An AI answer usually arrives with no provenance: one model, one shot, trust it or don't. That fails exactly where the stakes are highest — a teacher deciding a plan for a child with complex needs, an engineer diagnosing a system that won't behave. PFP Darwinism is a reasoning arena that shows its work instead.

A population of small models starts at once on your problem, built on NVIDIA Nemotron models served through Nebius Token Factory. Each starter gets a model-generated exploration brief covering a different region of the hypothesis space, generated against the other briefs so the population covers the space instead of photocopying one answer. Branches split at uncertainty (Plaus), mutate blindly into unexplored regions (natural selection), merge when their claims are compatible, and die when a falsifier breaks their claims. Every event — spawn, split, merge, attack, fail — is written to a hash-chained journal, so any run is replayable and auditable end to end. You don't just get an answer; you get to see how it earned its standing.

What makes it different:

Coverage by construction: briefs are generated against each other, and every descendant's brief diverges from its parent's — inherit the map, never the route. On our test problems, solo small-model answers were unanimous and wrong; the arena population contained the correct direction where solo never produced it.
Test before merge: a branch cannot be absorbed into a merge until it has faced at least one falsifier attack. Selection before recombination.
Race mode: a time/call budget with an anytime winner — a teacher-grade answer in 5 minutes, a sprint answer in 60 seconds, for cents per run.
Proof trail as training signal: the run logs double as distillation data — verified reasoning trajectories are what a small model grows on.

Live demo replays (education problem, 5-minute run; SSH diagnosis, 60-second race) are hosted on the demo page. Source is MIT on GitHub with setup instructions in the README.

Honest status

Run 16 (2026-10-27, education problem, v0.53) produced the first gauntlet-verified winner's-box survivor: a visual schedule with timed cues and brief predictable sensory breaks plus content choice within the shared thematic unit. It survived a falsifier test, a contradiction check, three gauntlet rounds, and a pre-exit check; every objection raised was possibility-only and no concrete break was found. Journal head 610947c1056e3cbd05a7249e85b558ed8c19d89302f19a7322b0c0530b3ad231, verifiable with run_arena.py --verify. The system is in active development toward the v0.5 milestone.