SScarTissue

RESEARCH EVIDENCE GATE

Decide before you ship.

A compact workbench for comparing weight-space repairs without turning an attractive metric into a safety claim.

Readiness ledgerResearch articleStored evidence only · no model runs
01 · DECISION GATE

No repair clears the evidence gate

Set a policy, inspect the reasons, and export the decision. The bundled study is deliberately conservative.

Research result

Load a valid bundle to evaluate repairs.

Accept, reject and insufficient evidence are research-policy outcomes—not production approval.

02 · STUDY SETUP
A

Study identity

Confirm what was measured before changing the policy.

Imports verify prompt and answer hashes, not the truth of declared controls or metric provenance.

Source hashes and limitations
B

Your research policy

Illustrative bounds, not safety recommendations.

Required study identity

Loading evidence…

03 · EVIDENCE

Compare the trade-offs

Lower harm and over-refusal, higher utility, fewer repair-selected scalars.

Point estimates · marginal intervals

Harm-ASR is majority-of-three judge harm, not refusal keywords. Utility is lm-eval strict-match. Calibration failed: attractive estimates cannot clear the evidence gate.

04 · INSPECT

Replay one prompt across every arm

Open the answer only when you need to inspect the underlying judgement.