Verified recovery across 12 target environments
Raw-RAG reuses every highly similar memory. Continuum reuses only six provider-attested same-cause memories, then acquires current evidence for all six counterfactual near neighbors.
CONTINUUM · TRANSFER FIREWALL
S3 PREREGISTERED · BEDROCK · 84 GITHUB RECEIPTS
Similar failure. New environment.
Reuse only when the cause transfers.
Six provider-verified source outcomes are each paired with a changed-environment same-cause target and a deceptively similar near-neighbor target. The model never sees the relationship, expected patch, causal signatures, or scoring policy. A server-owned provider attestation—not semantic similarity—decides whether memory may authorize an action.
Technical summary
Public gate—workflow + artifact + S3 seal
Continuum recovery—all changed environments
Safe same-cause reuse—zero diagnostic probes
Near-neighbor rejection—false transfer remains zero
Raw-RAG recovery—retrieval-only transfer baseline
Raw-RAG false transfers—near-neighbor cases
Paired exact p—same-cause probe reduction
Child workflows—18 + 12 + 18 + 36
Outcome gating separates useful transfer from semantic imitation
Interpretation
Continuum preserves stateless recovery while saving one diagnostic provider call on every same-cause pair. Against raw-RAG it prevents every registered false transfer and improves verified recovery by —.
Target attestations are shared benchmark inputs. The six avoided candidate diagnostics are real, but this experiment does not claim fewer total GitHub workflow runs or universal latency superiority.
Scope and metric definitions
| Metric | Denominator | Meaning |
|---|---|---|
| Same-cause verified transfer | 6 changed-environment targets | The admitted source memory is cited, no diagnostic runs, and the target remediation receipt succeeds. |
| Near-neighbor safe rejection | 6 similar-symptom targets | The source cause differs, memory is not adopted, one read-only diagnostic runs, and the target recovers. |
| Verified recovery | 12 targets per arm | The actual GitHub Actions remediation receipt reports success. |
| Receipt integrity | 84 provider runs | Unique run, artifact, and digest identities; exact source SHA; repository mutation and cleanup residual both zero. |
Preregistered execution contract
01SealChallenge and evaluator labels receive separate checksums and write-once S3 receipts.
02CalibrateEach source failure produces baseline, wrong-patch, and green provider receipts.
03AttestEach target emits a separate read-only causal signature from actual workspace facts.
04ConstrainOnly a compatible signature exposes memory fetch and its discriminated proposal tool.
05ScoreA separate remediation workflow and controller-only labels determine the outcome.
Robustness and failure evidence
The first policy run failed closed
All actions recovered, but three rejected memories remained in final citation arrays after current diagnostics. The gate correctly failed. The server now hides incompatible fetch tools and independently rejects any citation that is not fetched, admitted, and authorizes the exact patch.
The first transport run also remained non-evidence
A transient read-only artifact download disconnected before scoring. Bounded retry was added only to GET operations; provider dispatch and effects were never retried. The failed parent remains visible and was never promoted.
Checksum-bound lineage
| Parent workflow | — |
|---|---|
| Source SHA | — |
| Artifact ID / archive SHA-256 | — |
| Challenge SHA-256 | — |
| Labels SHA-256 | — |
| Commitment / seal receipt SHA-256 | — |
| Public result SHA-256 | — |
| Fingerprint overlap | — |
Limitations and next question
Bounded claim. This proves provider-attested transfer and rejection for six reviewed synthetic CI fault pairs with disjoint source and target fingerprints. It does not prove arbitrary repository repair, open-world semantic generalization, or that a separate target attestation is free. The next architectural question is whether the same receipt-to-action lineage remains intact when canonical source memory is persisted, vector-retrieved, and scope-filtered through CockroachDB in the live decision loop.