Demo refresh audit — 2026-10-08

These records distinguish useful numerical work, accepted scientific conclusions, agent mistakes and harness defects. The original campaigns are preserved. Fixes made after a cutoff are not credited to the autonomous worker.

Observed outcomes

Investigation	Budget and worker	Result
Kepler orbit	Luna medium; completed in 399.5 seconds	Accepted root falsification and fresh, independently supported finite repair
FLASH island coalescence	One hour; Luna medium	Four completed prospective trajectories rejected by a faulty path diagnostic; no accepted claim
Same FLASH question	One hour; Sol medium	Five baseline runs, two complete refinements and recoverable full-window output from a third; final processing and scientific review remained incomplete
FLASH → WarpX	Two hours; Sol high	Actual state transfer and completed spatial, timestep, particle-count and boundary controls; no accepted claim

The three plasma runs used a separate Sol-high reviewer context and a research director. Budgets, models and concurrent machine use differ from the historical six-hour DeepSeek/classic campaign. These observations do not establish a general ranking of models or workflows.

Fixed defects and tested improvements

Ancestor review ordering. The first Kepler attempt exposed a host scheduling error: repair support was repeatedly reviewed before its parent’s falsification had been accepted. The corrected scheduler processes ancestors first and leaves dependent reviews queued without repeatedly spending provider calls. The fresh Kepler run completed the full loop after that correction.

Discovery isolation. A compact audit has some of the filenames of a live study, but may omit its immutable commissioning snapshot. Auto-discovery allowed that expected integrity error to prevent the web interface from opening another study. Discovery now skips that invalid candidate; explicit loads and resumes still reject it. CI first reported one failure with 1,299 passes; the corrected commit passed 1,303 tests with seven skips, plus real WarpX, packaging and docs checks. Further changes have their own later validation.

Director-to-worker coordination. During the coupled study, the director repeatedly requested review preparation while a solver ran. The host had parked the worker after its preceding turn, and a continue decision did not wake it. The instruction was therefore not an immediate parallel task. After both studies ended, the director schema gained wake_worker, its packet gained the waiting state, and the GUI gained an explicit wake-up-request message. Tests verify that the worker resumes without stopping the solver or creating a replan gate, while ordinary waiting still avoids unnecessary provider turns. This host behavior is tested; an improvement in real-model completion rate has not yet been measured.

Reusable coordinate quadrature. Luna’s diagnostic mixed a cell-centre gauge with cell-boundary integration. The reviewer flagged earlier alignment problems but approved a repair that still contained the half-cell error. A post-run coordinate-consistent calculation reduces its apparent 0.18–0.25 discrepancy to approximately \(10^{-7}\) on the same fields. An optional, narrowly scoped helper is now in the FLASH skill, with analytic controls for unequal spacings and shifted origins. The original faulty source and its results are unchanged.

CUDA runtime qualification. Operator commissioning exposed WSL GPU-path, isolated CuPy/header-discovery and nested native-library identity issues. The installer and registration fixes passed an actual device kernel, openPMD readback and a sandboxed kinetic trajectory before the two-hour clock started.

What remains

• Analysis runtime and instrument requirements need a clearer separation. The Sol FLASH campaign attempted campaign plotting in a runtime without Matplotlib. It recovered with SVG, but changing the bound source at the end left its processing unapproved. The current required-capability rule applies to every evidence method/run, including postprocessing; ordinary lab.analyze records exploration. A future design should allow an appropriate reviewed analysis runtime while preserving verified upstream solver inputs and the operator’s instrument requirement. That rule has not been weakened here.

• Local resource limits should be easier to discover and configure. The enlarged kinetic loader exceeded the executor’s 4 GiB resident-memory default, not the GPU’s device-memory capacity. The worker successfully batched loading. The limit itself worked; local allocation visibility and configuration remain an improvement opportunity. The SSH execution-pool resource interface already has a different allocation path.

• Validate each case before finishing the scan. Sol’s \(S_\eta=500\) trajectory reached the primary flux window, but stride-two sampling discarded the last needed crossing. This was discovered in the final combined analysis. The FLASH guidance now recommends prospectively qualified endpoint handling and immediate per-case checks. No retrospective eligibility change was applied to the record.

• Keep figures separate from frozen numerical measurements where practical. Cosmetic or dependency repairs should not unnecessarily change the analysis implementation that must be reviewed. Test the exact runtime and final report command during commissioning.

• Leave time for review in the actual critical path. The coupled worker launched a fresh run too late, then cancelled it after comparing its measured cost with the remaining budget. A written reserve is not a demonstrated completion plan. The wake-up correction addresses one coordination problem, but does not guarantee good model planning or a completed investigation.

• Read receipts alongside worker prose. Both longer runs left stale statuses in RESULTS.md. The host’s execution and verdict records are authoritative; the exported worker reports retain their original wording.

The FLASH operating guidance collects the reusable lessons. The diagnostic correction, editorial plots and this audit were produced by the operator after the corresponding runs; they are not autonomous recoveries or accepted new scientific evidence.

Thirty-minute continuation trials

Both studies received separate 1,800-second continuations from commit 4d76a14f443b42d06dfec8d6539df7d7b01feabd, with unchanged original hypotheses, requirements and model/effort routes. No interpretation of the exponent was supplied. Frozen parent manifest, experiment, method and claim-review hashes were checked after both runs: none changed. The original records remain intact.

Observation	Island continuation	FLASH → WarpX continuation
Recorded executions	12 succeeded; 1 diagnostic cancelled	5 succeeded
New numerical work	Five base cases and two matched refinements completed	Fine 576² / 32-ppc full-window control completed in 566.6 s
Native input / cached / output tokens	4,527,873 / 3,925,376 / 20,873	4,066,383 / 3,548,544 / 33,144
Scientific outcome	Unresolved; no accepted claim review	Unresolved; no accepted claim review
Report outcome	Narrative retained; review timed out; stale pending-analysis passage remains	Quantitative narrative retained; review timed out

The director wake-up change was exercised by both models: report and control preparation proceeded while numerical jobs continued. The coupled director explicitly deferred the fresh-seed run when its measured cost no longer fit the compute allowance. These are observed behaviors, not a general completion-rate or token-efficiency benchmark.

The protected report phase also exposed two concrete host defects. First, entering finalization cancelled an analysis job even though its admission permitted execution into the drafting window. The worker reran it successfully, but the report-drafting call ended before incorporating its terminal result. Second, report review received a fixed 135-second call allowance. Both reviews hit that limit and the host marked finalization finished despite remaining wall time. The original reports and failed review status are preserved; no verdict was inferred from reviewer commentary.

Subsequent fixes, after the trials ended: preserve admitted diagnostic work at the numerical cutoff, wait for its receipt within the draft allowance and resume the writer to incorporate it. Donate unused drafting time to review and allow one bounded remaining-budget retry after a review transport failure. Ask report reviewers to return clear blocking corrections promptly rather than repeating the entire methods audit. Sixty-three focused tests passed, including regressions for each observed transition. These later corrections have not yet been validated by another paid plasma campaign; the two trials cannot establish their real-model completion rate.

The new field/evidence figures are operator renderings from the preserved arrays. Their plotting bundles retain source hashes and selection metadata; they add no claim approvals. The full raw simulations, original compact records and new continuation records remain distinct.