Simjecture 0.6.2rc1 — Projects and local workspaces

Simjecture 0.6.2rc1 adds optional projects and local workspaces for sustained research across conversations.

• Projects: group conversations, share instructions and reference files, and preserve the selected context for each agent turn and autonomous study.
• Workspaces: switch between local lab environments with separate conversations, project collections, saved model connections and machine registrations.
• Agent and CLI access: manage the same records through JSON CLI commands and explicitly scoped HTTP APIs. A browser is optional.
• Navigation: project pages collect related conversations; Resources groups tools, machines, connections and benchmarks. Existing users still open Personal without a setup flow.

Existing conversation folders, scientific records and study ownership links are retained. Local workspaces organise trusted research; they are not an operating-system security boundary. The hosted trial keeps its existing account isolation and single-workspace interface. Private operator assets are not included in this release candidate.

Install this prerelease explicitly with:

python -m pip install --upgrade 'simjecture[workspace]==0.6.2rc1'

The versioned one-command installer, source distribution, wheel, workspace bundle and SHA-256 checksums are attached. The stable release remains v0.6.1.

Validation: The release pipeline passed 1,339 tests with 7 skips, DSH, lint and fresh installation. Merged-source CI also passed documentation, numerical benchmarks, recorded demos and actual CPU WarpX qualification. The public PyPI wheel was downloaded, checked against release SHA-256 sums, installed in a fresh environment and exercised through the workspace/project CLI.

A local simulated-provider test timed out once during preparation; all seven session tests passed on unmodified recheck, and the independent full CI and release runs passed. The attached VALIDATION.json retains these observations.

Project/workspace guide · Implementation PR

Simjecture 0.6.1 — Research completion and visual plasma studies

v0.6.1 improves how bounded investigations finish and makes their numerical evidence easier to inspect.

• Protected reporting time: new minimal studies reserve time for a current report and separate report review. Completed reports and accepted scientific claims remain distinct; findings are falsified, supported within the tested scope, or unresolved.
• Research coordination: directors can wake workers for useful analysis while simulations continue. Ancestor reviews precede repair reviews, and report handling preserves admitted analysis jobs and uses remaining review time after a transport timeout.
• Visual plasma studies: follow real FLASH field evolution, the FLASH-to-WarpX patch, spatial pressure maps and evidence maps. Figures are available as PNG, PDF and SVG, with source records and reproducible plotting bundles.
• Modern documentation: refreshed README, research-loop and evidence guides, contributor entry points, current setup instructions and a separate historical archive.
• Tools and runtime fixes: Cylinder flow / Warp-LBM contributed by Zifei Meng, reviewed FLASH-to-WarpX transfer helpers, CUDA runtime isolation and identity checks, and clearer FLASH/WarpX/EOS/opacity guidance.
• Installation and interface fixes: preserve required validation data through source-to-wheel builds, keep large research archives outside the Python package, tolerate incomplete records during GUI discovery, and render the shared chat layout correctly from the first load.

The implementation passed 1,324 tests with 7 skips, actual CPU WarpX qualification, documentation/browser checks and clean packaging/installation checks. The publisher repeats the release checks before PyPI publication.

The preserved plasma investigations remain scientifically unresolved. Their report-review timeouts and subsequent harness fixes are documented; the new figures do not change those outcomes or imply an accepted claim.

Documentation · FLASH walkthrough · FLASH → WarpX walkthrough · Release acceptance

Upgrade:

python -m pip install --upgrade simjecture

Install the local workspace:

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

Simjecture 0.6.0 — Hosted research and a shared workspace

Simjecture 0.6.0 adds an optional hosted research service and a shared local/online workspace. Try it at https://simjecture.com.

The workspace has a persistent dark navigation rail, bottom monitor/account controls, account and workspace popups, a compact expanding chat composer, accessible inspector tabs, and peer navigation for Conversation, Study and Evidence & review. Concrete starter tasks cover double-pendulum integration, FLASH magnetic-island coalescence and an autonomous pendulum-period hypothesis study. Interactive and Autonomous research are explained on the home page.

