Autonomous Enterprises — Build the Means of Building
This additive source preview extends the existing capability workspace and its thirteen-section Evidence Docket.
It preserves Discovery Lab, Out-Evolve, the six research stages, original paper files and all existing mission events.
The maintained homepage links to successor/capabilities/#autonomous-enterprises in this source tree.
Public deployment requires the existing owner qualification and publication pipelines.
The constructor and checker are separately authored implementations of the selected 23-page manuscript. No audited manuscript code/results companion was obtained from the inspected public main tree, code search or recent release assets. This is an independent reimplementation, not reproduction of the paper's Python package.
What executes
- Exact shortest paths over optional durable capability stocks, actual recipe/stock witnesses, a separate decreasing-stock dynamic program, capital envelopes, the retention-law witness, dominance and bounded cover certificates.
- All 64 signed/layout candidates at each Walsh size 2, 4, 8, 16 and 32; all four convolution normalizations; every one of the 1,024 ordered convolution basis pairs; both complete 32-basis inverse qualifications.
- Inert, content-bound JSON ASTs with structural linear/bilinear semantics, exact bounded rational arithmetic and topological dependency admission. Generated data are never passed to an evaluator of arbitrary source code.
- Construction and acceptance in the existing mission chain, freezing before later input generation, five actual treatment paths, simulated work reconciliation and a separately checked inverse for a changed kernel.
- Single-use in-memory execution leases. An interrupted/reopened execution reservation is uncertain and cannot silently rerun. Unknown historical reservations block a new attempt in the same mission. A genuinely unstarted budget refusal records known zero production work while retaining its inconclusive attempt.
Reproduce and inspect
The repository pins Node 22.17.1. The original hosted candidate passed all 516 Node tests on that runtime, then failed
a workflow-policy assertion; its archive and logs remain retained. The current source passes 113 enterprise tests plus 47 owner metadata tests locally
on Node 24.19.0 and 22.23.3. Its complete n32 CLI journey and thirteen-section docket independently replay.
At public commit 72075cf, the exact-tree hosted capability workflow passed 576 Node tests, strict typing and the sixteen-check
enterprise browser journey on observed Chromium 155.0.8059.0. Its original archive is retained and hash-verified.
The browser harness also reviews every contrast-incomplete node in actual frozen, terminal desktop and mobile states;
use the current head's hosted artifact for its result. Local browser execution remains blocked by missing Python test
dependencies and Chromium IPC permissions; the rejected attempts are retained. Fresh exact-source hosted acceptance
remains required. No model, wallet, account, package installation or network access is needed for the deterministic study.
node --test examples/autonomous-enterprises/*.test.mjs
node examples/autonomous-enterprises/reproduce.mjs --output /tmp/autonomous-enterprises-evidence
python scripts/generate_successor_capability.py
python tests/test_autonomous_enterprises_source.py -v
The output directory must be empty. Retain the original generated files and SHA256SUMS. The complete workspace,
raw candidates and certificates are in workspace.json and foundry.json; evidence-docket.json uses the canonical
thirteen sections. capital-confirmation.json preserves every generated model and result. Scientific operator bytes
and arithmetic counts are deterministic; mission UUIDs, timing and environment observations are separate run identities.
The real browser acceptance is tests/browser/test_autonomous_enterprises_browser.py. It constructs and freezes,
restores an authenticated accepted-construction prefix in storage and reloads it, freezes the retained construction
once without revealing future inputs, then
supplies future inputs, invokes all five treatments, constructs a changed-kernel successor, downloads and independently
replays the actual docket, reloads the saved mission, and checks desktop/mobile, keyboard and pinned Axe accessibility.
The existing capability workflow retains its original checks and adds these tests and raw browser evidence.
Saving a new successor proposal is refused while an enterprise attempt is active, both by central history replay
and by the interface, so continuation and safe stopping retain their original version. New witness evaluations and
Discovery Lab appends are also held until finish or Stop. Continuation checks every remaining event plus a recovery
slot before work; an older near-capacity history stays readable and can use its remaining slot to Stop. An already
full historical chain stays readable and exportable, but cannot append another event.
Actual counted results and their denominators
| Comparison | Declared arithmetic work |
|---|---|
| Complete transform chain | 185,376 |
| Convolution construction and selected qualification, incremental | 592,132 |
| Inverse preparation and direct qualification | 78,016 |
| Same inverse preparation and inherited qualification | 30,912 |
| New convolution solely to qualify one inverse | 623,044 — choose the direct route |
| Specified isolated architecture, recomposed from executed receipts | 1,226,276 |
| Shared architecture | 808,420 |
| Independent same-primitive joint planner | 808,420 — tie |
| Complete executed comparison construction, including both inverse controls and dependency checks | 1,508,042 |
Operating arithmetic is 160/544/384 for the generated transform/convolution/inverse at dimension 32. The fixed fast specialist ties 160/544 and uses 96 operations for this particular inverse kernel. Its supplied implementation and mathematical authoring are not free experimentally: those costs remain unmeasured. The removed-capability control executes the direct definitions. The from-scratch and joint arms actually rebuild; they receive the same declared per-treatment allowance and retain their additional current qualification work.
