Skip to content

Project notice

1.37.1 — OUT-EXECUTE and simpler research publication

Combined software scope

This candidate combines the capability workspace and bounded Out-Evolve work described in the preserved 1.37.0 proposal with executable OUT-EXECUTE contracts, seven featured papers and an additive paper-publication lane. The owner approved qualifying this combined source as the next release rather than waiting for the unpublished 1.37.0 candidate. Version 1.37.1 was unused when prepared; no published release, tag, asset or earlier evidence is overwritten. This is not a documentation-only release.

OUT-EXECUTE in the existing capability workspace

The capability workspace now executes bounded finite nonnegative integer min-sum models. It constructs complete boundary profiles, independently verifies each retained contract and its evidence, distinguishes preserved from invalidated work, and assembles and accepts a current result. Pointwise equality of the complete profile can stop coordinator re-optimization; an equal scalar optimum or a self-declared hash is not a proof of behavioral equivalence. Current witnesses are still reconstructed and checked.

The existing project history retains source bindings, attempts, lineage, failures, unknowns, budgets and actual logical-work counters. Interrupted checkpoints and ordinary/unsaved recovery exports can be reopened without automatically running work. Stale evidence, source/evaluator/schema changes, non-equivalent profiles, forged terminal states and divergent histories are rejected or require new work. The pane shows the contract, change, reuse/invalidation decisions, acceptance and costs without creating a separate product or history store.

The implementation guide (repository source) specifies the typed interface, finite bounds, trust assumptions and accounting exclusions. USDC6, AGIALPHA18 and Fleet handoffs are inert, source-bound review drafts, not submitted jobs or executable authorizations. Existing Jobs/Fleet runtimes do not install this new adapter. Fleet qualification remains unprovisioned; no denied review is represented as cleared.

Reproducible evidence and honest comparisons

The local authored-study record retains exact compiler/checker/harness hashes, counts, accounting, reproduction command and fixture custody hashes. The 64 freshly authored fixtures produced 576 snapshots matching exhaustive optimization and 160 semantic stops. This is not reproduction of the manuscript's 576/158 experiment: its companion archive was not available. The complete model/result/archive files were retained locally, not embedded in the compact record.

The candidate and the actual equally capable incremental baseline tie at 13,926,222 counted operations; full enumeration uses 94,588,992 on that bank. A legitimate fixed-proof baseline is cheaper on the identical-input control: 45 then 23 operations, versus 137 then 208 for boundary construction/reuse. No universal speed advantage is claimed. Construction, qualification, reuse verification, comparison runs, archive reopening and output size are reported; reopened-archive profile comparisons are charged. Logical operations are not CPU instructions, wall-clock savings, currency or industrial benefit. Host JSON/hash/allocation overhead is not included in that operation unit; elapsed study time and serialized bytes are reported separately. Source hashes bind bytes, not external authenticity. Physical, industrial and general-AI benefits remain hypotheses.

Seven original papers and a bounded publication lane

The README and bilingual homepage feature TRANSFORMATIONAL INTELLIGENCE ∞, Out-Learn 📚, Out-Think 🧠, Out-Design 🎨, Out-Strategise ♟️, Out-Execute ⚡ and Out-Evolve ♾️. Accessible direct English PDF links, French/English summaries, larger-text controls and no-JavaScript reading preserve the flagship's free-start entry and history. All seven original PDF/TEX pairs and canonical URLs remain unchanged. Out-Learn describes constructed measurement protocols and reusable calibration under declared drift and measurement-error assumptions, not physical qualification. Out-Design studies equivalent implementations that preserve required services and release capacity for change in finite executable models. A research link does not imply an implemented or commissioned runtime capability. The TI foundation leads a linked reading index and responsive companion grid immediately after the free-start intro.

The paper-upload guide describes one catalog import and the normal pull request. A pre-change, fail-closed classifier examines every introduced commit. Only validated paper/catalog changes and exact bounded presentation/cache reproduction may reuse an immutable qualified software baseline. Mixed changes, unknown paths, workflow/runtime edits, unavailable history or failed proof take full software qualification. Installing this lane and this combined OUT-EXECUTE release therefore require the full lane, never paper-only acceptance.

Future paper overlays wait for their matching software release, retain separate paper-site provenance and cannot relabel software artifacts or roll back a newer site. Existing software tag, manifest and site-pack identities stay source-bound. Publication guards and the serialized Pages boundary remain mandatory.

The combined site retains all 997 published file paths, directory index routes and known anchors. The 28 served PDF paths retain their original bytes; the seven featured PDF/TEX pairs remain at their original repository paths. Additional locations are copies, not moves, renames or redirects. Normal HTML updates remain possible. Packaged and public acceptance check this compatibility floor against the exact new distribution, and finalization rejects missing or stale evidence. The additive PDF repair receipt is retained as a historical record, not a claim about the new software's qualification.

Qualification and preserved boundaries

The complete read-only capability acceptance workflow is now also an explicit dependency of Agent release packaging. It retains exact-checkout/source receipts, generated-byte checks, computation/history tests, the freshly executed four-fixture study, browser/recovery journeys and accessibility evidence. Separate standalone and release invocations cannot cancel one another. All prior release dependencies remain required, with unchanged permissions. The historical Integration test-job deadline increases from 75 to 90 minutes, retaining both Python versions, full coverage and all steps. Configuration/timing evidence indicates the older Python 3.12 run hit the 75-minute limit at 98% progress; no direct timeout annotation was available, so the cause remains a strong inference, not a verified annotation. This finite headroom does not turn that cancelled run into a pass or waive fresh qualification.

Before publication, the exact integrated source still needs configured PR checks and fresh source review, authorized merge, complete main-source Agent acceptance and historical CI, full security checks, packaging, public HTTPS readback and uploaded-byte verification. Prior local source reviews or old four-paper screenshots do not establish those results. Local browser launch and Semgrep fetch were unavailable/denied; neither was retried by another route. Consult the matching workflow, release manifest and retained reports for actual acceptance and delivery.

The original README, mission, flywheels, selected essays, flagship routes, laboratories, membership/contracts, wallet profiles and separate USDC6/AGIALPHA18 routes remain. No signing, funding, payment, activation, credential change, new deployment or operational authority is granted. The compatible live Jobs/Fleet release set remains incomplete. Existing research and historical release/deployment records retain their original scope.

See first use, release readiness, operator guidance and featured papers.