PFP Darwinism — AI that shows its work

AI that shows its work. Pass, Fail, Possible — survival of the fittest, for reasoning.

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. Branches split at uncertainty, mutate blindly into unexplored hypotheses, merge when their claims are compatible, and die when a falsifier breaks their claims. Every event is written to a hash-chained journal, so the whole run is a tamper-evident proof trail you can replay and audit — you don't just get an answer, you get to see how it earned its standing.

Watch a run

Education run — verified winner — A student with dyslexia, level 1 autism, and ADHD: group reading, disruptive stimming, no sense of when work starts and ends, a 504, and a refusal to do different work from the class. 12 branches, 88 calls, $0.03. Winner's-box survivor: a visual schedule marking work-period start and end with timed cues and brief, predictable sensory breaks, plus content choice within the shared thematic unit — so the student works on material that feels relevant without leaving the group's work. It survived a falsifier test, a contradiction check, three gauntlet rounds, and a pre-exit check; every objection raised against it was possibility-only, and no concrete break was found.

SSH race (60 seconds) — SSH works laptop→desktop but not desktop→laptop. 12 branches, 64 calls, $0.02, halted by the race budget. The falsifier attacked untested branches before any merge — test before merge, enforced in code.

Both replays render from the runs' hash-chained event journals. The education replay ends in a gauntlet-verified winner's-box survivor (run 16, 2026-10-27): journal head 610947c1056e3cbd05a7249e85b558ed8c19d89302f19a7322b0c0530b3ad231, verifiable with run_arena.py --verify <run.jsonl> --expect-head <head>. The SSH replay shows a best-verified answer at cutoff, labeled as such.

Why this matters

A teacher can't wait an hour, and can't afford to trust a black box with a child's plan. The arena answers in minutes for cents, and the answer arrives with its trail: which branches died, what broke them, what survived and why. That trail is also training signal — verified reasoning trajectories are the data a small model grows on.

How it works

Briefs, not photocopies. Each starter gets a model-generated exploration brief covering a different region of the hypothesis space, generated against the other briefs. Every descendant gets a brief that diverges from its parent's — inherit the map, never the route.
Falsification first. A branch cannot be merged away until it has faced at least one falsifier attack. Selection before recombination.
Race mode. A time/call budget with an anytime winner: at the deadline you get the best-verified branch alive, with its verification depth recorded.
Proof trail. Spawn, split, merge, attack, fail — every event hash-chained and replayable.

Built with NVIDIA Nemotron models (Nano 30B, Super 120B, Qwen3-235B) served on Nebius Token Factory. Source (MIT) · README and setup