Project notice

1.36.0 — Bounded transformational discovery

This is a draft source-release proposal following the 1.35.1 Computer Use reader and portable-history corrections. It does not identify a published release or a final compatible four-repository set. Exact-source hosted acceptance, packaging, public-site readback and the separately reviewed release selection remain required. Historical 1.35.x releases, records and source identities remain intact.

Installed discovery mechanisms

The native public-source pilot has three installed operators: representation, production and their complementary combination. It develops and freezes an executable source-clause selector, then compares it with unchanged-frame, isolated-edit and unrestricted controls under one fixed structural instrument. The native transformational-intelligence/v1 charter permits public, non-personal material and fixes both currency caps at zero. Structural coverage is evidence about this bounded task; it is not measured reader benefit, calibrated real-world success or an advantage over unrestricted discovery.

The public-preparation companion builds a short reading card from two pinned public Manager documentation snapshots. A separate source/prerequisite checker and exact historical import retain the original material and comparisons. The unrestricted control ties the constructed method; the card is a documentation aid, not deployment guidance or measured reader benefit. The existing real-model release process now requires the public pilot using its pinned llama.cpp/Qwen3-4B process. That exact-head result remains pending and is separate from the retained deterministic run.

An optional OpenAI-compatible loopback model adapter can propose a program in the closed selector language. Its one-call permit binds the request, prompt, configuration and source. A separately approved native campaign evaluates the proposal. Pause and cancellation revoke queued permits; an uncertain or interrupted call cannot be silently retried. Loopback fixtures exercise the actual HTTP interface, but no real model weights, model advantage or provider billing were qualified. The local no-charge declaration and reported model identity remain operator/provider attestations.

The separate ti-affine-recurrence/v1 mathematical driver executes scalar, cycle-detection and affine composition programs with an exact finite checker and an algebraic witness. All three solve the 864-case confirmation census. The frozen affine candidate uses 43,488 logical units, versus 30,204 for the stronger cycle control, so the result retains cycle detection. Including development and all comparisons makes the single-census net difference negative. This is a known elementary reduction with computational checks, not a new theorem or a proof-assistant certificate. Native source-selector importers do not accept this separate domain schema.

The complete-campaign policy study trains on two public constructed campaigns and confirms on four fresh campaigns with the same initial agent, archive access, charter, tools, permissions and instrument. Direct complementary search reaches the same structural coverage as full operator enumeration with eight fewer logical search units per campaign. Complete child campaigns, unsuccessful comparisons and costs remain in the signed Journal. This finite result does not establish independent sampling, held-out custody, calibrated reader benefit or general learning-to-discover.

The separate ti-source-package-audit/v1 driver executes all 64 combinations of six bounded choices: fact representation, retained information queries, dependency closure, report production, task batching and checked coordination. Its seventh axis proposes a real evaluator-disagreement case through the existing separately signed measurement process. The optimizer cannot authorize its own revision or carry confirmation evidence into a changed epoch. All methods still owe the same exact inventory.

Its six-file public-source fixture retains an unsuccessful frozen candidate: a declared stale-worker fault corrupts the unchecked handoff, while 32 of 64 confirmation methods deliver the exact inventory. The complete comparison charges 4,832,867 logical units; the unrestricted oracle uses 22,721 units. These are finite static parsing and local coordination mechanisms, not general discovery of arbitrary representations, experiments or evaluators. Source text is parsed without execution or network access. The source-audit guide records controls, costs, fresh-source custody, explicit approval, recovery and the separate measurement role boundary.

Persistent projects and reuse

The guided English/French workspace retains browser history and exact imported native evidence in a versioned shared TI project. Native requests, results, supplier inspections and Computer Use handoffs retain their original bytes and source attribution. Conflicting, malformed, older-prefix and unrelated imports are refused. Delayed imports cannot replace edits made while validation is running. Recovery exports preserve history; importing a project does not grant its embedded claims current authority.

Manager USDC6 and Jobs AGIALPHA18 routes remain separate inert proposal boundaries. Actual native free-campaign inspection is supported. Positive paid proposals are refused under the current zero-cap charter; there is no funded native TI route. Existing Computer Use permits, public-document controls and native approval boundaries continue to govern any separately authorized session.

Versioned transformation-library archives retain the full native package and instance evidence. Distinct configured verifier and controller signatures are required for local reuse promotion, with explicit scope, prerequisites, costs, failure conditions, retirement and rollback. Routing checks the approved portable behavior binding against the current environment. A selected entry grants no execution, access, publication or spending authority; a new task needs its own valid charter and native approval. Configured local role separation is not independent organizational review.

Migration and recovery proposal

Preserve the complete private backup, original exported JSON/HTML and exact old runtime before an upgrade. A native backup includes identity material: keep it private and never add it to a public evidence bundle. Test the new installation with a separate home; do not run two versions against one Journal. Keep the existing 1.35.1 reader fixes and prepare fresh permits after upgrading.

Import complete history before adding new results, inspect the retained project identity and source bindings, and preserve raw earlier evidence. A source, model, configuration, charter or instrument change requires fresh approval and, where applicable, a new measurement epoch. Historical signed reviews and supplier inspections do not become current approval by import. Cancel or pause work before changing its environment. Retain uncertain calls and interrupted reservations; do not replay them to obtain a more favorable result.

Recovery must preserve cancellations, accounting and append-only history. Restore into a new home with its matching runtime and verify the retained checkpoint before continuing. Do not replace a newer Journal with an older snapshot to erase a revocation or charge. Earlier exported Computer Use returns remain historical evidence. Recovery of an unexported legacy Journal missing its signed worker digest is not qualified; keep its original runtime and exports rather than infer attribution from current files.

Qualification still required

The release acceptance workflow requires the dedicated TI workflow before packaging. Its local gates exercise paper reproduction, the native campaign, model and mathematical controls, policy study, measurement/library boundaries, generated sources and the offline browser journey. Passing local tests does not establish exact-main hosted qualification or delivery of a saved package. Final release evidence must name the actual compatible Alpha, Fleet, Manager and Jobs commits and verify the package and served bytes. Live mixed-version recovery and independently measured outcomes remain separate acceptance work.

No smart contract, compiler profile, historical deployment, ENS configuration, on-chain authority, funding or activation change is proposed here. Ubuntu Hardhat creation retains hidden local ephemeral deployer-key entry; Ledger governance remains on Mac Studio. Live signing and deployment remain separately authorized. The original Transformational Intelligence paper and companion artifacts remain unchanged.

Implementation scope · Start here · Release readiness · Operations and recovery · Computer Use setup