactly as an offline run is.

Scenario storm_dfw, seed 7, 4 rounds, 3 mediator rulings.

The --live path was hardened to make this possible: role-agent calls run with Qwen3 "thinking" disabled (it otherwise spends thousands of reasoning tokens per call and times out on the large negotiation-state prompts), a per-request client timeout guards against a single slow call hanging the run, and the mediator's ruling application now drops citations to source ids that aren't in the regulation library (a live model can invent them) instead of aborting the whole run.