Back to projects

Published project

Proof Ranker

Small proof-control ladder for proof scoring, repair, verifier loops, and active-frontier experiments.

Claim discipline: Claims on this page should be treated as hypotheses unless linked to a receipt, benchmark, or reproducible artifact.
Claim
Setup
Result
Limit
Replay

What problem it addresses

Reasoning traces need cheap ranking, local failure detection, repair attempts, and verifier feedback before expensive model passes.

How it works

Publish a v0-v4 ladder: existence classification, path quality ranking, local tension failure detection, chain repair, and generator/ranker/verifier active-frontier loops.

What exists now

Published model ladder exists; repo-level routing still needs exact confirmation.

Evidence / receipts

Hugging Face artifacts v0-v4 form the public model ladder. Related proof-bank entries should attach exact verifier receipts as they land.

Limits

Small proof-control artifacts are not general reasoning capability proof. They are bounded verifier-loop experiments.

Next milestone

TS-Reasoner / TensionProofLM integration.

How to run / inspect

Install / open

Open the Hugging Face model cards and pair them with verifier scripts before use.

Inspect

Compare v0-v4 model cards and verifier-loop notes.

Links

Focus tags

Proof scoringTrace repairVerifier-backed loops

v0: proof existence classifier.

v1: proof path quality ranker.

v2: local tension and failed-step detector.

v3: chain repair.

v4: generator/ranker/verifier active-frontier loop.