Skip to content

Build the Next Discovery Lab

Implementation candidate for integration #4850. The additive interface is generated into docs/successor/capabilities/index.html#discovery-lab. Publication is established by exact-source release and served-byte evidence, not this source file.

The supplied virtual workcell has three or four binary controls. Its constructors produce a drift-cancelling measurement program, an updatable representation, a joint service realization, a robust conditional route, complete boundary contracts and an installed reusable experimentation platform. Independent finite checkers execute the corresponding obligations. The platform is frozen before screening, inverse-design and constrained selection queries. A stronger successor must preserve the predecessor's original domain and operations.

Reproduce

Use the repository's pinned Node 22.17.1. No credentials, provider, wallet or payment are needed.

node --test --test-concurrency=1 examples/successor-capability/*.test.mjs
node examples/successor-capability/discovery-lab-reproduce.mjs --output /tmp/discovery-lab-evidence
python scripts/generate_successor_capability.py
python tests/browser/test_discovery_lab_browser.py
python -m scripts.validate_discovery_lab --site docs --output /tmp/discovery-lab-packaged-evidence

The browser command uses the existing pinned Playwright/Axe setup in SUCCESSOR capability workspace CI. Its receipts bind actual page bytes, source manifest, browser version, accessibility results and screenshots. The CLI writes full reports, parent chains, a case summary and separately measured elapsed observations. Public cases are regression inputs, not protected holdouts. The packaged validator requires source files to match the exact Git checkout. After the normal release workflow builds and deploys that source, add --url https://montrealai.github.io/AGI-Alpha-Agent-v0/ and point --site at its matching site package to check actual HTTPS page/manifest bytes before and after the same complete journey. A local validator pass establishes no public deployment.

Contracts and custody

  • discovery-lab-charter/v1 freezes purpose, model scope, operations and per-phase limits.
  • discovery-lab-campaign/v1 contains six closed discriminated stages and all comparison/query results.
  • discovery-lab-artifact/v1 binds each stage to mission/revision, purpose, charter, model, full generated source closure and checker identities, predecessor digests, assumptions, executable data, attempts, costs, dependencies, invalidations and local execution mode. Structural validity grants no authority; current semantic checking remains mandatory.
  • successor-capability-workspace/v2 extends the original exact workspace validator and event chain. Original v1 events are preserved. Reserved starts, completions and interruptions live in the existing store; no competing history database is created. Import restores historical evidence, not authority.
  • successor-discovery-work-request/v1 selects one zero-cap USDC6 or AGIALPHA18 draft route. Its installed inspector checks the entire current finite campaign, ancestry, target stage and exact input artifacts. An unsupported or unprovisioned receiver remains explicit. Inspection never submits or settles work.

Resource account

Charter caps apply separately to construction and a full deterministic verification replay. The exact total reservation is shown before work: twice each additive cap, peak virtual workers, a 500 kB recovery/report allowance, a 120-second local elapsed cap and a 10 MB encoded-retention allowance. The latter covers the 2 MB workspace and an enforced 8 MB verification cache; it is not a measurement of JavaScript heap use. Runtime memory counts abstract resident words, separately from retained archive bytes.

The whole-history account sums every successful, rejected and inconclusive attempt plus recorded verification. It preserves unknown interrupted-work reservations. A separate current-session meter records actual import replays, failed replays, exact-content cache reuse and bounded cache retention. Export the account receipt to retain that session observation. Identical immutable evidence may reuse its already checked content under the same loaded runtime; this is neither a fresh experiment nor external acceptance.

Elapsed receipts are self-reported local observations; importing them does not repeat their measurement. Use each history entry's Show campaign control to inspect an earlier frozen result after a later failure. Selection does not run construction or change the edited charter; review exports still require current source and lineage qualification. Reload explicitly opens the latest retained result. An overrun retains the finite program but makes wrapper qualification inconclusive and prevents parent reuse or commission export. Acquisitions, logical work, encoded bytes and virtual workers are different units. Each zero-cap review draft reserves the ancestor checking caps plus 10,000 declared receiver-admission units and the shared DISCOVERY_LAB_INSPECTION_CHECKS 10,000-unit native inspection charge. The latter is charged once per inspection attempt, including framing failures; it is not measured VM instructions. Inspection memory charges four canonical UTF-8 footprints per owned snapshot plus finite verifier words, not JavaScript heap measurements. Effective remaining limits must be exact safe vectors no larger than the committed request. The inspector validates the complete envelope and its separate payload commitment before semantic replay. Its measured elapsed cap is a soft deadline checked before and between synchronous replays and at report construction: an overrun retains spent work and returns incomplete, without claiming hard preemption. A zero elapsed allowance starts no semantic replay. Canonical encoding captures data descriptors and bounded array length once. Each semantic validator checks and returns one inert parsed snapshot, so later caller reads or changes during hashing cannot replace the data that was validated.

Currency caps remain zero. No industrial or economic savings are inferred from these counts.

Evidence and boundaries

See paper-to-code map, source custody, foundation reproduction and scoped model review. The candidate release notes describe upgrade, recovery and delivery gates. The independently constructed complete cold reference uses the same declared primitives and objective. It may tie or win; component controls and fixed certificates preserve their own valid scopes.

Manager #1613, Jobs #4010 and Fleet #64 contain the receiver proposals. Their public checklists distinguish source review, current semantic inspection, installed qualification and release publication. This finite synthetic model establishes none of real materials performance, arbitrary invention, physical deployment, revenue, token value or general intelligence.