A Slack agent for fight-operations teams. It checks cross-jurisdiction suspensions and return-to-competition windows over source-cited commission records, corroborates against a live boxing record feed, and refuses to clear when fighter identity is statistically uncertain. Every decision lands in a tamper-evident, hash-chained audit ledger. A human always makes the final call.
The clearance engine's suspension-window logic is proven equivalent to an independently written safety specification over all dates and intervals, by the Z3 theorem prover. Not at build time. Right now, on this server, when you click.
The rule engine's window logic is Z3-verified against an independent
safety spec: an active suspension can never produce CLEAR. The proof found and killed
a real fail-open bug before launch.
The match threshold is conformally calibrated on thousands of query
variants built from the real fighter table. Two plausible candidates force a human
pick. Two real fighters named Bruno Silva is not an edge case here; it is the
point.
Verdict cards render only from the deterministic engine, never from model prose.
A live second source can tighten a verdict, never loosen it. The ledger
is append-only and HMAC-chained; you can export it to a Canvas and hand it to a
commission.
In Slack: ask about one fighter or paste a whole fight card; every bout lands on one board, banded CLEAR, DO NOT CLEAR, or NEEDS PICK with the blocking record cited. Click into the live Z3 proof or the hash-chained audit trail from any verdict.
On its own: a daily roster monitor re-checks every suspension window (the failure that killed Tim Hague was a lapsed window nobody re-verified) and pushes a deterministic digest to the ops channel. No model decides or phrases an alert. Quiet days send nothing.
Honestly scoped: curated, source-cited records, not an exhaustive registry. A CLEAR means no recorded suspension matched; commissions remain the source of truth.