Workbench source: b0dcc50336b1296caa022fb8b5e775d537e3930b, PR 46. Compiler: clean candidate 8951c5f8f22526e9cc1225a4dd47e63a6d420c4a, PR 1410. Both were built with Go 1.27.2; the compiler uses decision runtime 0.2.26. The installed public compiler remains 0.6.18 at 13d576ef.

clean-cli-dogfood.tar.gz retains the complete original output of sequential fixed-order and own-model CLI calls against the committed before/after examples. This includes both compiler interface receipts, structural facts, input rows, generated policy, execution and saved replay, final report, identities and timings. dogfood-api.sh and its log retain the exact comparison. Source, model and binary hashes did not change; workbench source was clean. An earlier macOS Bash empty-array harness error occurred before invoking the CLI and was corrected before this run.

Both calls reported three changes with identical original/new values and assessments: input-tightened, output-weakened, field-added. Actual changes were input-only observations (0/0 labelled checks). The policy's 17 authored selection cases checked 51 fields, all passing. Both modes attempted eight combinations. Generation model calls were 0/1; saved replay used zero new inference.

The existing model was models/graph-chooser-20261008/all-data-demonstration/qat_ternary/model.json, SHA-256 3c68205a660695103712e2115ce90accd298b406dced8ad5ad5c9b84bab0a202. It was neither retrained nor newly downloaded. Model prediction took 199,542 ns; record generation took 17,870,208 ns versus deterministic 18,215,625 ns. These are this program's observed phases, not general inference benchmarks.

Whole command	Elapsed s	User s	System s	Peak RSS bytes
Fixed order	1.64	0.40	0.26	87,834,624
Own model	0.77	0.40	0.23	86,622,208

These are one sample each with different order/cache conditions, including compiler/native child processes. Host CPU utilization was not measured. No causal speedup or arbitrary-code correctness is inferred.

tests-final.log: full native go test -race ./... passed (209.344 s) before the last surviving-field order fix. An earlier test fixture accidentally contained Go test text inside its Gooo source string; the failing log remains in the local work record and the fixture was corrected before the full passing run. field-order-red.log reproduces an addition hiding a simultaneous reorder. api-tests-publish.log verifies all API/native/race cases on the final source (6.127 s), including that regression. go vet ./..., formatter and executable go build ./cmd/... passed. The repository's pre-existing archived main-package fragments prevent claiming a successful whole-tree go build ./....

The batched policy test uses 27 distinct policy inputs repeated to 135 rows, crossing the 128-row boundary in deterministic and model modes. This is batching evidence, not 135 independent experiments. Proposed next operations remain data; the tool does not automatically repair callers or prove behavior compatibility.

pr1410-proof/ retains the original CI artifact 11594515290 from run 37878887257, attempt 1, together with metadata, extracted proof/provenance and verification. ZIP SHA-256 is 183b533d6ed6a12ca9080d0d5902ec17c2f5cfe5aa7d8b557319abdd9fe62d35. The six canonical checks and source/base tuple passed. This is not a merge record.

The PR's separate readiness run 37878887293 failed on Windows during text-operations saved replay: EXECUTE_1, 3,411,652,400 ns, timeout true, no stdout/stderr or recorded CPU time. windows-original.zip is original artifact 11593048454, SHA-256 c82361024185a05553de2375e812e179dec45626f4908d6e0c40ca8da955a5e8. The failed replay JSON was missing from that artifact and was extracted unchanged from the original job log into windows-first-failure.json.gz.

After that run ended, a single diagnostic dispatch of the same head was run as 37880504176. It also failed on Windows, at package construction with actual inputs: EXECUTE_1, 2,953,490,000 ns, timeout true, no stdout/stderr or recorded CPU time. Other three platforms passed. Original artifact 11594265997 is retained as windows-reproduction-original.zip, SHA-256 f6b00f2d651e70d61631d3dc1e7fe0c4281b2a73e9241dac36ef02c0a5f3c44f. Its original log JSON is windows-second-failure.json.gz.

Both calls used the existing two-second native execution deadline. These records do not establish the underlying cause, including whether startup was delayed. No timeout was extended, original run restarted, or failed check bypassed. Compiler PR 1410 remains open; workbench PR CI is in progress. The dispatch is diagnostic evidence and cannot authorize a release.

main188-proof/ retains the successful push/main run 37877948144 at 188e1c2ec3b0a85fdb07ff618693c7a0b45bba65, artifact 11593787476, original ZIP SHA-256 941f41baab930346978fa1199328bc2f9b54d30c6806de456761abf1f9eaae8d. The original archive digest, push tuple, six checks and provenance passed the existing frozen proof verifier. This closes the previously pending post-merge check for public 0.6.18; it is separate from the new interface candidate.

The initial b0dcc503 PR/push runs 37881317701/37881313997 both succeeded. After both ended, commit e14e5e67100b701f51f4dcc9c27de29f95003028 added one artifact path, out/ci-api-diff/, so CI also preserves the new original outputs. No executable, recipe, model or example changed relative to the clean CLI sample. Neither initial run was canceled or restarted, and the PR body stayed unchanged.

Final PR run 37881747837 and push run 37881743956 both succeeded at e14e5e67. The original 11,519,442-byte artifact 11594312641 had SHA-256 a656c11b2fa1ba38bff2323eb5c990c5d697c22151cde65c16990170622c5948. The archive was downloaded and its digest checked. To limit duplicate storage, workbench-pr46-ci/ publishes the original compiler/API output subset and artifact/run/job metadata; the complete archive remains in GitHub Actions and the local work record. The subset preserves original bytes.

Compiler identity, three classifications, 51/51 selection fields, actual-input 0/0, and saved zero-inference replay were verified from that original archive. With both runs successful, the live PR was clean and matched the expected head and main base. Normal expected-head squash merged PR 46 to fd53420e3851c46870f846086bc152418f936d4b; its tree equals the checked candidate, and its sole parent is the previously observed main 4f4fcf25. Local main was updated fast-forward only. merge.json and merge-verification.txt retain this.

Post-merge main run 37882197317 was still in progress at this update. This does not change the unresolved Windows readiness status of compiler PR 1410.