CONTINUUM · CLOSED-LOOP CI
REAL GITHUB ACTIONS · 54 CHILD RUNS · ZERO REPO MUTATION

A red workflow is not memory.
A verified green receipt is.

Six reviewed CI fault families are calibrated with a real red baseline, a real red wrong patch, and a real green repair. Stateless, append-all raw-RAG, and Continuum then face the same novel and recurrence fixtures. Every score comes from an exact GitHub Actions run and artifact receipt.

Public gateexact workflow + artifact
Child workflows18 calibration + 36 evaluation
Continuum recoveryprovider-verified
Raw-RAG recoveryappend-all memory
Stateless recoveryno memory
False promotionsraw-RAG / Continuum
Recurrence successraw-RAG / Continuum
Canonical precisionraw-RAG / Continuum
Provider p95raw-RAG / Continuum
End-to-end p95raw-RAG / Continuum
Lift vs raw-RAGdescriptive, percentage points
Lift vs statelessdescriptive, percentage points

Verified recovery, same twelve fixtures

Provider-verified success

The stateless arm is intentionally visible: this benchmark does not establish general recovery superiority.

Continuum
Stateless
Raw-RAG

Closed-loop contract

01Observe redExact fixture diagnostic, never an inferred success.
02Propose bounded toolOne of six reviewed discriminated patch actions.
03Execute externallyA disposable GitHub Actions child run applies the proposal.
04Reconcile receiptRun conclusion, artifact digest, head, cleanup, and mutation state are bound.
05Promote only greenFailed raw history cannot enter Continuum canonical memory.

Checksum-bound lineage

Parent workflow
Source SHA
Artifact ID / archive SHA-256
Challenge SHA-256
Population SHA-256
Public result SHA-256

Claim boundary. These are twelve source-defined synthetic fixtures with explicit diagnostics and six reviewed patch tools. The result proves real receipt-bound red→green closure and failed-memory isolation. It does not prove arbitrary-code repair, broad CI-population generalization, or an advantage over the stateless arm; stateless also recovered 12/12 and was faster at p95.