The complete retained local journey records 9,702,282 lineage units for construction, future treatments and successor.
Repeated history/admission replays are additional work recorded by session-replay-account.json; do not confuse
a selected 808,420-unit architecture with the full research or controller campaign. The session admission meter includes
bundle checks also named in future reports, so it must not be blindly added to descriptive historical totals.
Indexing, allocation, serialization, hashing, control overhead, rational bit growth, engineering, institutional review,
wall-clock overhead, money and energy are outside the declared metric or unmeasured, not zero.
Default reservations distinguish construction (2,000,000 units), each future treatment (2,000,000 units), all fifteen
treatments (30,000,000 units), current bundle admission, and verification. admissionLimit is a per-complete-replay
allowance, not a cumulative whole-session cap. The session account retains repeated known replay charges and failed
replays with unknown partial work. Search limits in the capital API apply per inner completion query, not globally
to a portfolio call. The finite caps do not claim hard JavaScript preemption or industrial-scale planning.
The independent 72-catalogue confirmation checks 32,256 completion values, 28,976 capital targets and 897,024 triangle triples. The paper reports 28,472 targets under its different generator. The disagreement is disclosed; the paper's 51 new tests and preceding 40 tests/ten output reproductions are not claimed here.
The separate synthetic cash example retains whole policies, actual modeled funding traces, the 5-versus-13 witnesses and the 12-unit reserve requirement when seven funding units are refundable. It establishes no collected revenue. The capital API admits bounded exact integer charges, at most ten capabilities/64 recipes/32 producers; complete tables are bounded to seven capabilities. It does not prove a rational-price, timed, consumable or vector-resource extension.
Ownership and remaining acceptance
| Lane | Existing owner work | Remaining dependency |
|---|---|---|
| Alpha | #4855 stacked on #4851; current owner parent 24d58598; #4853 retained | Exact-source hosted checks/review; qualified normal integrations and fresh main release |
| Manager | #1617 stacked on current owner #1616; #1613 retained; maintenance v1.19.6 is separate | Final installed Discovery/enterprise source and compatible profile; USDC6 only |
| Jobs | #4010; prerequisite rc.17 is now published | Final installed receivers and a later compatible profile; AGIALPHA18 only |
| Fleet | #64; held #61 remains separate | Eligible owner/custody review, final installed acceptance and compatible publication |
The new native inspection contract has additive source-only receiver candidates in the other three repositories, with actual local interoperability evidence. Installed source custody, durable receipt commit integration and release/profile qualification remain unqualified. The existing 39-record inventory remains historical; fresh approved RPC/ABI/role/token observations are NOT_RUN. The public real-task operational pilot remains NOT_RUN. Broader manuscript implementation, all-four-homepage placement, final native round trips, Windows-hosted checkout, fresh releases, downloaded-byte acceptance and public served journeys remain explicit owner-qualified gates. No maintenance release or local proof is represented as the complete four-repository launch.
The controller reconciles only simulated work and has no external payment, publication or transaction authority. Its imported evidence cannot approve spending, create customers, collect invoices or promote itself.
The configured review of the original integrated child found two real issues: unchecked bundle disclosure metadata and unrecorded partial work when successor qualification throws. The repair closes the complete bundle/disclosure frames and records failed successor replay. Retained fail-before/pass-after controls cover inert metadata alterations, a late stale-work-bound exception, a fully counted negative spectrum certificate and known-zero budget refusal. These source fixes require fresh review; the original findings are not erased or represented as organizational approval. The next configured review found a reload boundary where accepted construction could not continue to freeze. The repaired interface offers that safe transition and stops at the freeze before enabling future inputs. Its actual browser saved-prefix fixture fails on the previous page and passes all thirteen checks after repair; it does not claim an operating-system interruption test. The source candidate also preserves the newer owner's two inert checkout metadata files and the merged opaque-paper attributes/test, including exact original manuscript bytes.
The next configured review identified a P1 version boundary: saving a proposal during an active enterprise could strand that attempt. The current repair refuses that event in central replay and API preflight, and disables proposal editing until finish or stop. Three failing ordinary cases now pass along with the preserved legacy case on both local Node profiles; the actual browser checks proposal gating and recovery. Fresh configured review remains required.
The next configured review found that unrelated witness or Discovery events could consume the slots observed at enterprise start. The history-capacity repair blocks new unrelated appends while the attempt is active and checks remaining continuation plus recovery capacity before computation. Valid historical interleavings still parse unchanged. The retained nine-case regression fails four cases on the prior implementation and passes all nine after repair; it covers dense histories, pre-work refusal, safe Stop, legacy activity and unknown reservations. Current local Node and CLI results do not replace the fresh browser, hosted, installed-custody or release gates.