Back to projects

Published project

Proof Ranker

A bounded model ladder for proof existence, path quality, local failure detection, repair, and generator/ranker/verifier loops.

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

Candidate traces need cheap local ranking and repair signals before expensive or authoritative verification.

How it works

Train and publish a staged family of small classifiers/rankers, then keep the typed verifier outside the learned proof boundary.

What exists now

A published support experiment rather than a separate flagship programme.

Evidence / receipts

Public Hugging Face artifacts cover proof existence, path quality, local tension/failure detection, chain repair, and active-frontier loops.

Limits

These are narrow proof-control artifacts, not general reasoning capability.

Next milestone

Use only where it improves a measured TS-Reasoner pipeline.

How to run / inspect

Install / open

Open the model cards and pair learned outputs with typed verifier scripts.

Inspect

Compare the v0–v4 task boundaries and ensure no learned score is treated as proof authority.

Links

Focus tags

Candidate rankingFailure localisationVerifier separation

v0: proof existence.

v1: path quality.

v2: local failure detection.

v3: chain repair.

v4: generator/ranker/verifier loop.