Skip to content

$AGIALPHA Agent 1.4.0

This feature release makes the GitHub Pages entry point a useful workspace and repairs the complete demo catalog. It preserves all original paths, README material, flywheels and previous immutable releases.

Use the project from the page

  • Four editable browser workflows: exact bounded allocation, source-grounded research, job-priority scheduling and train/holdout forecast comparison. Results expose methods, checks, assumptions and limits.
  • Worker cancellation, mobile layouts, keyboard controls, reduced motion, reviewed report exports, optional device-local mission memory and native mission-file handoff to the installed agent.
  • Real pinned GPT-2 ONNX generation in a separate worker, with visible download/error states and offline reload acceptance. This is a small text-completion baseline, separate from the checked workflows.
  • Local verification of Ed25519 agent exports using an independently trusted public key. New exports carry canonical signed bytes to preserve Python floats and large timestamps in browser verification.
  • Searchable 26-entry catalog, current guides, operator installation commands, recovery instructions and explicit $AGIALPHA/identity/payment setup. GitHub Pages never requests wallet secrets or an agent token.

Correctness and preservation

Meta-Agentic demos retain lineage across repeated runs, preserve RNG independence and use the correct Pareto objective directions. v3's explicit offline provider and curriculum actually evaluate the identity fixture; other generated programs require the existing isolated Docker evaluator. All three lineage UIs read their actual database schema and export the selected range. The maintained Streamlit launcher works.

MuZero respects the step limit without rendering and closes its environment. AIGA's multilayer Hebbian updates work for individual and batched observations; unexpected errors no longer pretend a stub succeeded. OMNI has a working package entry point and propagates episode failures. Macro Sentinel reports lower-tail stress correctly. Marketplace has an explicit local dry run. GPT-2 uses a packaged checksum-verifying, atomic downloader, supports explicit offline models and fails visibly.

Legacy charts load the correct pinned Chart.js browser build. Their shared replay controls distinguish bundled samples, actual Python examples and optional paid provider-generated synthetic data. Python uses a complete checksum-pinned same-origin Pyodide runtime. Provider modes require an explicit model ID and report errors without silently replacing failed output with samples. Original Insight charts, tree and scripted logs remain a distinct preserved presentation; its D3 library, replay and tree bindings work.

Validation and distribution

Required gates retain the existing runtime, real local model, isolated code, local EVM, original contracts, full Python regression, platform smoke, types and quality checks. New gates cover catalog launches, optional CPU research backends and lineage UIs, independent browser-algorithm oracles, all replay pages, actual Python execution, mobile/offline workspace use and browser signature verification.

The full tested site is packaged separately from the Insight browser ZIP and deployed through GitHub Pages. Public acceptance verifies that tested version before release publication. Release assets include source, wheel, source distribution, browser ZIP, site archive, guides, validation evidence, manifest and SHA-256 checksums. Published 1.3.0 and earlier release assets are unchanged.

The manual 📚 Docs entry point uses that same guarded release workflow, repairing the legacy literal Docker-tag failure and avoiding competing branch-based deployments. Pages requires Source: GitHub Actions; no additional template or admin token is needed after the one-time repository setting.

Upgrade and boundaries

Follow OPERATIONS.md to pause, verify and back up your existing private agent state. Install 1.4.0 into a new environment and verify the existing state before resuming. There is no journal migration. Earlier exports still verify with the CLI; re-export them for the browser's canonical-byte verifier.

Optional CPU demo dependencies have their own hash-locked environment (requirements-demo-verified.lock), validated on Linux/Python 3.12. Keep it separate from the lightweight agent installation. Historical cloud, finance and sovereign deployment templates remain preserved research material with documented limitations. No mainnet transaction, treasury spending, formal security audit, general intelligence or profit guarantee is claimed. See CAPABILITIES.md, DEMO_VALIDATION.md and PAGES_GUIDE.md for scope and recovery.