Hosted users can sign in with GitHub, upload scientific inputs, retain their work and delete their own conversations. Guests have temporary per-tab leases and cleanup. Deletion stops active writers before erasing files and survives service restarts; usage counters remain. The official benchmark leaderboard, rankings, cost/time charts and downloads are available online, with model testing and import disabled there.

A separately confined Linux executor enables general Python, custom solver inputs and private FLASH application builds. Scientific jobs use dedicated UIDs, Landlock, seccomp and resource limits; the gateway uses a private Unix HTTP socket. This execution profile shares networking. Binaries, inputs, source and execution provenance remain recorded.

Analysis, plots, compilation and inspection now appear in Commands; numerical solver/integration runs appear in Simulations through an explicit execution category. Output files no longer determine the category, and hosted commands show their own retained outputs. Tool selectors offer custom study preparation alongside starter presets.

Actual P40 checks cover FLASH 1D/2D/RZ, a private FLASH resistive-MHD build/run, WarpX CUDA 1D/2D/RZ/3D, ITER diagnostics, editable inputs, compilation/reuse and private-file denial. Live MiMo tasks produced a double-pendulum calculation and correctly separated oscillator integration from subsequent analysis. WarpX photon emission and absorption also ran through the confined launcher; the guide records their tested regime and missing aluminium tables. These software/physics-operator checks do not establish a full aluminium Z-pinch radiation-yield model.

This release also includes the preceding research director, adaptive-run recovery and simulation-file provenance repairs.

Install on Linux/WSL:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.6.0/install.sh | bash

Python package: pip install 'simjecture[workspace]==0.6.0'. Hosted operator: pip install 'simjecture[public]==0.6.0' and follow the deployment guide.

Details: https://github.com/tomzhu0225/simjecture/blob/main/CHANGELOG.md
Deployment: https://github.com/tomzhu0225/simjecture/blob/main/docs/how-to/public-trials.md
Validation: https://github.com/tomzhu0225/simjecture/blob/main/docs/testing/0.6.0-acceptance.md

Simjecture 0.5.4rc2 — research recovery and clear simulation outputs

This preview fixes the compact-context error that prematurely stopped a research campaign and makes selected simulation files easier to interpret.

• Research recovery: derived context, journal and routine report errors recover with visible diagnostics. Primary scientific-state validation and final-report publication remain required. Early stop causes and unused budget remain visible after expiry.
• Simulation files: generated run outputs appear separately from copied or modified inputs. Command files and cross-run reports have shared-file labels. New runs record input hashes; older snapshots retain their provenance limits.
• Monitoring and review: expose role-level cached/uncached usage and submission-to-result latency; defer unchanged healthy-job strategy reviews within a 15-minute bound while retaining shorter operational checks and evidence-triggered reviews.
• Documentation: refresh the README, installer links, citation metadata and FLASH guidance, with a detailed campaign audit and rc2 acceptance record.

Install on Linux or WSL:

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

Python-package users can install simjecture==0.5.4rc2 after the publishing workflow completes. 0.5.3 remains the latest stable release. Existing research files and application versions are preserved.

Validation: 1,196 Python tests passed with seven optional-runtime skips, including recovery and simulation-categorization browser tests. All 14 DSH/MCP checks, lint, JavaScript syntax, public schemas, strict documentation, recorded CLI demos, package/archive checks and a fresh installation without model credentials passed. GitHub CI also verifies the full source and real WarpX qualification before publication.

The linked recovery continuation reached its recorded deadline with no supervisor/director error. It recorded 27 successful diagnostic/pilot jobs and 14 failed attempts; the reconnection contribution remains unresolved. This is a harness repair, not a scientific completion or lower-cost claim.

See PR #11, the rc2 acceptance record, campaign audit and changelog.

0.5.4rc1: Research director and execution repairs

Long-running numerical experiments now receive bounded strategy reviews during waits. The research director can stop named experiments, retain available partial results and require a recorded replacement plan or reasoned challenge. New minimal studies enable it by default, with visible GUI settings, decisions and confirmations; independent scientific methods and claim review remain separate.

This preview also includes:

• Live JSON/FLASH timing through local and SSH workers, measured execution costs and receipt-backed advisory progress.
• Reserved-slot MPI helpers, explicit trusted native-process workers, OverlayFS source manifests and frozen exploratory analysis.
• Durable replans across later continue reviews, recorded stop-control failures, reviewer recovery and discovery of externally launched studies without restarting the GUI.
• Explicit read-only inspection for minimal research reviewers while preserving tool-free review in structured/frontier workflows.

Install the preview on Linux or WSL:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.4rc1/install.sh | bash

Python-package users can install simjecture==0.5.4rc1 after the publishing workflow completes. 0.5.3 remains the latest stable release. Existing study policy, hypotheses, deadlines, research files and older application versions are preserved.

Validation: 1,167 Python tests passed with seven optional-runtime skips, including director/recovery and desktop/mobile GUI cases. All 14 DSH/MCP tests, strict documentation, archive checks and a fresh installation without model credentials passed. Full source and real WarpX CI checks pass before publication. See PR #10, the candidate acceptance record and changelog.

The ongoing Al campaign has completed two exploratory 40 ns 3D trajectories and exercised director stop/replan control. Scientific attribution remains unresolved; director token overhead remains substantial. This preview makes no scientific completion or lower-cost claim.

Simjecture 0.5.3 — stable workspace, SSH experiments and benchmarks

Stable release of the 0.5.3 candidate series. Minimal remains the default;
structured and frontier remain selectable. Existing research data and older
program folders are preserved by the versioned installer.

• Run recorded numerical experiments on local or SSH workers with CPU/RAM/GPU
  reservations, automatic machine setup, 30-second availability checks and recovery
  after interruptions. Research agents and independent review stay on the coordinator.
• Publish the measured coding-agent leaderboard with task-specific time/cost
  rankings, failure-inclusive cost bars, consistent model colours and a Pareto plot.
  Custom trials and community grades remain separate; the benchmark analyzes
  recorded diagnostics and does not launch new simulations.
• Repair concurrent ledger writes, idempotent replay, reserved-CPU time limits,
  transport recovery, input validation and conversation races. Retain request-level
  usage records, native DeepSeek tool-message continuity and review evidence retrieval.
  Preserve scientific review gates, deadlines and the benchmark task pack 0.3.0.
• Connect conversation, selected study, and Evidence & review with shared
  navigation. Study links retain their selection across reloads, mode changes,
  and monitor Back/Forward navigation.
• Return from study details to the owning conversation, and prepare continuations
  there by default. Standalone CLI studies and read-only inspection remain available.
• Guard study rendering and continuation dialogs against stale requests, newer
  navigation, and repeated submission. Clarify that simulation output alone is
  not an independently accepted scientific finding.
• Release pending continuation previews when navigating away, so a stalled old
  request cannot block a new action. Ignore its late reply even after returning
  to the original route; retain the currently open dialog.
• Use shared light/dark logo artwork in the workspace and monitor, with SVG/ICO
  favicons, accessible home links and packaged font attribution. Keep assets
  self-contained and preserve the existing content-security policy.
• Make the README and documentation entry points workspace-first, move terminal
  automation into a dedicated guide, and group dated reports in an archive index.
  Simplify contributor setup and checks, and keep release facts in the changelog
  and acceptance records.

Release checks and limitations.

Archived version DOI.

Simjecture 0.5.3rc4 — audited harness and interface repairs

Simjecture 0.5.3rc4 integrates the independent rc3 audit, with confirmed repairs
to evidence integrity and numerical execution, worker and accounting improvements,
and modest interface polish. Minimal remains the default.

• Concurrent SQLite writers now preserve the event hash chain. Canonical JSON
  prevents distinct replay payloads such as true and 1 from being accepted
  as identical. Indexed campaign heads and replay reduce query overhead.
• Reserved threaded experiments receive CPU-time allowances scaled by their
  assigned CPUs. A real eight-thread computation that rc3 killed early completed
  within the same wall allowance after the repair. This is not CPU affinity or
  an aggregate MPI quota.
• Reduce targeted worker-poll overhead, retain full reconciliation for admission,
  verify completed replay uploads before skipping remaining chunks, and preserve
  uncertainty after malformed transport replies.
• Stream provider usage fields with much lower memory overhead; improve transport
  errors, validate finite/nested inputs before persistence, and avoid extreme-range
  interpolation overflow.
• Fix conversation navigation/submission races, draft preservation, duplicate
  benchmark submissions, button states and connection-save errors. Add shared
  theme styling, keyboard skip links, accessible labels and bounded dialogs.
• Reconcile documentation and qualify historical reconnection conclusions against
  their registered tolerance band. Preserve original records. The authorship
  section remains explicitly a discussion draft.
• Keep installed dependencies and local state out of source distributions,
  including builds from Git worktrees.

Validation: 974 Python tests passed with 7 optional-runtime skips; 14 DSH/MCP
tests passed. Lint, schemas, strict documentation, clean distributions and a
fresh-home installer passed. CI also passed the bounded WarpX CPU qualification.
Independent browser review covered 20 desktop/mobile light/dark views.

These fixes do not change the scientific search strategy, prompts, completion
rules, review gates, task pack 0.3.0 or published model measurements. Synthetic
overhead measurements do not establish model token savings or improved long
research outcomes. No new long scientific campaign or paid-model trial is claimed.

Install this explicit preview:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.3rc4/install.sh | bash

Or install the Python package:

uv tool install --upgrade 'simjecture[workspace,tui]==0.5.3rc4'

This is a prerelease; stable/latest remains 0.5.2.
Scope and validation,
original audit log.

Simjecture 0.5.3rc3 — measured AI benchmarks

Simjecture 0.5.3rc3 publishes the measured coding-agent leaderboard and improves its public presentation. The benchmark evaluates analysis of recorded plasma diagnostics; it does not launch new FLASH or WarpX simulations or establish general research ability.

• Two tasks, with actual results from 42 scheduled configurations and 252 scheduled attempts. Only configurations that reached inference appear in public rankings.
• Finish-time bars, cost bars and a cost–time Pareto plot with distinct model colours and a model legend. Dashed lines connect only efforts of the same model and coding agent; frontier points are individually outlined.
• Unfinished runs appear in red and keep their recorded costs or labelled lower bounds. Remove old pilot/unranked panels and the completion/time plot; methodology is a documentation link.
• Run custom models with installed CLI agents or configured API connections. Current community/local submissions remain separate from owned measurements; full audit records are retained.

Validation passed: 868 Python tests (7 optional-runtime skips), 14 DSH/MCP tests, source lint, strict documentation, browser checks, WarpX CI qualification, and a fresh-home installer test without an API key or agent. Both the main CI and PyPI publishing workflows completed successfully.

Install this explicit preview:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.3rc3/install.sh | bash

Or install from PyPI after the publishing workflow completes:

uv tool install --upgrade 'simjecture[workspace,tui]==0.5.3rc3'

This is a prerelease. Stable/latest remains 0.5.2. Benchmark results and limitations, preview scope.

Simjecture v0.5.3rc2 — SSH workers and refreshed research interfaces

Simjecture 0.5.3rc2 adds local and SSH experiment workers for minimal studies and refreshes the Machines and Research tools interfaces. One coordinator retains the research agent, hypotheses, evidence and independent review; headless workers execute experiments in parallel. Simote remains optional.

• Connect a worker with an SSH address and password. Detect hardware, persistent paths, non-root accounts and execution environments automatically; use advanced settings or agent-guided preparation for custom solver setup.
• Check availability every 30 seconds, show placement and jobs, reserve CPU/RAM/GPU capacity, recover existing jobs after connection loss or coordinator restart, and confirm process-tree cancellation.
• Inspect registered instrument source through read-only helpers and retrieve recorded artifacts with SHA256 verification. Preserve machine selections in continuation drafts.
• Use compact Research tools cards, with build lists, paths, diagnostics and logs in a separate Details panel. Detected local builds remain distinct from an unavailable optional managed runtime.
• Retain rc1's campaign audit repairs, request-level usage, native DeepSeek tool continuity and experimental Simjecture Bench 0.1.0.

Real-host checks used two P40 machines, concurrent numerical jobs, FLASH RZ/MPI/HDF5 and simultaneous WarpX jobs on distinct GPUs. Two-field automatic preparation passed on both hosts. These are infrastructure and bounded commissioning checks, not new scientific conclusions.

Local release validation: 823 Python tests passed, 7 optional-environment checks skipped; 14 DSH tests passed. Ruff, JavaScript syntax, public schemas, documentation, clean archives/checksums and fresh-home installation also passed. The publishing workflow repeats the release checks before uploading to PyPI.

Preview: stable/latest remains 0.5.2. Execution pools currently cover minimal-mode numerical studies; remote inference-agent hosting, Simote profile import and Slurm worker adapters remain separate. CPU counts are admission reservations, and cooperative PRoot is not kernel/network isolation.

Install this preview, preserving existing research data:

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.3rc2/install.sh | bash
# Python package only:
uv tool install 'simjecture[workspace,tui]==0.5.3rc2'

The publishing workflow attaches installer assets and uploads the Python package after validation.

SSH worker guide · Release scope and validation · Changelog

Simjecture v0.5.3rc1 — campaign repairs and scientific benchmarks

Simjecture 0.5.3rc1 repairs campaign recovery and usage visibility and introduces Simjecture Bench 0.1.0, an independently versioned experimental scientific coding task pack. Minimal remains the default; existing session policy, completion guards and independent scientific review remain in force.

• Preserve native DeepSeek tool-call/result history and private reasoning continuity, selected through controlled adapter trials.
• Pause exhausted-credit/quota failures with an actionable resume path and the original deadline.
• Interpret alternative instrument requirements consistently; retain recent evidence in bounded review contexts and support hash-verified reviewer retrieval.
• Record request-level tokens, cache/reasoning details when reported, retries and unknown usage; show live breakdowns in the terminal and browser.
• Add Benchmarks to the workspace: prepare a conversation with your chosen agent, then independently grade delivered results.
• Package portable CSV energy accounting and recorded RZ HDF5 analysis tasks, varied numerical holdouts, immutable input checks, saved-result and numerical plot-data validation. Passing is a finite coding result, not approval of a physical claim.
• Export Harbor schema 1.3 tasks with an unprivileged agent, declared time budget, separate verifier and oracle controls. A checksummed standalone Bench pack is attached.
• Publish the matched MiMo 2.6 Flash/Pro and DeepSeek V4.1 Flash pilot, including cache usage, interrupted requests and setup/continuation caveats. Single trials are not a leaderboard.

Install the preview

curl -fsSL https://github.com/tomzhu0225/simjecture/releases/download/v0.5.3rc1/install.sh | bash
# Python package:
uv tool install 'simjecture[workspace,tui]==0.5.3rc1'

The versioned installer preserves existing research-data folders. 0.5.2 remains the stable/latest release.

Validation and scope

Local full suite: 782 passed, 7 optional runtime skips; all 8 benchmark controls passed after the dependency-preflight addition; 14 DSH adapter tests passed. Clean-source Ruff, strict Sphinx, JavaScript syntax, distribution contents, fresh-home installation/GUI and fresh wheel CSV preparation passed. Both exported task configurations validate against Harbor 0.23.0.

The release host has no Docker runtime, so a full Docker/Harbor trial was not exercised locally. The exporter is experimental. Interactive benchmark preparation does not impose a chat deadline; controlled timed comparisons use the external runner. Recorded diagnostics include no FLASH binary/source or EOS/opacity tables. Current benchmark scoring does not certify scientific prose, causal explanations or research completion rates.

Benchmark guide · Model pilot · Acceptance record