Platform and website update — 2026-10-07

This update adds a practical machine-labor entry point to the complete demo observatory. All existing demo directories, command decks, guides and original Mermaid blocks remain available. Hypernova PR #3904 was merged after all 48 workflow runs on its exact head succeeded and both review findings were resolved.

Changes

• A homepage route for buyers, workers and independent reviewers, plus a concrete progression from useful digital work to coordinated scientific and infrastructure research.
• A browser-local planner for ten work categories. It exports exact proposals, tasks in the existing computer-work schema, and operator handoffs. Draft edits invalidate prior downloads. No worker, transaction, source fetch or external effect is triggered by the planner.
• Editable drafts can be saved and reopened locally, including unfinished work. Import accepts only bounded planner fields, invalidates generated downloads, and never imports execution authority. Switching categories preserves a custom objective; an explicit reset restores the suggested objective.
• Exact six-decimal proposal arithmetic, explicit USDC/AGIALPHA deployment distinction, source URL validation, input/output limits and independent content acceptance criteria.
• A stable readiness index and published computer-work guide. OpenClaw's native Codex readiness configuration is documented against current official sources; ChatGPT Work remains an operator-led route.
• The $40T/year opportunity is explicitly a user-supplied planning assumption. No universal automation capability, realized stellar infrastructure, buyer acceptance or production certification is asserted.

Dependency refresh

The fresh root production audit initially reported 1 critical, 8 high, 36 moderate and 22 low affected packages. Targeted compatible patch updates move proxy-addr from 2.0.7 to 2.0.8 and source-map-js from 1.2.1 to 1.2.2 across affected npm locks and the CULTURE pnpm workspace. The root audit after the npm refresh reports 0 critical, 7 high, 36 moderate and 22 low.

All 20 nested npm production lockfiles report zero critical and zero high findings. The refreshed CULTURE pnpm production workspace reports zero findings at every severity. One-Box retains two low findings and Validator Constellation v2 retains five moderate findings. The machine-readable audit binds every result to its exact lockfile SHA-256.

The remaining high findings are the previously documented braces and node-forge dependency chains, including legacy Web3.Storage dependencies. They are not waived or hidden. Counts reflect registry advisories at the time of execution and may change. See the dependency review for migration and compatibility requirements.

Primary advisories: proxy-addr IP trust validation and source-map-js indexed map validation.

Verification results

• Complete site build: 76 demo pages, 337 guides and 265 preserved diagrams.
• All 22 site tests pass, including actual adapter-schema compatibility, the two security regressions and bounded editable-draft round trips.
• Full browser verification passes with 31 check groups, zero diagram failures and zero automated accessibility violations across the tested widths.
• Gateway and orchestrator builds pass; all 47 computer-work boundary tests and 12 browser/provider/runtime tests pass.
• Documentation links verify across 288 Markdown files. The checker now decodes percent-encoded filenames correctly, and a missing incident-response link is corrected. Pages CI runs the documentation check before publishing.
• Locked dependency installation and the Node/npm preflight pass.

Reproduce validation

Use the pinned Node/npm toolchain in START_HERE. Install the independent CULTURE workspace as described in the website guide.

npm ci
npm run ci:preflight
npm run demos -- --check
npm run site:build
npm run site:test
npm run site:qa
npm run build:gateway
npm run build:orchestrator
node --test apps/orchestrator/dist/apps/orchestrator/__tests__/computerWork.test.js
node --test demo/One-Box/computer-work/test.cjs
npm run compile
npm test
npm run release:check-size
npm audit --package-lock-only --omit=dev --json

The audit is expected to report the unresolved findings rather than claim a clean production tree. Website tests include actual adapter-schema compatibility for all ten work categories, exact proposal units, invalid inputs and downloadable bytes. Browser QA checks stale-draft invalidation, input injection, keyboard navigation, responsive layouts, accessibility and the no-JavaScript path alongside the full preserved collection.

Local and CI validation establish only their measured behavior. Live worker commissioning, independent security review, authorized release trust, buyer use and real settlement remain deployment-specific gates in the readiness index.

Evidence-review follow-up

The evidence reviewer adds a complete local delivery-inspection route at /review/, linked from the homepage's reviewer route and the job planner. It compares the independently supplied admitted task and expected job/deployment with the persistent worker receipt, recomputes each artifact's SHA-256 and UTF-8 byte count, and records the exact input-file fingerprints. Artifacts are displayed as text; downloads preserve their content and append .txt. Criterion-by-criterion findings export as an unsigned assessment. No upload, worker dispatch, provider authentication, buyer acceptance or settlement is inferred.

The gateway previously accepted malformed UTF-8 by replacing invalid bytes. This was reproduced against the prior source with a local HTTP response. Transport now rejects malformed encoding; task/evidence strings reject unpaired Unicode surrogates, and the adapter enforces the 128,000-byte artifact limit in UTF-8 bytes. Rejected worker output retains the dispatch replay barrier. The exact multibyte boundary remains accepted.

Local checks on the follow-up pass: 583 contract tests, 63 gateway/orchestrator/provider/runtime checks, 28 website/model tests, both TypeScript service builds, the contract-size gate, toolchain/lock preflight and documentation links across 289 Markdown files. The site builds 76 demo pages, 338 guides and all 265 original diagrams. The evidence tests consume actual adapter-generated local HTTP receipts and cover tampering, substituted tasks, replay, invalid encodings, unsigned provenance and acceptance requirements. Browser checks cover the real file/download journey, text injection, edits during asynchronous reads, five viewport widths, accessibility and the no-JavaScript path; their results and screenshots are emitted to the Pages CI evidence artifact.

