terrizoaguimor

Hyphae Consensus makes AI-proposed code changes accountable. Before a change ships, three roles argue it out on Meta's models through the Meta Model API: a Proposer (muse-spark-1.3) drafts the change and its claims, a Critic (muse-wonder-1.0) attacks them, and an Evidence Evaluator (muse-wonder-1.0) judges every claim against records from Hyphae Agent Memory. The runtime, not a model, retrieves those records and verifies their recall proofs against a pinned anchor; a fixed policy, not a vote, turns the judgments into a recommendation; only a human can authorize. Every event is hash-chained, and anyone can re-check the chain in the browser. What we found: in our 6-run final battery, every first proposal had 1 to 3 claims that the stored records contradict, and none of those runs ended "supported" without revision. In the recorded run F03, the Proposer asked to raise payment retries to 10 while the project memory says retries are capped at 3; the Critic objected, the Evaluator confirmed it, the revision went back to 3 attempts with jitter, and the run still ended "contested" with 6 open objections instead of pretending to be done. When a model breaks the protocol the run fails closed: in run M04 the Evaluator left a contradiction unjudged and the runtime refused to issue any recommendation. Try it: https://consensus.terrizoaguimor.dev. Guided examples are free and need no account. The 53 recorded runs (1,103 records) are open to inspect. Live runs on preset tasks use the owner's key under a fixed budget, Turnstile and judge codes.