AutoSeal

The question

A passing process or benchmark verdict can be misleading when candidate-controlled configuration changes which tests execute.

How it works

• Use a typed process runner, content-addressed receipts and separate proxy/audit views.
• Reproduce a disclosed failure on a pinned public case with explicit legitimate, invalid and exploit-bearing controls.
• Require fail-closed review of candidate identity, control roles and preservation checks before considering a repair.

Evidence & results

• On one selected SWE-bench Verified instance, seven native/adapter controls and separate semantic audits are archived.
• The audit accepted two valid fixes and rejected five other controls, including disclosed false-reward variants.
• A separate tiny training diagnostic established a real optimizer update but no difference between paired evaluator arms.

Scope & next steps

• Known disclosed controls, not blind discovery, a benchmark-wide attack rate or general efficacy.
• No repair promotion or downstream model improvement is claimed.
• The local execution setup is not a reviewed hostile-code security boundary; independent personal reconstruction remains open.