The full browser verification passes 34 check groups, with zero diagram failures and zero automated accessibility violations on the tested surfaces. The separate CULTURE browser journey also passes its lifecycle, export, responsive-layout and accessibility checks.

The current root dependency audit remains 0 critical, 7 high, 36 moderate and 22 low; this change does not suppress or remediate those existing findings. Read-only GitHub inspection on 2026-10-07 found main at 16cd23de77670b5fe5e75246654e59ccd6a28fb5, all 17 associated workflows successful, protected: false, and no active repository rulesets returned. Documentation now distinguishes intended CI checks from enforced merge protection. Branch rules require separate administrative configuration; no repository security policy is weakened by this update.

The prior validation counts above describe the earlier update. Follow-up results establish their stated local behavior, not certification of all historical demo claims or readiness for unattended mainnet operation.

Delivery and release boundary follow-up

The planner now rejects malformed Unicode and UTF-8 before producing a task, disables stale exports immediately when an editable draft starts opening, and preserves edits and keyboard focus made during a delayed import. Four stale ASI Take-Off walkthrough anchors were repaired; generated-site checks now verify linked sections as well as files. The full collection, original sources, visual design and all 265 Mermaid blocks are preserved.

Computer-work admission rejects unsafe profile/state ownership, writable shared configuration, non-regular profiles and symlinked state. Profile admission opens once without following symlinks, then validates and reads the same file descriptor; replacing the pathname cannot substitute unchecked content. Newly created journal-directory entries are synced before a dispatch can occur; a failed durability check cannot create a claim or contact the worker. The setup guide records permission migration and actual runtime commissioning requirements, including current OpenClaw/Codex and ChatGPT Work distinctions.

The generic agent:validator rehearsal uses the current domain-bound contract commitment, keeps private deployment-scoped reveal secrets before broadcast, and reconciles them after restart. Exclusive durable intent files prevent automatic duplicate sends after an uncertain result. Burn-required jobs resolve their own confirmed registry receipt before preparing a commitment; saved receipts are revalidated without replacement before either broadcast. Tests exercise distinct burn receipts for two real jobs through restart and reveal, corruption, concurrent events, changed rounds/chains, lost responses, timeouts and reorganizations. An explicit fixed decision is mandatory and remains a rehearsal input, not independent content review. Other service paths require their separately documented recovery and commissioning checks.

The orchestrator and gateway now share the current six-argument reveal interface, domain-bound commitment encoding, confirmed burn-evidence lookup and actual Solidity round-getter layout. Burn receipt searches use bounded pages and adapt to smaller provider limits without omitting older history. Reveal scheduling uses the chain's round deadlines and rejects expired or unrepresentable timer delays. After a mined commitment, transient deadline-read failures retain bounded scheduling retries; an uncertain transaction is never automatically resent. Existing commitments and unreadable saved records require reconciliation instead of replacement with a fresh salt. The gateway's automatic validator reads and verifies the original committed specification and abstains on computer work before notification, staking or voting; worker-supplied labels cannot authorize an automatic vote. Real-contract regressions exercise both verdicts, the actual orchestrator methods and gateway helper state reload. These protocol repairs do not retrofit the rehearsal journal into every service.

The gateway's manual and automatic validator paths share chain, module, round and canonical-block scope. A confirmed later round can atomically archive its predecessor and create a new vote even when a contract reset reuses the nonce. Locked scope/hash comparisons reject delayed confirmations from an older round; prior reveal flags are not carried into the new record. Both paths claim the same durable reveal intent before broadcast, then persist the returned transaction hash before awaiting confirmation. Manual recovery uses the authoritative saved vote and canonical receipt/event evidence to recover a lost response without resending. Automatic handlers reconcile records when processing selection/submission events, retain bounded read-only retries and cancel stale callbacks. Duplicate selection callbacks preserve active work and retry budgets. Uncertain broadcasts are not automatically resent. Legacy, corrupt and orphaned records remain blocked for reconciliation; missed-event discovery and live restart commissioning remain deployment responsibilities.

Validator records now use canonical identifiers, protected non-symlink directories/files, bounded descriptor reads and strict UTF-8. JSON is limited to 1 MiB with depth/node limits, and deletion uses the same lock as writers. The optional absolute VALIDATION_STORAGE_DIR supports a deliberate private-volume setup; default filenames and location remain compatible. The gateway runbook explains permission migration and uncertain-state handling. CodeQL's medium network-to-file heuristic reports the intentional persistence of network-derived JSON; these records remain untrusted data and are neither executed nor served as active content. The source-to-file flow stays visible to analysis, with no suppression.

Production releases now require a fresh audit of every tracked npm and pnpm production lockfile. The gate retains raw registry evidence, lockfile/manifest hashes and source identity, and rejects critical/high findings or missing/invalid evidence. Its live check covered 22 lockfiles: 21 met the severity gate; the root retained 0 critical, 7 high, 36 moderate and 22 low affected packages. No advisory was waived, and no unsupported dependency override was added. Fresh registry and primary-advisory checks still found no patched compatible braces or node-forge release. This gate is independent of successful CI and Pages deployment.

Local validation passed 669 contract tests and 89 compiled runtime/provider/packaging checks. Website validation passed 36 model/site tests, 37 browser check groups and the separate complete CULTURE browser journey. All 265 diagrams rendered; the tested surfaces had zero diagram failures, browser errors, external requests or automated accessibility violations. Contract-size, pinned-toolchain, lock-integrity, formatting and documentation-link checks passed. The PR's exact-commit checks provide the final integrated validation record; these local observations do not grant release signing trust, independent security approval, buyer acceptance or live settlement authorization.