0.5.3rc1: campaign audit repairs

This candidate implements the first correctness and observability repairs from the stagnation audit. It does not claim better scientific completion rates from a new ten-hour run.

Included

  • Provider payment/quota errors are distinguished from temporary disconnects. Affected studies pause with an actionable explanation. Resume uses the existing deadline; it does not silently buy another allocation of wall time.

  • Instrument prefixes are explicitly evaluated as alternatives. The same meaning appears in method/claim packets and the monitor. Availability is not qualification.

  • Recovery briefs bound long plan text and preserve current experiment/method/review references. Recent evidence is prioritized ahead of old plans.

  • Methods/progress reviewers can request up to three JSON excerpts per round, for at most two retrieval rounds. The host verifies declared outputs and hashes, limits size, and records responses. Missing data cannot grant approval. Claim reviewers continue to receive their existing full claim-evidence packets.

  • Builtin provider calls receive independent request receipts before step-level aggregation. Completion retries, failures, cache/reasoning fields when available, and prepared context sizes/hashes are recorded. No prompt contents, credentials, or private reasoning text are added by the usage logger.

  • The terminal and web views show reported token usage and provider retry wait. The monitor additionally separates worker, reviewer and memory usage, and shows unavailable details explicitly. Usage appears during a worker turn, not only after it.

Accounting limits

Request records describe client completion calls. The OpenAI-compatible SDK may internally retry transport once; those internal network attempts are not separately instrumented. A request whose response is lost can have unknown usage even if the provider billed it. Cached/reasoning totals are incomplete when any response omits those details. Historic step-only records remain readable and are not added again when request receipts exist for the same turn.

The display is not an invoice. Currency estimates, enforced spending/token caps, session persistence, and enforced spending/token caps remain follow-up work. The later model comparison is a pilot, not a completion-rate estimate. This release does not change the worker’s 24-step session policy or silently apply repairs to old scientific results.

Regression scenarios

  • Actual sanitized 402 Insufficient Balance envelope pauses without reconnecting or changing the deadline; ordinary transient errors still retry.

  • FLASH alone satisfies the existing FLASH-or-WarpX identity requirement.

  • Twenty recorded experiments plus verbose plans retain current evidence and method information; even a 2 KiB brief retains the latest receipt references.

  • A reviewer fetches a previously omitted residual and retains a negative methods verdict. Mutated, undeclared and oversized artifacts cannot pass as verified data.

  • Multiple completion-check calls are counted separately; duplicate stream receipts and legacy step events do not double count them. Interrupted requests remain visible.

  • A nested reviewer request updates live totals without overwriting another review.

  • Browser coverage checks credit recovery text, unknown cache/reasoning counters, request counts and alternative instrument labels without JavaScript errors.

Acceptance commands and final results are recorded with the local release artifacts under artifacts/release-0.5.3rc1/. Release publication remains a separate operation.

Validation results (2026-09-29)

  • Full Python suite: 764 passed, 7 skipped. Skips require optional solver installations or a local PRoot environment; no skipped case is claimed validated.

  • DSH adapter suite: 14 passed. Public schemas and three recorded demos verified.

  • Ruff, JavaScript syntax checks, and the Sphinx documentation build passed.

  • Planted and PIC benchmarks/campaigns, package verification, blinded search and the budgeted offline campaign passed.

  • Browser fixture passed with no JavaScript errors. An 80-column terminal check confirms request/token information is visible on a second status line.

  • Fresh-home installer check passed with no agent credentials or system changes.

  • Two real DeepSeek requests exercised the synthetic methods-review protocol: missing result requested, FLASH accepted as the alternative instrument family, then readiness rejected for a residual of 0.13 against a 0.02 threshold. Usage: 2,784 input, 717 output, 1,280 cached input tokens. This is a small protocol check, not a model comparison or plasma qualification campaign.

Follow-up: builtin adapter optimization

The nineteen-trial adapter benchmark selected native tool-message history for builtin DeepSeek workers. It reduced completion retries and token usage in independent confirmation while retaining the completion guard and existing session/summary policy. The initial validation above predates this follow-up; its additional regression results and receipts are recorded under artifacts/adapter-overhead-20260929/.

Follow-up: model pilot and task pack

The matched model pilot compares MiMo v2.6 Flash/Pro and DeepSeek V4.1 Flash on fresh FLASH tasks and CSV controls, reporting cache usage, interruptions and continuation/setup differences. One trial per main arm does not establish a model population ranking.

The preview also includes Simjecture Bench 0.1.0: portable recorded CSV/RZ inputs, numerical holdouts, independently reexecuted reducers, immutable-input checks, saved-result checks and quantitative plot-data validation. The sidebar’s Benchmarks entry prepares a conversation without starting a paid agent and exposes grading after delivery. Scientific prose/claim approval remains outside the finite coding score.

Both exports validate against Harbor 0.23.0, schema 1.3, with separate verifier configuration. Docker/Harbor end-to-end trials were not run on this host; native Bubblewrap oracle and negative controls, plus the real browser flow, passed.

Final source validation on 2026-09-30: 782 passed, 7 optional skips, 14 DSH adapter tests passed, strict Sphinx build passed. The benchmark controls include both recorded-data oracles, independent closed-form RZ/CSV checks, varied-energy holdouts, constant-output/no-op rejection, malformed saved results, mutated inputs and GUI preparation/grading without a model request. Installer and distribution checks are performed from a clean tracked-source snapshot before publication.

The clean wheel/source bundle contain the benchmark data and exclude the private operator preparation directories. Fresh-home installation/GUI checks passed without credentials or host package changes; a fresh wheel installation listed and prepared the CSV task. An additional dependency-preflight regression verifies that RZ grading reports a missing h5py extra as a setup error before attempting code execution.