Read-only diagnostic workflows per 12 cases
The six novel cases still require evidence. All six recurring exact fingerprints reuse a prior green provider receipt.
Twelve opaque CI incidents share deliberately non-identifying red summaries. The answer key is checksum-addressed and write-once sealed before the first model call. Stateless, raw-RAG, and Continuum receive the same bounded tools; only provider-verified exact-fingerprint memory can skip a real read-only diagnostic workflow.
The six novel cases still require evidence. All six recurring exact fingerprints reuse a prior green provider receipt.
Continuum, stateless, and raw-RAG all recovered 12/12 after the server compiled evidence into discriminated proposal schemas. The confirmed result is information value: six of six paired recurrence cases avoided a provider probe, exact p=0.03125.
Token cost did not fall: memory search and fetch used more model input tokens. This is an external-call and recurrence-latency result, not a universal cost or accuracy claim.
| Parent workflow | — |
|---|---|
| Source SHA | — |
| Artifact ID / archive SHA-256 | — |
| Challenge SHA-256 | — |
| Labels SHA-256 | — |
| Commitment / seal receipt SHA-256 | — |
| Public result SHA-256 | — |
Claim boundary. This is a preregistered twelve-case synthetic population over six reviewed CI fault families, an exact-fingerprint recurrence contract, and a real GitHub Actions provider. It proves bounded information-value and outcome-gated promotion for this population—not semantic transfer to changed environments or arbitrary repository repair.