0.5.4rc2 scope, checks and limits

This preview repairs the premature supervisor exit found in the adaptive aluminium campaign and separates simulation outputs from copied project inputs. It includes the rc1 execution controls. Minimal remains the default, with independent methods and claim review.

Research recovery

  • Compact context eviction accounts for every removable section, including progress and execution costs. The previously failing 6000-byte context is covered by regression tests and replay against the original remote records.

  • Derived context, journal, routine report and navigation failures produce visible diagnostics and a recovery context. Primary scientific-state errors still fail validation, and the final report must persist before completion is published. Successful recovery clears stale warnings.

  • Status retains an early stop cause and the unused budget after deadline expiry. The monitor reports cached/uncached usage by role and submission-to-result latency separately from execution time.

  • New studies can defer strategy reviews of healthy, unchanged work to at most 15 minutes. Operational checks continue during deferral; changed evidence, pending replans, transport problems, stale telemetry or deadline risk restore the shorter cadence. Existing saved policies retain their behavior.

  • FLASH guidance covers build-specific radiation units, common comparison windows, native energy crosschecks, NumPy JSON conversion and preservation of partial raw data after analysis failure.

See the campaign audit for the original failure, measured overhead and the repair’s scientific limits.

Simulation file categories

The interactive monitor displays Run outputs first. New runs record input hashes, sizes and timestamps, so unchanged copied inputs and modified inputs can be distinguished from generated outputs. Copied inputs remain available in a collapsed section. Command files and cross-run reports have shared-file labels; project reports remain accessible through Project files.

The input manifest remains on disk without being repeated in polling and agent context. Output indexing prioritizes generated files under its display limit and tolerates files disappearing during a live scan. Older snapshots retain their files and disclose their limited input provenance; changed inputs cannot be reconstructed retrospectively when no original hash was recorded.

Verification

Release checks cover the full Python suite, desktop/mobile browser regressions, the DSH adapter and real MCP profile, lint, JavaScript syntax, public schemas, strict documentation, packaged source integrity and a fresh installation without agent credentials. Exact results are recorded in the release notes.

The repaired private generation also ran a linked continuation on the original research coordinator with numerical jobs on its SSH worker. The continuation used only the 4 h 14 min of budget lost at the parent’s early exit. It reached its recorded deadline on 2026-10-04 at 13:18 UTC+8, reporting budget_exhausted, no supervisor/director error and no active experiments at the subsequent check. It recorded 27 successful diagnostic/pilot jobs and 14 failed attempts. This is an observed recovery on one campaign, not a guarantee against other faults. Original parent records were preserved.

The simulation categorization regression uses two runs with a copied prior output and shared project report. Backend tests also cover modified inputs, limited output indexes and legacy snapshots. The local neon-tube demonstration was inspected through the running GUI; its selected case showed 16 run outputs separately from 39 copied inputs.

Limits

  • The aluminium continuation has no accepted methods or claim. It mostly audited retained data and tested instrumentation; it produced no new complete 40 ns trajectory. Boundary-energy accounting, spatial qualification and reconnection attribution remain unresolved. Reaching the wall deadline is not scientific completion.

  • Native usage counters include cache reads but do not provide invoices or complete per-request counts. The changed review cadence is not a demonstrated reduction in spending or an improvement in scientific judgment.

  • File categories identify provenance, not numerical validity. Shared reports can combine multiple runs, and older input manifests have explicit limits.

  • Existing studies, hypotheses, deadlines and scientific acceptance gates are preserved. Cooperative process execution remains a trusted-host option rather than an OS security boundary. This release does not qualify a new solver, complete a WarpX patch investigation or reproduce an experimental result.

Install this preview on Linux or WSL:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.4rc2/install.sh | bash

Stable 0.5.3 remains available. The versioned installer preserves research files and earlier application folders.