From Claims to Proof Bank

Most research websites fail at the same point: a strong claim appears, but the path behind it is foggy. There might be a repo somewhere, a model card somewhere else, and a result buried in a note. That is not enough for a serious archive.

The Proof Bank is the correction. Every entry repeats the same fields: ID, status, claim, setup, result, limit, reproduce, related repo, related artifact, updated date, and confidence level. Repetition is the point. Strong claims should be boring to audit.

A concrete example is contradiction localization. TS-005 says the narrow claim is that contradiction can localize as residual/provenance tension in a small graph. The setup is 9 nodes and 16 constraints with one planted contradiction. The result is that the bad edge ranked first by residual energy and relief-if-removed, and removing it reduced tension near zero. The limit is small and planted. That is mature evidence language: useful, but bounded.

CIG extends that logic. Claims, evidence, sources, confidence, contradiction, and revision should be graph objects, not loose paragraphs. Proof Ranker extends it in another direction: proofs become traces that can be scored, repaired, and checked against verifiers. TensionLM extends it into model internals, where telemetry can expose pressure that ordinary output text hides.

What exists now is the route map and the initial proof entries. What is still missing is full replication depth: commits, seeds, exact scripts, artifact hashes, and public benchmark comparisons. The archive gets stronger every time a sentence becomes a receipt.