Release readiness — 1.36.0 source candidate
Version 1.36.0 adds bounded transformational discovery and the shared inert TI project. The release proposal and implementation register separate installed mechanisms from research and qualification gaps. The public-source pilot has three installed operators; separate local-model, mathematical and complete-campaign studies retain their own evidence scopes. An offline TI workspace, original native example and complete guides are included in the package proposal.
The reviewed 1.35.1 Computer Use reader and history fixes remain in the source ancestry. A fresh exact-main acceptance run must pass all existing gates plus TI local/package/public readback before delivery. The current source label, local test reports and historical passing runs do not identify a published 1.36.0 release. No model-weight advantage, calibrated outcome gain, general seven-axis capability, paid lifecycle or compatible four-repository release set is established. Original paper bytes, URLs, historical release records and all commissioning and authority boundaries remain unchanged.
Previous release scope — 1.35.1
Version 1.35.1 repairs the public-document reader's cancellation and wall-clock boundary. The maintenance notes describe disposable DNS/HTTP supervision, conservative interrupted-read accounting and fresh source-bound permits. Existing source, regression, Docker, package, browser, documentation, public HTTPS and uploaded-byte gates remain mandatory. A matching published manifest and exact-commit workflow establish delivery; source preparation alone does not. Historical releases, deployment records, on-chain authority and separate commissioning gates remain unchanged.
Previous release scope — 1.35.0
Version 1.35.0 adds the Computer Use flagship journey and project-bound native returns described in the release notes. Level 3 develops and compares strategies; Level 4 tests a different problem formulation while preserving the original outcome and acceptance criteria. The primary mission develops a source-grounded improvement proposal for the released public Computer Use runtime, with a reproducible validation plan. Research and mathematics lanes request substantive candidate work. Beneficiaries and expected value are explicit unverified hypotheses. Prepared attempts, historical results, owner reviews and existing catalog work remain in one portable project history. These additions do not modify published v1.34.0 evidence.
Browser preparation performs no model call. Execution requires a configured native model, an immutable container image and the exact reviewed task/provider/image/worker/context permit. The static page has no remote computer-control service. A retained result or owner's review confers no signer trust, independent qualification, operational authority or settlement approval.
Release publication requires all existing exact-integrated-source, full regression, CI, packaging, generated-identity, browser/recovery/offline/accessibility, public HTTPS and uploaded-byte gates. The complete flagship gate must exercise mission preparation, native-return boundaries, strategy/reframing preservation, review, combined archive recovery and English/French layouts. The default must use real pinned public source, preserve the original useful outcome through alternative strategies and avoid invented benchmark, novelty, proof-checking or economic claims. Historical passing checks do not qualify a later source. The actual matching workflow, manifest and checksums establish acceptance and delivery; this document does not assert a hosted full-suite or public-site pass.
Native Docker and live-provider qualification of the added flagship path remain separate operational gaps. Labeled model/actuator doubles and scripted browser checks establish their tested software contracts, not a completed task by a live model or production container isolation. An authorized provider/container trial must retain exact source/model/image identities, screenshots, usage and cost evidence, and review of the actual work product. Patch execution, performance measurements and formal proof checking require separately connected and qualified tools; browser inspection and a visible proposal do not establish those results. Unknown costs and interrupted effects remain unknown. Full desktop work, private profiles, credentials, payments and messages remain outside this adapter's supported scope.
The proposed $40 trillion/year market is the owner's estimate. No universal computer-work capability, independently protected superiority, realized economic benefit or paid settlement is established by this release scope. Mainnet records, pinned historical wallet profiles and separate USDC/$AGIALPHA routes retain their existing commissioning and authorization boundaries.
Previous release scope — 1.34.0
Version 1.34.0 combines the useful Alpha flagship with the 18 October Mainnet deployment record. The release notes describe the combined source and upgrade boundaries. A prepared version, a passing feature branch and a published website are separate facts. The matching release manifest, successful exact-commit acceptance and public served bytes identify what has actually been delivered.
