Campaign friction log

Record entries while the campaign is running. Preserve the earliest actionable failure and distinguish product defects from model behavior.

Time	Stage	Category	Earliest actionable evidence	Immediate disposition	Durable fix-forward artifact	Independent verification	Status
2026-09-12T13:12:00Z	RESEARCH	missing-identity	Worktree launcher reports 1.0.0 while source release is v1.2.1.	Continue because module path and commit are exact; retain mismatch.	Ticket if still present after campaign.	Worktree module path and commit check.	open
2026-09-12T13:14:00Z	RESEARCH	manual-workaround	Repo-scoped controller tools unavailable in this session.	Use verified worktree CLI fallback.	Environment attachment explicitly deferred.	Tool catalog inspection.	open
2026-09-12T13:18:00Z	RESEARCH	missing-identity	Product and Windows GPU views succeeded; Docker Desktop lacks nvidia-smi and a stale WSL distribution could not attach.	Do not infer empty lane; use managed owner inventory.	Unrelated WSL repair deferred.	Recipe-container inventory plus Windows telemetry.	open
2026-09-12T13:15:00Z	RESEARCH	ambiguous-output	Healthy incumbent remains loaded but current registry no longer resolves its selector; one older same-family container is exited unhealthy.	Preserve exact labels/image/revision and retained healthy container for restoration.	Added guarded models recipes stop/start support so a registry-independent exact container can be retained and restored.	349 focused CLI/model lifecycle tests plus real incumbent stop/start dry-run.	resolved-before-mutation
2026-09-12T13:41:04Z	BASELINE	model-behavior	Incumbent returned 107–126 of the requested 128 exact code words in all five strict capacity requests.	Preserve the 0/5 strict result; use the same observe-mode workload only for descriptive latency.	No product change: strict mode already failed closed and retained sanitized response timing.	Incumbent deterministic preflight and repeated quality suites passed independently.	open-model-defect
2026-09-12T13:47:41Z	DOWNLOAD	operator-error	First live cache-inventory retry used unsupported --json-out and failed before inspection.	Read focused help and rerun with supported --output.	No code change needed; CLI usage error was explicit and non-mutating.	The corrected inventory showed one growing Signal partial and the download later verified exact completion.	resolved

Live fix-forward addenda (2026-09-12)

• The earlier lifecycle row was provisional, not a general-release approval. Independent review found missing-registry, fingerprint handoff, readiness, deadline, admission, and transport-contract gaps. Regression tests reproduced and fixed them through the managed CLI/MCP path. General merge/release remains held on the shared host-lifecycle lock and process-identity allowance contract; see the tracked retained-lifecycle/evidence-gaps ticket. Serialized campaign use was conditionally accepted. A live restart preview refused the active Qwopus peer with no mutation.
• Windows reports the pre-existing camera utility as a compute client. It is protected starting state, not a model process to terminate. The explicit baseline-PID allowance keeps the Docker-peer and 1024 MiB memory guards; immediate process identity/start-time verification is required for restoration.
• Signal and Swift each failed all five original strict 128-word requests. Swift still failed 5/5 with speculation disabled, ruling out speculation as a sufficient fix for that case. A separately declared 32-word diagnostic then passed 5/5. Preserve the original failure; never rank repetitive cap-hitting output as useful speed. Matched Signal no-spec also failed all five128-word requests, so removing speculation is not a sufficient fix for either model.
• Signal and Swift reasoning scouts each passed 9/10 and exhausted the same computer-science question's 5120-token total budget. A larger-budget retry fixture preserves the original question and independent answer key.
• Qwopus reasoning scout passed 1/10 under exact output syntax: eight visible answer-format failures and one reasoning-budget exhaustion. Disabled mode passed 24/30 repeated attempts, with two questions exhausting the1024-token visible budget. Reinforced syntax plus2048 visible tokens passed7/10; the mixed mitigation is diagnostic, not a matched performance cell.
• The first full test run stopped after5332 passes/340 skips on the intentionally changed MCP public-schema snapshot. Updated the snapshot after reviewing the typed allowance/stop/start contract; focused foundation tests passed. A fresh full run is in progress. No raw benchmark failure was overwritten.
• The documented old eval benchmark run spelling failed at parsing; focused help selected the supported eval benchmark capacity command. External-only quality normalization also emits warnings for unselected built-ins. Both are recorded in the retained-lifecycle/evidence-gaps ticket, not disguised as model failures.

