081-Literature disposition is explicit and claim-grained

Type: ../types/adr.md · Status: accepted

Status: accepted Date: 2026-08-26

Context

Direct source grounding already separates a named source's retained evidence from the target's interpretation. It does not answer a different question: whether external literature makes a local artifact redundant, leaves an independent transfer or synthesis, or shows that the artifact was selected for assessment on a false premise.

The literature-disposition workshop exposed three recurring forces. Sources selected from a topic label miss the exact claims that decide overlap. A document-level overlap judgment spreads source warrant across claims the source does not establish. And inbound link count identifies an inspection and rewiring surface, not which claim consumers import. The completed cohort also supplied a negative control: one target did not make the alleged literature-repeating claim and was removed from the cohort without an artifact edit. A reusable procedure can therefore be adopted without encoding the cohort's predominately keep-shaped artifact outcomes as a prior.

A bounded Pirolli rerun adds a procedural warning. After source reconstruction and target-claim judgment were separated, three of four support verdicts became narrower. The rerun also changed prompts, roles, and intermediate representations, and it did not test false narrowing. It makes bilateral isolation a useful diagnostic control, not an established default.

The proposal Operational destinations for literature-grounding findings left one separate choice open: whether missing-prior-art discovery should run during every ordinary write. Such a search has no natural completeness boundary, and model recall or search rank can nominate a source without warranting a novelty or subsumption verdict.

Decision

  1. External-literature disposition is an explicit multistage branch. A standalone instruction governs source selection and disposition when the request explicitly asks whether literature duplicates or subsumes a claim-bearing artifact, or asks for a keep, rewrite, thin, merge, retire, or cohort decision on that basis. cp-skill-write-multistage loads the instruction conditionally and supplies reconstruction, audit, promotion, and validation. Ordinary grounding, synthesis, and revision do not load it.
  2. The decision unit is a live claim region. The procedure inventories the target before selecting sources, admits sources against exact propositions, separates source-established content from the local transfer, synthesis, application, counterexample, or boundary, and attempts the smallest faithful replacement. A merely verbal remainder does not preserve an artifact. The available outcomes include cohort removal, keep, rewrite, thin, merge, and retire; none is the default.
  3. Bilateral isolation is a conditional diagnostic control. Use separate fresh target, source, and comparison contexts when the user requests an independence control, a previous comparison is challenged for cross-side contamination, or the task prospectively evaluates the method. Freeze the two independent inventories before comparison and give the comparator only those outputs. The target-blind source worker invokes cp-skill-ground itself; the skill's ordinary context: fork from a target-bearing caller does not establish isolation. Do not make the control routine until a prospective matched test measures both over-attribution and false narrowing.
  4. Inbound links are impact candidates, not warrant. Each inbound use is classified by whether it imports the affected claim, another claim, or nothing material. Raw count may price inspection and rewiring but cannot decide semantic propagation or artifact value.
  5. Candidacy and verdict authority stay separate. A user-named source, model recollection, search result, or citation graph may create a reading assignment. For each admitted source-side claim, the literature procedure invokes cp-skill-ground rather than reproducing its ingest, Quotes, or snapshot decisions. A user-supplied canonical URL may take that skill's missing-ingest path. An agent-nominated untracked URL requires user approval before it is passed to the skill. The target cannot change until direct source evidence is read against the exact claim through the resulting Quotes-or-pinned-snapshot path.
  6. Ordinary writing does not search for missing external prior art. Its intra-KB near-duplicate check and named-source dependency guard remain bounded as designed. cp-skill-write routes an explicit literature disposition request to the multistage branch without asking for a second confirmation. No absence of candidates within a bounded assessment certifies global novelty.
  7. No new assay, validator, command, or propagation evidence note is added. A validator cannot discover an unnamed source. A report-kind assay lacks a chosen trigger and search budget and would still carry only candidacy authority. The observed propagation trace motivates semantic inbound classification, but this decision does not depend on its numeric rate, so git remains its audit path.

