eval(deltawire): prepare 72-run preflight and semantic oracles

[bigboateng]

Superseded experiment provenance. This conflicted PR is retained only for historical and raw evidence. The clean integration is PR #106. The final v4 evidence commit is c200dc0e.

Purpose

This lineage established fail-closed DeltaWire treatment delivery and semantic verification before any larger evaluation. It never authorized or launched the 72-run experiment.

Evidence summary

Overhead control: matrix-medium/r1

One real adjacent B0/D1 pair completed and both arms passed. D1 had the skill configured but did not activate or invoke DeltaWire. D1 reported 39,027 input-plus-output tokens versus 24,397 for B0, approximately 60% more. This pair is excluded from confirmatory estimates.

Historical probe v1

Delivery passed, but semantic verification failed: 999 {id,value} records were produced instead of 500 ordered {index} records. Derived status: delivery pass, semantic fail, end-to-end fail.

External-contract probe v2

Exactly one D1-only probe ran. Semantic verification passed, but end-to-end delivery failed because the structured transcript lacked gemini skills list evidence and the frozen artifact list did not retain .deltawire/config.json. No pair was launched and v2 was not rerun.

Recovery probe v3

Exactly one D1-only probe ran. Semantic exact match, discovery, activation, lifecycle, hashes, retention, and independent contract verification passed. The frozen parser rejected which deltawire as binary-path proof, so delivery and end-to-end status failed. No release or pair was launched.

Measurement-recovery probe v4

Exactly one D1-only range-large-v4 probe ran on tbench-c4d using the frozen v4 manifest and binary.

• Harbor 0.20.0 completed without exception with reward 1.000 in 261.440381 seconds.
• Semantic exact match passed for all 500 canonical {index} records.
• Structured discovery and activation, public-contract access, DeltaWire lifecycle invocation, config/schema/plan/state/output retention, schema integrity, and both contract checks passed.
• The deterministic verifier-phase environment generator produced a passing receipt: expected path/realpath, executable regular non-symlink binary, exact version and zero exit, binary SHA, and config/schema/public-contract hashes all passed.
• Harbor could not retrieve .deltawire/environment-receipt.json from the main task container after the verifier wrote it in the verifier container. The independently retained environment receipt was therefore missing.
• Fail-closed status: environment fail, treatment-use pass, semantic pass, end-to-end fail.
• Diagnostics: 934,798 input tokens, 10,344 output tokens, 750,963 cache tokens, and 261.440381 seconds wall time.

No probe-release.v2 was created. The authorized range-large/r1 pair was not launched. Under the one-shot stop condition, v4 was not repaired or rerun and no v5 was created.

Validation and publication

• Infrastructure workflow passed at 5a8ee779: run 29923012991.
• Frozen manifest-head workflow passed at 55b8c499: run 29923193058.
• Final evidence-head workflow passed at c200dc0e: run 29924758152.
• Full local validation passed on the clean integration projection: Go tests/race/repeat/vet, determinism, 12 canonical contracts, 96 mutation failures, plan checks, fidelity/release/rerun gates, binary parity, matrix/readiness, and production cleanliness.
• All 22 checks on clean integration PR eval(deltawire): integrate treatment-fidelity preflight #106 passed across Linux, macOS, and Windows; GitHub reports the PR as cleanly mergeable.
• Exact credential-byte and generic secret scans found zero credentials in v4 evidence.
• Raw trajectories, logs, and expanded output remain only in this evidence lineage; PR eval(deltawire): integrate treatment-fidelity preflight #106 contains implementation and compact provenance records.

Readiness

READY_FOR_72=false. Remaining blockers include the incomplete six-run canary, no final confirmatory manifest, and no explicit 72-run approval. The v4 pair and 72-run experiment were not launched. No savings, cost, statistical-significance, or family-wide claim is made.

[bigboateng]

Superseded by the clean integration PR #106, now merged as e3e085d. This PR remains immutable experiment provenance. Final v4 evidence: c200dc0. Passing v4 validation: https://github.com/operatorstack/intelligence-flow/actions/runs/29924758152. No post-hoc v4 release was created; READY_FOR_72 remains false.