Assess a claim-bearing artifact against external literature

Type: kb/types/instruction.md

Decide the artifact's smallest faithful disposition from direct source evidence without mistaking topical similarity, model recall, or backlink count for a verdict.

Use this procedure only for an explicit retrospective literature-disposition question. Ordinary writing, a named-source grounding check, and a general request to improve an artifact do not trigger it. This procedure does not certify novelty: a bounded search that finds no source means only that no candidate was found within the recorded boundary.

Run the procedure inside cp-skill-write-multistage. Use that workflow's workshop for the records below and its reconstruction, audit, source-dependency guard, promotion, and validation stages for any live edit.

Composition contract

cp-skill-write routes an explicit literature-disposition request to cp-skill-write-multistage, which loads this instruction. This instruction owns source candidacy, source-versus-target comparison, and artifact disposition. For every admitted source-side claim it invokes cp-skill-ground, which owns ingest resolution, retained Quotes, the snapshot-required route, any call to cp-skill-ingest, and conditional routing to re-ingest.md. Do not duplicate or bypass that grounding protocol here.

These artifacts are composition siblings. Before changing their trigger, inputs, results, mutation authority, or stop conditions, inspect the other direct callers and callees under the instruction collection's edit-time rule.

Steps

  1. Fix the assessment boundary. Record the target, the alleged externally established subject, the intended consumer, and a finite source-candidacy boundary. Prefer a user-supplied corpus or source. Otherwise record the exact queries, venues, dates, or result limit used to look for the strongest plausible host. Ask the user only when materially different reasonable boundaries could change the requested disposition.

  2. Inventory the live claims before selecting sources. Divide the target into claim regions at the finest grain at which a consumer could rely on one commitment without importing another. Include the title claim, load-bearing premises, mechanisms, scope boundaries, and operational consequences. For each alleged literature overlap, first decide whether the target actually makes that claim. If it does not, record cohort removal and do not edit the artifact merely to fit the assessment premise.

Use bilateral isolation only when the user explicitly requests an independence control, a prior comparison is challenged because one side may have shaped the representation of the other, or the task is a prospective evaluation of the comparison method. When it applies, add target-claim-inventory.md, source-reconstruction.md, and isolated-comparison.md to the multistage workshop checklist. Launch a fresh target worker now to write target-claim-inventory.md from the target and its contracts without seeing the source, source reconstruction, or prior comparison. Do not make bilateral isolation the default from the existing single bounded case; it changed verdicts but did not isolate which procedural difference caused the change.

  1. Nominate sources against exact claims. For each in-scope claim region, record the proposition a source would need to establish, contradict, or bound. A user-named source, model recall, search result, citation graph, or topical resemblance may nominate a reading assignment; none supplies verdict evidence. Admit an exact tracked ingest immediately. A canonical URL explicitly supplied by the user as an assessment source is also admitted: that input authorizes cp-skill-ground to resolve or create its ingest. If the agent nominates an untracked URL, record it as a reading assignment and obtain user approval before admitting it, because passing it to cp-skill-ground can create a source artifact. Do not invoke cp-skill-ingest directly and do not change the target from a snippet, paraphrase, or remembered source.

  2. Ground each admitted source use through cp-skill-ground. For the ordinary branch, invoke it once per source-side proposition with Target: <exact ingest path or authorized canonical URL> and Claim needed: <source-side proposition or question>. Do not include target prose or a target-specific transfer argument. The grounding skill owns ingest resolution, any permitted ingest creation, Quotes sufficiency and append, and the snapshot-required route; do not reproduce those decisions here.

When bilateral isolation applies, the fresh target-blind source worker must be the grounding caller. Give it only the admitted source identifiers, exact source-side questions, source collection and ingest contracts, and an instruction to invoke cp-skill-ground; do not give it target prose, the target inventory, or a prior comparison. Its grounding fork then inherits a source-only conversation. Have the worker write source-reconstruction.md from the returned direct evidence and its stated limits.

Consume its result as follows:

  • quotes sufficient or quotes added — read the returned ingest's complete ## Quotes section and use only its retained verbatim extracts as source support. Record every quotes added side effect with the ingest path.
  • snapshot required — use the verified name-paired snapshot and apply semantic/grounding-alignment; preserve the exact marker on any resulting ingest link.
  • any stop or blocker — do not assign a disposition that relies on the unresolved source-side proposition.

Record the source-established unit and its limits separately from any Commonplace transfer, synthesis, application, counterexample, or boundary.

When bilateral isolation applies, launch a third worker with only the two frozen outputs, the intended consumer, and the assessment question. Have it write isolated-comparison.md, testing both charitable over-attribution and false narrowing. It marks missing information instead of opening either live input. Use that comparison as the source-versus-target input to the remaining steps.

  1. Build a claim-region disposition table. Give every live claim region one row with:

  2. the incumbent claim and scope;

  3. the direct source evidence and exactly what it licenses;
  4. the target-local remainder, if any;
  5. the intended consumer's warranted inference or action;
  6. the strongest existing local artifact that could carry the remainder;
  7. the smallest faithful replacement path; and
  8. the provisional disposition.

Keep source-established and target-local units separate. Evidence for one claim, conjunction, or scope does not spread across the document.

  1. Try to eliminate the incumbent. Construct the strongest replacement from the admitted source plus existing local artifacts. A replacement is faithful only when it gives the intended consumer the same warranted inference or operation over the same scope. Keep a local remainder only when it changes that use and cannot be recovered from the replacement path. A verbal difference, domain renaming, ungrounded transfer, or subject narrowed until the predicate is true does not earn a separate artifact.

  2. Inspect inbound uses semantically. Find tracked inbound links with one backlinks query. Classify each as imports affected claim, imports another claim, or incidental. Use the first class to test whether the proposed replacement preserves real imports. Use total link count only to estimate rewiring work; it neither warrants the incumbent nor measures the inherited claim's blast radius. No inbound links do not by themselves justify retirement.

  3. Choose the outcome from the rows, without an outcome prior. Use these meanings:

  4. cohort removal — the target does not make the claim that admitted it to the assessment; leave the artifact unchanged and correct the cohort record;

  5. keep — an independent contribution survives the replacement test and the incumbent already states and supports it accurately;
  6. rewrite — an independent contribution survives, but its wording, scope, attribution, evidence, or separation from source-established material must change;
  7. thin — source-restating regions can be removed while a smaller independent, citable contribution remains;
  8. merge — an existing local artifact can carry the surviving contribution with a cheaper faithful path;
  9. retire — direct sources and existing local artifacts replace every warranted use, leaving no independent contribution.

A mixed target may receive different region-level outcomes, but the final artifact action must preserve one coherent contribution or type-appropriate purpose. For a merge that changes another artifact, use the multistage workflow's user-decision gate. For retirement, read and follow Retire an artifact, including its approval stop after the inbound inventory. Keep relocations pure when an outcome changes only the artifact's address.

  1. Execute and verify. Record the final disposition, evidence paths, search boundary, unresolved coverage, and inbound classifications in the workshop. Use the remaining multistage steps for an authorized rewrite, thin, or merge. Validate every changed artifact. A keep or cohort removal produces no target edit. After a changed artifact validates, suggest cp-skill-connect for broader graph discovery. Never report the result as proof that no prior art exists outside the recorded boundary.

Verify

  • Every admitted source was selected against an exact live claim and passed through cp-skill-ground from a tracked ingest or authorized canonical URL.
  • Every claim region has a source-established unit, a local remainder, a smallest faithful replacement, and a disposition.
  • The replacement test could have produced a non-keep outcome; verbal distinctness did not decide it.
  • Inbound links were classified by imported claim before they affected impact or rewiring decisions.
  • Candidate-generation evidence never decided overlap, novelty, or disposition.
  • The result states its search boundary and does not certify global novelty.

Relevant Notes: