Skip to content

Project notice

1.23.0 — Complete demo catalog and repeatable evidence

This release makes every catalog entry discoverable and gives repeat experiments a safe, single-command path. It includes the Sovereign session cleanup prepared for 1.22.1. All original demos, research narratives, presentations, media, manuscript bytes and flowcharts are retained.

Start here

Follow the installation guide, then run:

alpha-factory demo list
alpha-factory demo info finance_alpha
alpha-factory demo check finance_alpha
alpha-factory demo run finance_alpha --output-dir ./demo-runs --new-run

--new-run creates and prints a unique directory under your selected output directory. Repeating the command retains earlier evidence. Leave it off when you intend to reuse persistent history. The walkthrough describes the output and next steps for each supported path; the complete inventory distinguishes 17 finite offline demos from nine entries with separate acceptance or prerequisites.

Improvements

  • Generate both complete inventory tables and current launch headings from the executable catalog, and reject drift in CI. All 26 entries now show their current modes and requirements.
  • Link Governance, MuZero Planning, MuZero × MCTS × LLM and Insight v1 to their own maintained experiences. Preserve Decision Studio and its eleven cases.
  • Serve actual image bytes in generated guides while preserving linked badges, source links, code examples and Mermaid diagrams.
  • Retain validation results for every finite command, including timeouts and missing prerequisites; save the report before returning failure. Inventory-only checks do not claim execution.
  • Check all 17 finite demos both normally and with --new-run from the installed wheel outside the repository, with Python network calls blocked and prior output bytes compared afterward.
  • Clear Sovereign drafts and review notes when locking, retain saved mandates, and isolate delayed unlocks and file reads from later sessions. See the session cleanup notes.
  • Cancel obsolete release runs that cannot pass the current-main guard, preventing them from blocking newer releases. Deployment serialization and every acceptance gate remain in place.

Validation and delivery

The dated audit records the local evidence and its limits. The release workflow gates publication on the exact committed source: dependency advisory audit, full Python regression, Python 3.11–3.13 runtime/lab matrices, native training, installed-wheel demos, real Chromium journeys, Docker isolation, manuscript/preservation checks and public HTTPS acceptance. It packages the tested site without rebuilding it, binds receipts to the commit, and re-downloads uploaded draft assets to verify their checksums before publication.

Keep release-manifest.json, SHA256SUMS and the versioned validation archive with the source, wheel and operator guide. Inspect skipped and expected-failure tests in their JUnit reports; they are not passing qualifications. The publication workflow's final result is the evidence for this release, not results copied from an earlier commit.

Breaking changes and upgrade

No removed demos, journal migration, signed-packet change or mandatory output-path change. --new-run is additive. Machine consumers should recognize agialpha.demo.validation.v2, which adds overall status and explicit separate-acceptance coverage. Consumers of the earlier report schema should update before parsing new reports.

Pause the old agent, verify and retain a private backup, install matching 1.23.0 assets in a fresh environment, and restore into a new private home. Verify identity and journal head before resuming. Keep the previous environment and backup for rollback; never run two versions against one home. See release readiness for exact scope and upgrade procedures.

The supported scope remains bounded demos and private single-operator operation. Optional external integrations, independent validation, long-running production operation, mainnet deployment and paper-level AGI/ASI claims retain their separate evidence requirements in the manuscript map. Passing finite examples does not establish those claims.