The user-supplied or separately authorized URL path in decision 5 narrowly amends ADR 078. Its ordinary-writer boundary remains: a generic named-source dependency with no tracked ingest stops for a separate ingest run. The exception applies only inside an explicit literature-disposition assessment, where source acquisition is itself part of the authorized operation.

Considered alternatives

Run external prior-art discovery during every ordinary write. This would place source discovery near the point of authorship. Rejected because a universal search has no justified stopping rule, would turn a bounded writer into an open-ended survey, and could make an empty result look like a novelty certificate. Explicit routing captures the cases where the operator actually needs the comparison.

Add a deterministic provenance check or a report-kind prior-art assay. A deterministic check can validate known provenance but cannot find an unnamed source; named dependencies already have a guard. An open-ended assay could nominate sources, but its trigger, budget, and candidate-consumption path are unsettled, and its model output cannot decide the verdict. Both remain possible future additions if those missing choices acquire evidence.

Put the full procedure inside the multistage skill. This would guarantee a loading path but enlarge every multistage read with a specialised retrospective branch. Rejected in favour of a standalone instruction that the skill loads only on the explicit trigger. The instruction is shipped with the framework in installed projects, so the conditional route is not checkout-local.

Make bilateral isolation the default comparison shape. This would prevent source wording from shaping the target inventory and target wording from shaping source reconstruction in every assessment. Rejected because one confounded four-claim rerun does not estimate the control's benefit, cost, or false-narrowing risk. The adopted diagnostic triggers preserve the option while a prospective matched test remains open.

Retain an omnibus methodology note or the producing workshop. Either would preserve the reasoning but give the procedure no binding consumer. Rejected: the operational rules belong in an instruction, this architectural selection belongs in an ADR, and git retains the change narrative and unused measurement.

Generalize the cohort's keep-shaped results. Rejected because an outcome frequency does not warrant an outcome prior. The membership failure supplies the required negative control, and the adopted procedure requires an adversarial replacement attempt before any keep.

Free choices left open: whether a later prior-art assay is useful as a bounded candidate generator; which artifact classes, queries, and budgets would trigger it; whether a prospective matched test justifies making bilateral isolation routine; and whether repeated propagation measurements eventually warrant a durable evidence artifact.

Consequences

An explicit literature-disposition request now reaches one executable method: claim inventory precedes source selection, source evidence stays scoped to the claim it supports, a local remainder must survive an adversarial replacement, and inbound users are inspected by imported claim. Keep and cohort-removal results can complete without touching the live artifact; merges and retirements retain their existing user-approval boundaries.

An independence challenge has a defined stronger route rather than an informal rerun: source-blind target inventory, target-blind source reconstruction, then comparison of frozen outputs. That route costs two additional clean contexts and can miss a distributed relation present in the live inputs, so it remains a diagnostic until prospective evidence measures both error directions.

The method is deliberately expensive when invoked. It may require source capture, grounding, claim-by-claim comparison, and semantic inspection of inbound links. Ordinary writing does not pay that cost, but it can still rediscover known external ideas when no one asks the prior-art question. The system makes no novelty claim from that silence.

Operativity path: harness selection loads the promoted cp-skill-write or cp-skill-write-multistage body. The ordinary writer routes an explicit request; the multistage skill then reads assess-a-claim-bearing-artifact-against-external-literature.md from the source-checkout or installed-framework instruction tree and treats it as a binding decision contract. That instruction invokes cp-skill-ground for each admitted source-side proposition; grounding may invoke cp-skill-ingest for an authorized canonical URL or return the conditional re-ingest route. The multistage promotion and validation stages carry any selected edit into the library. No validator enforces the semantic disposition.

Where the decision stops applying: it covers explicit retrospective comparison of claim-bearing artifacts against external literature. It does not require continuous novelty surveillance, guarantee exhaustive literature coverage, or grant source suggestions verdict authority. A future bounded discovery path can extend the candidate surface without reopening the claim-grained verdict procedure, provided it preserves that authority boundary.

Relevant Notes: