Skip to content

Project notice

Release acceptance and reproducibility

Version 1.10.1 separates direct evidence from mocks, simulations and unavailable integrations. The release workflow (agent-release.yml) must complete its gates before publishing assets. Its run URL and tested commit are recorded in the published release manifest; JUnit reports, browser screenshots and integration evidence are retained as workflow artifacts and release evidence.

The Ascension implementation guide covers the new white-paper lab. Independent Node tests check exact accounting, encryption/tampering, policy gates and analytic dynamics. Real Chromium acceptance exercises all three scenarios, funding/redemption, Council settlement, native mission downloads, Nova-Seed recovery, policy changes, mobile layouts and offline reload. It also verifies the original PDF checksum. Both site builds and the public deployed URL must pass before publication.

The 1.2.0 release remains available unchanged. Patch 1.2.1 adds a required complete gallery rebuild: legacy documentation repair must retain the exact sandbox-host script hashes, every generated demo must load, and the rebuilt site must advance its simulation online and offline under normal browser security. Both minimal and full-asset gallery builds must pass. The generated site receives the complete browser distribution so documentation source exclusions cannot break its precache. Workflow-only repairs also retain distinct evidence from both Python matrix jobs.

Before main-branch packaging, a separate gate requires the historical integration and PR CI workflows to succeed on that exact main commit. A failure, cancellation, timeout or skipped workflow blocks publication. Their final run snapshots are included in the validation archive.

Full-repository and changed-file pre-commit hooks are required gates, including the locked browser ESLint environment. Their complete report, exit status and any proposed formatting are retained. The PostgreSQL ledger is exercised against a real temporary Docker database after TCP readiness; TypeScript meme mining uses the locked compiler. Both integration checks must run without skips.

Runtime and integration evidence

  • The runtime/control suite runs on every supported Python version, covering mission execution, signatures, corruption, review races, pause, idempotency, provider failure, coding opt-in, wallet control, confirmation/finality, invalid transfer receipts, exact integer accounting and replay rejection. RPC attack cases use fixtures.
  • Real local inference: pinned Qwen3 4B GGUF via llama.cpp, six exact quotes verified, operator review/export and backup/restore. This is actual inference, not the repository's legacy SDK stub. It is a public sample task, not a general intelligence benchmark.
  • Real local EVM: deployed ERC20 mock at the canonical address on an ephemeral Hardhat chain 31337; actual signed wallet proof, transaction, confirmation delay, receipt, integer balance, duplicate rejection after restart and verified backup/restore. No public-chain funds.
  • Solidity suite: 36 tests passed against shipped sources, including two new real identity rejection tests. The initial 34-test baseline used an always-true identity stub; it was not evidence of identity enforcement. The initial local Insight browser suite reported 6 passed, 1 skipped; hosted execution reported 5 passed, 1 skipped before the additional source-host test; TypeScript compilation passed; ESLint had zero errors and 9 existing warnings.
  • Initial broad Python regression: 893 passed, 12 failed, 100 skipped, 5 expected failures. Configuration precedence, loopback guard, compiler discovery and missing preview-asset causes were fixed; focused recheck passed 44 with only two notebook-kernel environment failures remaining.

Required hosted gates

The initial workspace kernel prevented Chromium and ZMQ notebook sockets. Those local failures were not passes. Hosted CI runs the full offline Python regression, real Chromium console interactions, the legacy web client tests, sandboxed Insight simulation online/offline from both npm and manual builds, the legacy container services, and real Docker isolation plus persistent operator-container restart. Each must succeed before publication. Optional test skips and expected failures remain visible in JUnit; they do not establish unavailable hardware, cloud credentials, optional integrations, live trading or deployed mainnet behavior.

The Insight cache-update check rebuilds after separate host-only and worker-only changes, requires new page precache revisions and matching CSP/worker integrity, then restores and reproduces the original build. The page revision covers its policy-complete template, the bundled worker template and all precached assets before filling the worker hash; hashing the final page and a worker that embeds that page's hash would create a circular dependency.

Both browser projects have dependency audit gates, with JSON reports retained in release evidence. The Insight audit includes its entire npm graph, including build tools and bundled browser libraries; its prior production-only audit omitted those development-labelled dependencies. The repaired dashboard audit covers production dependencies. Other legacy development environments can still have advisories; this is not a claim that every historical dependency is clear. The local full-graph audit records the exact lockfile hash, Node/npm versions and zero advisories; release CI independently reruns it.

The Python coverage gate retains the repository's existing scope: demos, legacy backend agents/memory and temporary test copies are excluded in both collection and reporting. The earlier expanded report measured 74.97%; applying the already configured scope measured 81.72%. The release retains its XML and raw coverage data so this distinction remains inspectable.

The runtime matrix covers Python 3.11, 3.12 and 3.13 using the hash-locked minimal environment. The larger historical suite runs in its original locked Python 3.12 development environment, without service credentials, external Python sockets or model downloads. Run locally with:

python scripts/check_python_deps.py
python -m scripts.run_local_tests --junitxml=regression.xml
python -m scripts.check_agent_preservation
python -m scripts.validate_agent_ui --output ui-evidence
python -m scripts.validate_agent_sandbox --output sandbox-evidence.json
python -m scripts.validate_agent_chain --output chain-evidence.json

Prefetch Insight browser assets with FETCH_ASSETS_SKIP_LLM=1 npm run build in its browser directory before the offline regression; its network guard intentionally blocks downloads during tests.

The last three require installed Playwright Chromium, Docker, and compiled Hardhat dependencies, respectively. They fail if their required capability is missing; they do not silently skip. The model validator additionally requires the independently downloaded server binary and pinned GGUF; its exact SHA and invocation are in the evidence JSON. The Qwen GGUF is downloaded independently; the full browser release ZIP includes the separately pinned quantized GPT-2 ONNX weights.

check_agent_preservation.py checks all 2,125 original paths, the entire original README text and flywheels (only CI badge queries and the exact live main-commit Health/Smoke/Integration image URLs may change), and equality of shipped Solidity sources with the copies compiled by the contract tests. Packaging verifies wheel metadata, installs outside the repository in a clean virtual environment, checks dependency consistency, and exercises the installed CLI and packaged web assets.

This evidence supports the bounded capabilities in CAPABILITIES.md. It is not a security audit, regulatory approval, investment-performance result or proof of AGI/ASI.

Additional 1.3.0 acceptance

The full-gallery job now requires real local ONNX text generation before and after a network-disabled reload, with matching deterministic continuations, the actual WASM backend and no page errors. The minimal-gallery job continues to require working offline simulation. Browser API-response fixtures are reported separately from actual ONNX inference. Model provenance is in scripts/browser_model_manifest.json.

The standalone Smoke Test matrix covers all nine OS/Python combinations; exact-main smoke success is included in the historical-CI publication gate. AIGA's direct-file startup test can no longer turn a service failure into a skip. README preservation permits CI badge query maintenance and the exact live main-commit Health/Smoke/Integration image URLs. It rejects static passing replacements, forced Health colors, unrelated checks or branches, and any removed flywheel or other historical text.

Version 1.4.0 additions

The release additionally requires the complete demo catalog, actual optional CPU research backends and lineage UIs, every canonical/mirrored legacy replay interaction, same-origin Python execution, preserved Insight Plotly/tree behavior, independent browser-algorithm tests and the complete Pages workspace. The latter includes edited inputs, native mission validation, review/export/import, explicit memory, tamper rejection, native Ed25519 exports, cancellation, mobile layouts and an offline root reload. Full-asset acceptance requires actual local model generation both online and after an offline reload.

The same tested site is retained as a checksummed release archive, deployed through GitHub Pages and verified at the public HTTPS URL before publication. The manifest binds source, version and tested commit. See demo validation and Pages operation and recovery.

Version 1.5.1 hardening

The operator Python dependency gate checks every package in requirements-agent.lock using pip-audit==2.10.1 and the current PyPI advisory feed. Missing packages, version mismatches, duplicate records, skipped checks, reported advisories and scanner failure block the release. The archive includes python-dependencies/pip-audit.json and a dated summary with the exact lock SHA-256. This scope does not include every preserved historical dependency environment or the host operating system.

The operator and CPU demo locks now use cryptography==50.0.1. Runtime acceptance rechecks signed identity, evidence, exports, corruption rejection and recovery on Python 3.11–3.13. Publication tests reject a stale main commit before any release mutation and again after uploading/re-downloading assets. A matching check immediately precedes Pages deployment. These checks reduce publication races; GitHub branch updates and deployment visibility are separate API operations, not one atomic transaction.

The CI watchdog now requires the intended SHA, rejects stale endpoint responses, and independently checks the repository run list when needed. Pending runs and unsuccessful reruns cannot report a pass when the bounded waiting period expires. All check failures remain visible.

See release readiness for supported operation, remaining research boundaries, maintenance and the clean-environment upgrade/recovery procedure.

Version 1.5.2 badge correction

Independent public verification after 1.5.1 found that the Health check-runs image aggregated several watchdog invocations attached to one commit. Earlier failures kept that image red after a newer watchdog run passed. Health now uses GitHub's latest workflow badge on main. The original check records remain available, and exact-commit acceptance still gates deployment and publication. The follow-up release repeats the full acceptance pipeline and public SVG verification.

