Bold Discovery · Hostile Verification · Neutral Assessment. Three roles hand off in sequence: Agent A is responsible for the highly divergent, high-confidence main push — it may propose bold conjectural bridges and does not stop researching out of fear of hallucination, but every unproved step must be flagged; Agent B does not generate, only hunts for problems — specifically the first illegal step, checking quantifier order, domain, uniformity, error terms, tail behavior, local-to-global jumps, counterexamples, and "proves-too-much" red flags (an argument strong enough to also derive a known-false statement); Agent C neither generates nor challenges, only renders a neutral academic judgment — what is currently the strongest defensible claim, what completion level it has reached, whether it is novel, whether QCI is closed, whether it's ready to submit, and what objections a referee would most likely raise.
Hunts specifically for the first illegal step; checks at minimum include:
Judges only:
If a case uses TRP, the suggested per-round presentation format is as follows (not yet applied to any existing case):
Round 001
Agent A — Candidate
Status: COMPLETE
Agent B — Audit
Status: GAP_FOUND
Agent C — Assessment
Level: 2
Next:
Repair Lemma L3
Single-Agent Protocol, Swarm Protocol, Blind Re-Derivation Protocol, Human Referee Protocol, Multi-Lab Replication Protocol — names only for now, see the protocols overview.