Novelty Coverage

A read-only, search-grounded, per-novelty-claim collision detector for a doc-shaped artifact (a design paper, a research doc, a spec). It opens the OUTCOME axis's novelty surface that neither cross-source-review (csr), primary-source-verification (psv), nor blueprint-crafting's researcher reach: csr converges a doc's PROCESS axis but does not search for prior art; psv FETCHES each cited source (backward, cited) but does not hunt uncited prior art; the researcher GATHERS sources forward for authoring (the opposite direction). prior-art-search is the missing backward, uncited-prior-art collision check — it SEARCHES the prior-art corpus for each novelty claim and reports collisions honestly.

Core Positioning

Search-grounded, not recall-based

A collision / uncited-relevant finding MUST cite a fetched QUOTE from found prior art, not model recall. A finding that cannot ground a fetched quote of the found prior-art text is downgraded to inconclusive, never asserted. This is the load-bearing invariant (workspace rule 3 + L1), enforced by a self-gate (NC-I5). The found text is itself model-extracted from search results — a weaker oracle than psv's fetched authoritative source (see the two-layer caveat below).

Per-claim collision coverage, not a novelty verdict

prior-art-search NEVER emits novel_confirmed or "the doc is novel." Its sole top-line signal is collisions_under_known_coverage: C clear-under-search, N collisions, U uncited-relevant, I inconclusive of M extracted novelty claims. I claims (and the M=0 case — no novelty surface) ESCALATE to a human. This is an honest coverage disclosure, NOT a novelty verdict. The irreducible absence-of-evidence limit: you cannot prove a negative ("no prior art exists"), so a doc with zero found collisions is "no collision found in the searched corpus," NOT "novel." A symmetric "novelty-convergence" skill that loops to a novel_confirmed boolean is the category error — it flattens a COVERAGE disclosure ("no collision found in the searched corpus") into a VERDICT ("novel"), i.e., absence-of-evidence treated as evidence-of-absence; prior-art-search is explicitly NOT that. (The broader two-axis frame — process × outcome — is the spec-gaming paper §4.2; prior-art-search lives entirely on the outcome axis, so its error is outcome-internal, not a process/outcome conflation.)

Additive to csr and psv; a single pipeline

prior-art-search is ADDITIVE to csr (process axis) and psv (cited-source verification); it does NOT merge with either. Unlike csr's same-family + different-family multi-round debate, prior-art-search is a SINGLE pipeline — extract novelty claims → search prior art → collision verdict per claim → coverage disclosure. The oracle is the searchable prior-art corpus, not a different-family model, so it does NOT port csr's heterogeneous substrate. psv and prior-art-search compose: a high-stakes doc may run psv (are the cited sources accurate?) AND prior-art-search (are the novelty claims collision-free against findable prior art?). They do not merge.

Two-layer oracle weakness, stated honestly (workspace rule 3)

prior-art-search's oracle is the searchable prior-art corpus, WEAKER than psv's fetched primary source at TWO layers (the load-bearing honesty caveat, proposal §3 / ADR §8 Q4):

1. Comparison-side weakness (shared with psv). The claim-vs-prior-art comparison is model-judged — clear-under-search means the model found no collision, not that none exists. This is psv's tier-2 partial-under-comparison weakness.
2. Selection-side weakness (prior-art-search's own). A search result adds a selection weakness on top of the comparison weakness: ranking bias, recall limits (you cannot search the whole corpus), and the found prior-art text is itself model-extracted from the search return, not a fetched authoritative byte-range.

So prior-art-search is weaker at two layers, not "weak oracle vs strong oracle." This is precisely WHY it NEVER emits novel_confirmed and why its top-line is a coverage statement, not a verdict. clear-under-search means "no collision in what was searched," not "novel." The weaker oracle is the irreducible limit of novelty-checking short of a human.

Two blind-spots, stated honestly (workspace rule 3)

prior-art-search's own model-performed steps carry blind spots disclosed rather than hidden: (1) novelty-claim extraction inherits the extractor's blind spots — claims neither author nor extractor identify as "novelty" are absent from M; (2) the claim-vs-prior-art comparison is model-judged — clear-under-search means the model found no collision, not that none exists. The C/N/U/I/M counts are conditional on what was extracted, searched, and compared, not a completeness guarantee. A different-family extraction leg is a deferred mitigation (partial).

Scope

• In: extract atomic novelty claims from a doc (each: location + search target); search the prior-art corpus per claim; emit a per-claim collision verdict (collision/uncited-relevant/clear-under-search/inconclusive) grounded in a fetched quote of found prior art for collisions; emit a collision-record (collisions_under_known_coverage) + a collision-findings packet.
• Out: code review (parallel-development); spec/arch-design authoring (blueprint-crafting); doc process-convergence (cross-source-review); CITED-source verification (primary-source-verification); research gathering (blueprint-crafting researcher); a novelty verdict — whether the doc's novelty is REAL or significant — (human).

Scope Guard (entry-time detection)

• Deliverable is source/test code → route to parallel-development.
• Deliverable is authoring/rewriting a spec, arch-design, iteration-plan, or research → route to blueprint-crafting.
• The request is "drive adversarial multi-round review to convergence on this doc" → route to cross-source-review (process axis).
• The request is "do this doc's CITED claims hold against their sources?" → route to primary-source-verification (psv = backward-CITED; prior-art-search = backward-UNCITED).
• The request is "does this doc OVERCLAIM novelty — collide with prior art it failed to cite?" → prior-art-search (this skill).
• The request is "is this novel / significant / the right contribution?" → outcome-axis, human only.