All telemetry and logs cited here are diagnostic; a passing code test does not qualify a model. Final dispositions must be reconciled before publication.

Windows Pi regression investigation

The second full run stopped after7961 passes/350 skips with Pi exit134. A bounded isolated-process reproduction retained the earliest error: Assertion failed: ncrypto::CSPRNG(nullptr, 0). The test's minimal environment omitted SystemRoot; supplying only that Windows prerequisite changed the same startup probe from exit134 to a successful RPC response. The fixture now preserves that one OS value without inheriting credentials.

The real smoke then exposed an independent transport defect: Windows select() cannot poll subprocess pipes (Python's official select documentation). A new real-child pipe regression failed before the fix. The Windows path now uses stdlib _winapi.PeekNamedPipe and reads at most the reported available bytes; POSIX behavior is unchanged. Six focused tests, including actual Pi0.85.1 history/fork/resume/model controls, passed. Independent review and a fresh full run remain required. Durable code and tests are in anvil_serving/workbench_app/pi_rpc.py and tests/workbench/test_pi*.py.

Segmented HTTP test-reader correction

The next full run stopped after 2,103 passes and 323 skips. Authoritative router stderr showed the intended 504 startup_timeout; the test retained only headers because its unframed 0\r\n\r\n search matched the last request-ID header ending in hex zero. A deterministic segmented-socket regression reproduced this before the fix. The test helper now searches only the body for that final-chunk marker; other marker searches and production router code are unchanged. All 19 request-runtime tests passed. A new full run is in progress. Durable artifact: tests/router/test_request_runtime.py.

Final operational dispositions

The time-ordered entries above retain what was known at each failure; this section supersedes their open campaign statuses without deleting failures.

• Installed-version drift, unavailable session controller tools, and stale WSL attachment are deferred product/environment gaps in the tracked lifecycle ticket. Exact module/commit identity and the managed CLI provided a verified execution path; no raw Docker mutation or unrelated WSL repair was used.
• Managed retained-container restoration completed with exact camera process identity/start-time checks, no competing candidate, and a healthy incumbent. See restoration.json, recovery-admission.json, and restored-state.json. General lifecycle release remains held on shared locking and immutable process-identity attestation; bounded campaign acceptance is not release.
• Signal no-spec's larger 9,216-token retry still exhausted its reasoning budget without an answer. This also changed speculation versus the original scout, so it is not a one-variable budget comparison. Qwopus syntax/budget mitigation improved strict passes to 7/10 but left wrong answers and a cap. Minitron retained wrong-answer and output-adherence failures. These profiles are rejected as replacements under the tested contract, not silently fixed.
• The revised 32-word diagnostic passes for incumbent, Signal no-spec, Swift no-spec, and Qwopus do not erase the original 128-word failures. Swift's warm cache is separate from the other cold-cache populations. CPU contention was not controlled; no diagnostic is promoted to a finalist latency result.
• Independent Sol review accepted the bounded Pi pipe correction and HTTP test-reader correction with no blocking findings. Focused gates passed; final broad regression verification is recorded separately when complete.
• A publication regression caught this new dossier's missing canonical Review narrative label after 2,918 passes and 325 skips. Restoring the required dossier structure passed all 19 benchmark-documentation tests. Evidence-consistency tests also caught an unfinished summary and a terminal-newline mismatch in the recipe copy; both were reconciled against native artifacts and the exact registry, not by weakening tests.

Reusable operational lesson: restore a registry-missing incumbent by its retained immutable container only after independent admission and identity checks, then prove functional health. The engineering-learning skill is unavailable in this session; this evidence-linked invariant and the tracked ticket are the durable fallback.

Final verification closure

The final full suite passed 8,117 tests with 391 skips in 381.30 seconds. Ruff, 22 focused documentation/evidence checks, strict MkDocs, 669-file Markdown link validation, five template JSON parses, skill validation, and the full 1,164-file CLI reference audit also passed. The skipped tests do not prove missing live client, full-context, or model-quality gates.

Independent evidence review caught a missing dedicated Signal license source entry. The pinned publisher card declares Apache 2.0; the source registry now retains its exact URL, date and limited interpretation. Windows checkout conversion could also change the two canonical recipe bytes while evidence copies are byte-preserved. Exact .gitattributes entries now preserve those registry bytes. Native artifact line endings and intentional registry terminal blank lines are retained for their hashes; whitespace validation explicitly accepts CR-at-EOL and those terminal blanks. No evidence values were changed.