Factory docs, home
Page navigation

Historical qualification: Historical source 7cf4b26e: Core 22/22 and Dashboard 26/26 focused checks passed. Scoped evidence is recorded in the historical local qualification record. Historical source 7cf4b26e passed the 3,286/3,286 full checks and all 36 installed families. Independent bounded local evidence acceptance passed; exact-final-head CI is recorded separately from this frozen local proof.

Status: In progress. Reviewed local candidate, 3 October 2026. Independent source and retained installed stage evidence passed. Bounded concert evidence retains its independently reviewed snapshot. Historical source 7cf4b26e passed the 3,286/3,286 full checks and all 36 installed families. Independent bounded local evidence acceptance passed; exact-final-head CI is recorded separately from this frozen local proof. Separately commissioned semantics retains native quotes, original allowance and explicit shadow policy; v1 collection remains a no-model workflow. Historical reviewed scope.

Intelligence: none — Caller ports propose bounded semantics; native quote support and current authority checks decide retention. This module has no model transport.

runResearchSemanticJob(request, ports) records one explicit semantic commission over selected retained native v1 collections. inspectResearchSemanticJob(job, ports) reads historical output and separately checks current native quote support. Both need caller-owned storage, actual native Research readers and an explicitly qualified caller fence. Importing the entries creates no store, reads no clock and contacts no provider. The module has no transport, scheduler, Factory registration, live model adapter or admission operation.

Original scope and authority

The caller supplies Product, semantic commission, values, operating policy, a source-classification catalogue, original allowance pins, a commissioned logical window, fixed Budget and an explicit predecessor watermark or null. The callback is sealed when the synchronous fence returns or throws, and its exact result must be returned unchanged; the contact result may be an authentic native Promise. The fence must resolve the actual records, authenticate READ, bind this Journal identity and serialise its callback with the same physical lock used by current commission and policy writers. There is no allowing default. Fresh mutation and contact require unchanged current authority. Settlement closes an original exposure under READ; it neither authorises contact nor renews its deadline.

The runtime includes the passive native Research/Discovery/Journal reader closure; it is not an I/O-free dependency graph. Importing it performs no collection or contact. An additive constructor-registered native reader seam captures original methods without invoking accessors; prototype forgeries, structural substitutes and method overrides are refused.

A native reader is the actual createResearchJobs API for the named producer and store. The job compares original admission and collection identities, traverses bounded evidence pages to establish capture membership and calls native verify on each selected quote. It never calls native admission, collection, cancellation, citation generation, transport or cache text acquisition. Native source text is untrusted. A verified citation proves retained quote bytes, not the truth of an interview or claim.

Source classes come from the original caller catalogue: customer voice, competitor, press and first party. A collaborator cannot relabel them. Rejected or unverifiable quotes never support findings or proposals. Support is checked before contact and again after a reply; inspection observes current support without changing historical Signals or their original checked time.

Shared allowance and durable ownership

One actual original allowance record supplies at most seventeen calls, 65,536 reply bytes and an exclusive absolute deadline. Every job and revision alias referring to its owner/store/record identity shares one cumulative balance. Changing that record's revision, digest or body conflicts; a new job ID does not reset it. A job Budget can only narrow spending and deadline. Unknown exposure keeps its whole reservation; invalid grammar, unsupported citation and unknown exposure retain the full charge. Only a valid canonical measured reply releases unused reserved bytes, exactly once. Calls are never refunded.

Cache ownership, shared allowance debit and original job dispatch are separate Journal commands. All three must be positively established before contact. A lost acknowledgement is unknown commit and is reconciled through the same deterministic command IDs. A retained dispatch is in progress until its original deadline, then unknown; it is never contacted again. SQLite and a collaborator are not one transaction. Cancellation is cooperative, and an in-process collaborator may retain independent effects or leave unobservable rejections beyond the shared observation helper's limits.

A triage first offer is owned by the original logical window, source identity and content hash. A synthesis first offer is owned by that window and the sorted accepted Signal identities. The original job durably pins its complete accepted-set synthesis offer before ownership/contact. Recovery resumes that offer; a currently unsupported source cannot turn an unknown offer into a fresh subset. Lookup precedes new collaborator parameters. The complete effective input retains objective, Guidance, values/policy/catalogue, exact quote and native ownership, original authorisation, collaborator ID/policy, reply limit and original checked time. Changed meaning or substitution under the same work conflicts with zero contact. Cached accepted, rejected, noise and unknown outcomes remain attributed to their original analysis owner.

One immutable root receipt per Product anchors the bounded commissioned window independently of caller-chosen commission and lineage names. A later lineage must name an authenticated retained predecessor; a null predecessor cannot start a renamed root. One bounded commissioned window is supported. A genuinely new lineage requires an authenticated new commission naming its predecessor and changed semantic subject; a renamed job or model is insufficient. Subsets, reordered or renamed old Signals cannot create new synthesis work. There is no rolling window, automatic renewal, lifetime duplicate index or caller cursor mutation.

Output and bounds

The result retains Signals, exact source observations, labelled author inferences or open questions, inert proposals, attempt exposure and a successor manifest. An exact compatible no-change window contacts neither collaborator. Missing triage leaves evidence untriaged; missing synthesis produces no proposal and parks accepted evidence. Source observations must reproduce a verified native quote exactly. Inferences and proposal usefulness still require independent semantic review.

Shadow assessment is optional and uses only an actual caller policy with explicit freshness, source, customer and new-evidence numbers. No lifecycle threshold is supplied by default. Competitor-only support is always parked. The only possible eligibility is eligible-validation-only; every result retains admission: "not-performed" and semanticReview: "required". No Slice or grant is created.

Whole overflow refuses. Request, effective input and result each fit 64 KiB, depth 32 and 4096 structural slots; four collections, sixteen selected citations/Signals/findings, eight proposals, 512-byte statements/rationale and 64 native evidence pages are the profile. Select one quote per source/content capture. Per-call reservation is at most 2048 triage bytes or 32768 synthesis bytes. The allowance retains at most 17 lifetime slots under 32768 bytes with worst-case settlement headroom, at most 36 mutations and 4096-byte events; no slots are evicted or reset. Job state fits 196608 bytes and cache state 131072 bytes. The caller's clock and cancellation are explicit.

These are technical reply/call limits, not token prices, money limits or guarantees about collaborator effects. Local deterministic proof does not establish live Research, model selection, semantic quality, release, customer benefit or feature admission.

Source: docs/agents/research-semantics.md