Version 1.6.0 Insight Atlas acceptance

tests/browser/insight_engine.test.mjs covers exact integer allocations, bounded inputs, scheduling resource exclusivity, independent quorum votes, training-only selection, holdout checks, forged/stale proof rejection, reviewed Chronicle promotion, revocation, recovery and typed-graph integrity.

python -m scripts.validate_insight_atlas exercises all three expeditions, eighteen native research mission exports, actual evidence/recovery downloads, unsafe architectures, duplicate and stale promotion, malicious source URLs and text, four viewport widths, mirrors and offline reload/recovery/replay. Minimal/full builds and public HTTPS acceptance retain screenshots and insight-atlas.json in the validation archive.

The release finalizer refuses publication without every required public Atlas journey. Results establish local model behavior; economic hypotheses remain unverified. See the field guide.

Final-manuscript acceptance (1.8.0)

The manuscript gate checks the original 198-page PDF, canonical Markdown and all 31 referenced figures against immutable SHA-256 commitments and the pinned upstream revision. Release assets contain the paper, source manifest and complete manuscript archive. The implementation map separates local engineering results from claims requiring independent, stressed or delayed real-world evidence.

Compounding Lab acceptance covers positive transfer, changed-regime failure, no-archive ablation, full prior-learning and validation-call accounting, human review, imported timing provenance, fresh re-review, strict JSON/ZIP replay, unknown-field and prediction tampering, mobile 320/390 px, keyboard access, WCAG A/AA automated checks and real offline recovery. Native Python and browser JavaScript implementations must agree exactly. Both gallery configurations and the final public HTTPS site must pass.

The native request-boundary tests send independently specified ASGI headers and body chunks. They verify that oversized, falsely declared, duplicate-framed, incomplete and timed-out uploads leave no mission or journal mutation. Valid fragmented JSON still creates the requested mission. Authentication, hostile-host and early rejection responses retain the security headers. These tests complement actual container and browser integration, rather than claiming a formal penetration test.

Use the matching release manifest and validation archive for final counts and skipped-test details. The previous candidate at 2afee418e1142becd4baa94a282b83b648ade359 passed 1,176 regression tests with 54 optional skips, 4 expected failures and 82.83% coverage before the additional request-boundary tests. The final release reruns the complete matrix on its own commit; a prior candidate does not authorize it.

Version 1.8.1 accessibility correction

Changing a Compounding Lab scenario or its costs now clears both the visible forecast and its accessible description. Browser acceptance asserts that no prior prediction survives either reset. The same minimal/full-gallery and public HTTPS checks remain required before publication.

Version 1.9.0 decision acceptance

Independent JavaScript tests exercise portfolio/procurement oracles, energy accounting, resource conflicts, held-out isolation, evidence holds, CSV handling and replay tampering. The browser validator calculates all eleven cases, downloads their dossiers and verifies each through the standalone CLI. It checks real editing, imports, infeasibility, save/restore/clear, four export types, mobile layout, automated accessibility and offline recalculation. These checks run on both build profiles and the exact deployed public site before publication.

Version 1.10.0 operational audit

Additional oracles check deadline-feasible schedules against faster late alternatives, productive-time staffing, non-transferable shift capacity, backlog conservation, coverage caps, numeric evidence boundaries, repeated prerequisites and UTF-8 input bounds. Browser acceptance checks staffing failure, keyboard focus and archived dossier replay. The finalizer rejects public Studio evidence with missing journeys, cases, browser failures or any mismatch in origin, source commit, release version or calculation version.

Version 1.10.1 operator boundary audit

tests/runtime/test_execution_boundaries.py uses two real engine instances and an expired lease to check that late success and late failure cannot alter a replacement attempt. A loopback HTTP server exercises a stalled body, a body that continually sends fragments, successful JSON completion and connection closure at the configured deadline. These are transport and lifecycle tests, not a claim of model quality; the separate real-model gate still performs generation, review and signed export.

The same suite checks malformed authorization, refusal of a backup beyond the restore size limit, cleanup after a partial archive write and preservation of an existing destination. Standard recovery tests continue to verify identity, signatures and journal contents after restore. The patch preserves the Decision Studio calculation version and requires all existing public browser acceptance checks.

Release acceptance also runs FULL_INSTALL=1 FETCH_BROWSER_ASSETS=0 bash codex/setup.sh in a fresh virtual environment and requires pip check. This complements the separate hashed operator installer and avoids relying on packages retained from previous developer sessions. Source version, package version and catalog release must agree even when no installed distribution metadata is available.

Wheel verification now uses real Ed25519 signatures in both the loader and CLI tests. An unchanged allowlist entry cannot authorize changed wheel bytes, another public key, an invalid signature or malformed base64. The four CLI tests execute without the previous OpenSSL-related skips. This verifies the signature boundary; it does not qualify the historical hot-load integration for production use.