The flagship produces joined catalogs, unmatched rows, discrepancies and reusable local tools. Its optional Gym constructs bounded strategies, retains losses and tests a frozen plan on a new public synthetic case. Discrepancy-action reframing and checked native workbench returns remain bound to the exact project. These are local engineering results. Protected Specialist evaluation, independent review, realized economic advantage and authorization remain separate obligations. Supplied economic estimates retain their assumptions; they are not measured savings.
The five current Mainnet records retain finalized creation/binding observations, source provenance and scoped graph verification. Publication approves no funding, activation or wallet migration. There is no global pause. The older 29-address ecosystem snapshots and existing wallet profiles remain pinned to their historical deployments. Use the current record's exact deployed commit/tree for reproduction; the combined publication source has a separate identity.
Publication requires the exact integrated source and generated identities, all preserved CI and
release dependencies, the complete flagship/SUCCESSOR browser, recovery, offline and accessibility
checks, and deployment-directory asset checks. Early version preflight must reject a tag
that already identifies another commit.
Historical ci.yml, pr-ci.yml and smoke.yml must pass on the same current-main commit before
packaging. The tested site must then pass public HTTPS identity and workflow checks; final assets
must be uploaded, re-downloaded and checksum-verified before publication. Existing tags and release
assets are not overwritten. No final-source or public-site pass is established by this source document.
Local and public acceptance evidence have separate captures. The original validation ZIP stays unchanged;
public-validation packs retain every public report and screenshot under the same 450 MB per-asset limit.
The release manifest's validation_evidence inventory and SHA256SUMS identify both sets.
Independent custody, providers, market access, Fleet commissioning, external dependency qualification and Mainnet activation retain their separate gates. Original research, diagrams, laboratories, earlier release records and the frozen deployment history remain preserved.
Earlier candidate records
Version 1.33.0 combines the reviewed local SUCCESSOR journeys described in the release notes. The exact combined source, generated identities, complete required CI and SUCCESSOR browser/accessibility dependency must pass before packaging. Current-main publication, public HTTPS identity and uploaded-byte verification remain required. This is a local preview using supplied evidence. Independent custody, providers, market access and Fleet adapters remain uncommissioned. A source version alone proves none of those outcomes.
Version 1.32.0 is the integrated additive candidate described in the release notes. Exact combined-source tests, all required CI, current-main packaging, public HTTPS verification and uploaded-byte checks remain mandatory. Historical passing feature branches do not qualify a later merge. The local supplier/Registry proof used actual Chromium with scripted provider responses and test-token settlement; Docker/live-provider commissioning and real financial transactions remain unqualified. The pinned CPU world-model gate must execute every objective test without skips.
The sections below preserve earlier source-specific records and are not new release evidence.
Version 1.31.3 brings Nova-Seed Studio into the public homepage, preserves French/larger-text reading choices and offers explicit local start, recovery and execution handoffs. All eleven Studio steps have a 132-layout language, viewport and text-size acceptance matrix. The same source, private operator, contract, complete-cycle, public HTTPS and uploaded-byte release gates remain required. See the Studio usability notes. The archived Mainnet deployment remains restricted; this release introduces no qualified replacement deployment or live-provider commissioning.
Version 1.31.2 validates and bounds Mainnet RPC responses in the Hardhat preparation and verification scripts. See the maintenance notes and operator guide. The exact-commit release, public website and uploaded-asset gates remain required.
Version 1.31.1 adds the English/French computer-use study and its compact offline pack while preserving the 1.31.0 product scope. Read the maintenance release notes for the publication checks and evidence boundaries. The manifest and matching validation archive identify the delivered commit and actual acceptance.
Version 1.31.0 combines Nova-Seed Studio, authorized native work, recovery hardening and inspectable successor cycles. Start with the English/French first-use guide, then use the release notes for scope and upgrade instructions. Local preparation does not require a wallet or paid provider. Publishing, funding and execution have their own prerequisites.
This is the release scope and verification record, not proof of publication. At the start of this integration,
PRs #4780 and #4781 had merged into the Nova-Seed feature branch, while #4779 still targeted main
and #4782 required reconciliation with that branch. The resulting combined source needs its own
verification. The release manifest, exact-commit workflow results, tag and public-site acceptance
identify what was actually delivered; a source version or a successful feature-branch run cannot do so.
| Evidence | How to read it |
|---|---|
| Historical 80-case record | Original requirement texts and snapshot decisions: 10 passed, 65 partial, 5 blocked; retained without changing their source identity |
| Historical cycle handoff | Real bounded three-job/two-cycle evidence, focused passes and recorded broad-suite/type-check failures at the named earlier sources |
| New integrated verification | Results must name the combined commit and actual environment; prior results qualify only their original source and scope |
| Release and public website | Require matching tag, manifest, checksums, release assets and served-site acceptance; preparing these files is not publication |
During the October 14 integration, local checks reported 148 passing MARK Node tests, 26 passing native
MARK selections and successful configured type-check scopes covering 31 and 32 files. These are local
integration observations, not final-source CI or a full-repository type-check pass. The inspectable-cycle
branch is included in the full Agent release acceptance candidate path; its actual run and the
integrated main results must be checked at the delivered commit.
The complete 80-case product scope and operational Mainnet commissioning remain separate from a software release. No new production profile, physical-wallet trial, live provider execution or external storage qualification is established by the historical fixtures. Per-edge observed actions, shared allocations and complete principal-return/cost accounting retain their documented limits.
Version 1.30.0 adds the MARK wallet with a durable public transaction journal, explicit nonce-bound previews and finalized original/replacement reconciliation. Targeted codec/provider tests, rendered injected-provider acceptance, generated-site acceptance and public HTTPS acceptance are additional gates; all previous integrated release requirements remain. Browser fixtures do not establish a physical-wallet trial, independent contract audit, live settlement or permission for new funding. See release notes.
Version 1.29.0 adds five bounded local work adapters, a portable English/French workbench and optional isolated browser Computer Use. Start with Create my Ω, retain its actual artifacts, or use the native work interface. The Computer Use guide explains the explicit provider/image permission and its separate commissioning requirements. The release notes describe the addition and the stopped-run recovery correction. Publication requires the exact integrated source, actual container execution, installed distribution, five-family public browser acceptance and all preserved release gates. The manifest records their results.
Version 1.28.1 strengthened contract compilation and exact-build verification. The warning and verification guide separates current-source checks from the archived October 11 deployment. Publishing this release changes no deployed bytecode and does not authorize a migration or public funding. See the release notes.
Version 1.28.0 added a bilingual build-your-own SUCCESSOR Ω workshop using Alpha's actual candidate engine and AGI Jobs' pinned work-order compiler. Visitors can compare four bounded programs, inspect 21 draft jobs and save the complete lab offline. The shared Ethereum mainnet $AGIALPHA token and all 29 recorded contract addresses are featured alongside the commissioning boundaries. See the release notes.
The earlier development-toolchain maintenance record retains its dated audit findings and remaining compatibility work. Existing dependency, runtime, security and public-site acceptance gates remain required for this integrated publication.
The repository guide maps the maintained launch paths, source areas and validation boundaries. This release features a personal SUCCESSOR Ω lab on the homepage: name it, compare real bounded programs, keep its evidence and download a complete offline copy. English/French, larger text and the download/reopen/rerun journey have explicit acceptance gates. The general Start here link still opens the simple planner; its existing Choose → Review → Keep experience and readable exports remain. Make your first plan or read the onboarding guide. The existing SUCCESSOR mission lifecycle, repository repairs and release-installer preflight remain. The SUCCESSOR guide, architecture and pack restoration describe that lifecycle's boundaries. A revision is publishable only after the exact integrated commit passes every existing and added gate. The published release manifest, rather than this source guide, identifies the accepted commit and recorded results.
The computer-use study reader is an editorial addition.
Both rebuilt galleries and public Pages must pass its exact-asset, English/French, large-text, print,
download and offline-reopening checks. The release includes computer-use-study-report-pack.zip and
retains the reader's acceptance report in its validation archive. These checks qualify publication of
the report; they neither rerun its original experiment nor establish measured human time savings.
The complete catalog lists all 26 entries and their prerequisites. Seventeen finite demos have installed-wheel acceptance, including separate repeat runs that retain earlier evidence. See the release notes, walkthrough and dated audit. Original research and diagrams remain preserved.
Sovereign adds a private portfolio → schedule → evidence workflow, explicit result-bound reviews, durable recovery and independently verifiable exports. Its public page exposes recorded native fixtures; local operator approval is distinct from independent validator consensus. Start the Sovereign workspace.
Finance Alpha adds a standard-library paper-research terminal, exact ledger/replay checks, Python 3.11–3.13 calculation gates and real browser/notebook acceptance. See the finance guide for the supported local scope and the separately unverified legacy/testnet integration.
The supported deployment profile is a private, single-operator agent with a persistent signed journal, plus a self-contained public browser workspace. Version 1.26.1 records the earlier mainnet deployment and its completed ownership handoff, with a separate read-only verifier. Version 1.26.0 added readable/printable first plans, exact settlement checks, a full contract-toolchain advisory gate and unsigned mainnet preparation. The final 198-page manuscript is the versioned research specification. Installation, execution, review, evidence export, recovery and publication have explicit acceptance gates. The manuscript map records implemented behavior and the research obligations that still need external evidence.
The additive Ascension protocol has an expanded local-EVM suite and mainnet preparation/verification tools. The recorded mainnet deployment has verified ownership and predates the current settlement fixes. Its participant commissioning remains pending; current-source deployment needs its own independent review, real-token fork evidence and operator configuration. The public Protocol Desk shows the executable lifecycle and explicit fixture boundaries. Existing agent state requires no migration.
Choose a starting point
| Your goal | Start here | What you receive |
|---|---|---|
| Create and recover an Ascension project | English/French first use, then complete-cycle guide | Exact project plans, private recovery and explicit native execution, review and continuation controls |
| Run and keep a useful local Ω | Five-family workbench | Bounded repairs, benchmarks, checked examples, public-data reports or numerical calculations with offline artifacts |
| Configure screenshot and keyboard/mouse work | Computer Use | Explicitly permitted disposable browser sessions, observations and stop/reconciliation controls |
| Make a first plan without technical setup | Guided first experience | English/French Choose → Review → Keep journey, larger text and text/JSON downloads |
| Construct and evaluate a bounded successor | SUCCESSOR Ω | Browser or signed native lifecycle, matched generation-two trials and empty successor proof/authority until newly earned |
| Review a governance proposal | Governance Workbench | Nine conditional gates, reproducible dossier and input-bound verification jobs |
| Make an operational decision with your records | Decision Studio | Eleven cases, editable data, constrained plans, proof backlogs and replayable dossiers |
| Measure whether a learned capability helps new tasks | Compounding Lab | Frozen policy, four measured arms, learning/review costs, negative controls and a complete Evidence Docket |
| Turn a claim into reviewed work and retained capability | Proof Bloom | Claim → jobs → evidence → review → promotion → memory, with revocation and native signed returns |
| Explore opportunities and test an agent firm | Insight Atlas, choose an expedition | Exact scenario ledger, training/holdout comparisons, claim dossiers and replayable capability history |
| Explore the original white-paper vision | Ascension Lab, choose a scenario, select Discover | Computed portfolio/schedule, encrypted recovery capsule, modeled economics and downloadable evidence |
| Solve a bounded task with your own inputs | Browser workspace, choose a mission and edit its data | Checked research, allocation, schedule or forecast; explicit review and export |
| Keep a persistent identity, private journal and operator controls | Install the agent from the matching release assets | Five native mission kinds, signed evidence, recovery, optional inference and verified payment receipts |
| Explore earlier experiments | Demo catalog | Preserved launch paths with explicit execution modes and optional requirements |
Pages needs no sign-in, wallet or API key for its built-in computation. The guided first-use page keeps entries only on the current page; download your result before reloading. The lightweight workspace works offline after installation; model text generation needs its separate initial download. Export work you want to keep. Browser storage is not a substitute for a downloaded recovery capsule and its separately retained passphrase. See privacy and offline recovery.
What the release gates establish
| Area | Required evidence | Practical limit |
|---|---|---|
| Correctness | Independent algorithm/accounting tests, full Python regression, strict types, Solidity checks | Bounded input domains and finite tests; no general correctness proof |
| Security controls | Authentication/origin checks, measured 512 KiB request limit and 15-second body deadline, private files, signed journal, tamper tests, real Docker isolation | Private single-operator host; not a public multi-tenant service or perfect sandbox |
| Dependencies | Hash-locked installation, full operator Python advisory audit, existing browser audits | A dated advisory snapshot; historical environments and host OS need separate maintenance |
| Recovery | Signed backup/restore, pause, corruption detection, persistent container restart | Backups contain secrets; archive limit 256 MiB; independent checkpoints detect rollback |
| User experience | Homepage clarity and preference handoff, English/French first-use completion, larger text, keyboard and download checks, real Chromium workflows, mobile 320/390 px, cancellation, offline reload, public HTTPS checks | Browser acceptance covers Chromium; no comprehensive accessibility certification or other-engine certification |
| Delivery | Same tested source/site, exact-commit CI, versioned tag/assets, upload re-download checksums | Hosting and GitHub remain trusted services; checksums do not establish GitHub-enforced release immutability or make separate API writes atomic |
| Manuscript fidelity | Original 198-page PDF, canonical Markdown and 31 figures checked against a pinned source commit | A paper is a research specification, not proof that every proposed scientific capability has been established |
| Transfer evidence | Frozen A-only learning, unseen B tasks, failure/ablation controls, full prior learning charges, cross-runtime replay and all 13 docket sections | Bounded synthetic forecasting; independent evidence, strongest-agent comparisons and calibrated α-WU remain HOLD |
| Preservation | Earlier 2,125-path baseline plus the 2,731-path pre-SUCCESSOR inventory, README/flywheels, original paper checksum and full catalog | Retention of an experiment does not certify its optional integrations |
Download release-manifest.json, SHA256SUMS, the versioned local validation archive and every
public-validation pack listed by validation_evidence from the
published release you select. The manifest identifies
the tested commit and workflow. JUnit reports distinguish passing, skipped and expected-failure tests.
The Python audit includes package names/versions, scan time and the exact lock digest; failed or skipped
audits block publication. Retain the matching checksums with prior release assets as recovery checkpoints.
Homepage acceptance
python scripts/validate_homepage.py checks the generated homepage and the handoff to the first-plan
experience. The release workflow runs the same acceptance against public Pages before publication.
Both reports must identify the exact packaged commit and version. The checks cover English/French
content, larger text, keyboard access, narrow-screen layouts, help, preserved advanced entry points,
and language/text-size preferences carried to /start/ through ?lang=fr&text=large.
The make-plan buttons must open the task-entry section directly; the general Start here link keeps
its introduction. Planner acceptance checks readable essential text and hides reading controls until
initialization succeeds, including no-JavaScript and module-load failure cases. Explicit review, accurate
downloadable results and input-specific error recovery remain required. These gates supplement all
earlier release checks.
A defined test is a requirement, not a recorded pass. Use the published manifest and matching validation archive for actual results. Automated accessibility checks and responsive layouts still need to be complemented by usability studies with older and nontechnical participants, additional browsers and assistive technologies.
Recover a release workflow after a retry
- Open the failed job and identify its actual error. A historical-CI waiting timeout does not mean the test suites failed: inspect the separate CI, PR CI and Smoke runs for the same full commit SHA.
- Wait for every required run on that commit to complete successfully. Repair a test failure through a reviewed source change; the new main commit must pass its own checks.
- For a timing-only failure, use Re-run failed jobs on the original release workflow. Successful test jobs retain their evidence. The historical waiter rechecks all three required workflows.
- Confirm packaging, public Pages acceptance and publication all succeed. A successful test run alone does not establish that the public website or release assets have been published.
The regression job allows 60 minutes for the complete locked test suite, cold strict-type checks and repository validation. The historical waiter reserves five additional minutes in its enclosing job for evidence upload and cleanup; its bounded wait accommodates runner queues and the historical test jobs' 60-minute execution limit. Deadlines still fail closed: successful tests cannot replace an incomplete job. The successful waiter supplies its unique attempt-specific artifact name to packaging, and the download verifies its SHA-256 digest. The report must identify the same commit and successful main-push runs of every required workflow. Packaging rejects missing, failed or stale reports. Package and public-site evidence downloads also use the producing job's unique artifact name. Artifact IDs are opaque identifiers: a larger number does not prove a newer attempt. Earlier attempt artifacts remain available as diagnostic history.
Upgrade without losing your work
- Pause the old agent. Verify its journal, make a private backup, and retain its checksum and journal head separately. Keep its existing environment and state directory.
- Download matching assets from one published release and use
install_agent.pyto create a new environment. It checks wheel/lock checksums, installs hashed dependencies and runspip check. - Restore the backup into a new private home. Verify identity and journal head, inspect your configuration and confirm the agent remains paused.
- Stop the old process before using the restored home. Resume deliberately and test representative missions. For rollback, stop the new process. Restore a pre-upgrade backup with the prior environment only if it still satisfies the latest independently retained checkpoint. Later work or revocations must not disappear through rollback; keep the operator stopped until a compatible current recovery path is verified.
Existing Ed25519 identities, native journal records, Nova-Seed capsule/readable-kit formats and v1 Bloom Chronicle events keep their original meaning. The private full-project recovery format is a separate versioned envelope; it does not rewrite native mission history or reactivate imported authority. Retain or separately reprovision the local operator configuration, trusted identities, private signer files and role-approval tokens. A public project export does not back up those secrets. Do not run two versions on the same home or discard your pre-upgrade backup. Exact commands, Windows paths and container operations are in the operator guide. For SUCCESSOR private recovery, verify the separately retained checkpoint using the installed verification procedure. Portable institution transfer is a different workflow: it restores history with active proof and grants empty. Native upgrades never deploy or migrate contracts; the contract deployment/migration guide describes the separate local rehearsal and external-chain prerequisites.
Boundaries that still require separate evidence
The Ascension economics, Council and governance are browser protocol simulations. The runtime verifies local-EVM payments; the recorded mainnet deployment establishes contracts and ownership, not commissioned mainnet job operation, deployed minting authority, live DEX activity or an independently audited production token economy. Paper-level AGI/ASI, formal invariants and physical/economic guarantees remain research claims; the latest manuscript map records the remaining obligations; the original white-paper map retains its mathematical corrections.
Long-running service reliability, public internet exposure, additional browsers, comprehensive accessibility conformance and organization-specific load targets have not been qualified by these release gates. Operate within the documented private-host scope. Keep your browser, host, Docker and backups maintained; review new advisories and preserve a tested rollback path for each upgrade.
Repository administration
The September 27, 2026 audit found main branch protection disabled and no repository rulesets. The release workflow still gates deployment and publication, but that does not protect direct pushes. The connected repository API supports the code/release work and does not expose administration writes. An administrator with repository settings access must enable the repository rule.
In the repository, open Settings → Branches → Add classic branch protection rule. Set the branch
pattern to main, enable Require status checks to pass before merging and Require branches to
be up to date, and select the GitHub Actions checks Lint (ruff) and Smoke tests. Keep force
pushes and deletions disabled, then save. These are the check-run names reported by GitHub; the
workflow-qualified labels in preserved historical README instructions are not the API context names.
The current CI enforcement guide and helper use the corrected names.
An administrator with an existing appropriately scoped token can alternatively run
python scripts/verify_branch_protection.py --apply --branch main and then the same command without
--apply to verify. Do not put a token in source, a command argument or a support transcript.
Curriculum Lab acceptance
The Meta-Agentic AGI v3 guide defines the finite grammar and independent review boundary. Release acceptance requires exact Python/JavaScript runs and bundles, answer-leakage and sandbox regressions, royalty accounting checks, wheel launch, mobile/browser accessibility, hostile imports, offline replay, preservation hashes and a public receipt bound to the release commit and exact assets.