The guard is soft (remind + route, not refuse). Phase A relies on explicit invocation (/skill:prior-art-search).

Quick Start

You are the orchestrator. The legs are read-only; the driver runs the single pipeline (NOT a multi-round debate). Run in order.

1. Frame — identify the artifact + its novelty surface. Enumerate the core novelty claims (the load-bearing "X is new / first / has no prior art" assertions to collision-check). Declare scope.
2. Novelty-claim extraction (NC-I2) — invoke the subagent tool with agent="solidforge:novelty-claim-extractor" (fresh context); it reads the artifact and enumerates atomic novelty claims, each tied to a location and a search target. Atomic = one novelty proposition per claim (compound split). A novelty claim is an assertion of newness ("X is a new framing," "Y is the first to," "Z has no prior art"). Distinguish from factual/citation claims (psv's domain — verify, don't collision-check) and interpretive claims ("this is significant" — NOT search-adjudicable; escalate to human, tagged inadmissible). (Extractor blind-spot caveat: claims neither party identifies as "novelty" are absent from M.)
3. Multi-source search + collision verdict (NC-I3) — for each novelty claim, search the prior-art corpus (arXiv, web, public repos) for work that ALREADY made the claim, then compare the claim against found prior art and assign one of:
   • collision — found prior art already makes this claim (the doc's novelty is overstated); cite the colliding source + QUOTE the found prior-art text (severity blocker; evidence MUST quote the found prior art).
   • uncited-relevant — found relevant work the doc didn't cite (not a direct collision, but a coverage gap); QUOTE the found prior art (severity warning).
   • clear-under-search — no collision found IN THE SEARCHED CORPUS (NOT novel; just no collision surfaced). Counted only, not a finding.
   • inconclusive — search couldn't cover the claim (paywalled corpus, ambiguous query, nothing findable); escalate to human (severity coverage). Open sources (arXiv abstracts, public repos) are searched without credential; web-search APIs use a separately-named HTTP surface (NOT csr's _ANTHROPIC_AUTH_TOKEN LLM-token namespace — see install.md). (Comparison blind-spot caveat + the selection-side weakness.)
4. Coverage disclosure (NC-I4) — emit a collision-record: collisions_under_known_coverage = C clear-under-search / N collisions / U uncited-relevant / I inconclusive of M = C+N+U+I extracted novelty claims; plus a collision-findings packet (collision / uncited-relevant / inconclusive findings; clear-under-search claims are counted, not listed). M=0 (no extractable novelty claims — a doc making no novelty assertion, or all novelty claims interpretive) and I>0 both ESCALATE to a human. NEVER novel_confirmed. A collision/uncited-relevant finding without a fetched quote is downgraded to inconclusive.

The honest signal is the coverage disclosure — exactly what was searched and what collided, and what could not be covered. Whether the doc is genuinely NOVEL stays human.

Authority Chain

• docs/proposal.md = authoritative design (master; CSR process-axis SUBSTANTIVE-CONVERGED + psv N=9/K=0 verified, 2026-07-31).
• docs/iteration-plan.md = execution blueprint (Phase A work-items NC-I0–NC-I6; on conflict, proposal.md wins).
• docs/design-decisions.md = ADR log (the non-obvious decisions: never-novel_confirmed, single-pipeline-not-debate, the false-positive structural-proxy, the two-layer search-oracle weakness, the relation to psv/researcher, the credential surface).
• docs/proposal.convergence.md = the proposal's cross-review trail.

Coordination with cross-source-review / primary-source-verification / blueprint-crafting / parallel-development

• cross-source-review stays the PROCESS-axis convergence engine. prior-art-search is ADDITIVE and non-merging: csr's substantive_converged does not imply prior-art-search's coverage, and vice versa.
• primary-source-verification is the complementary outcome-axis leg. psv = backward-CITED (fetch each cited source); prior-art-search = backward-UNCITED (search for prior art the doc did NOT cite). They compose (a doc may run both); they do not merge. prior-art-search reuses psv's schema-field pattern (collision-findings extends psv's doc-findings — a strict superset) and the install.md credential REASONING — copy-patterns, NOT code (workspace rule 7).
• blueprint-crafting stays pure (process-axis). Its researcher is FORWARD gather ("find sources about X for authoring") — the opposite direction from prior-art-search's backward collision. They do not merge.
• parallel-development owns code review. prior-art-search is doc-shaped, not code-shaped.
• files_touched boundary: this skill owns prior-art-search/ + the repo-root agents/novelty-claim-extractor.agent.md (mirroring psv's solidforge:claim-extractor). It does not modify csr, psv, bc, or pd core logic (prior-art-search is additive). Terminology (workspace rule 10): collisions_under_known_coverage is the one top-line term; novel_confirmed is the forbidden term and never appears as an emitted field.

Self-Checks (Definition of Done)

Maturity note (Phase A build): all self-gates below ship and pass green (NC-I0–NC-I6 converged; docs/iteration-plan.run-record.json carries the verdict). The live N≥3 dogfood ran 2026-08-01 on 3 real artifacts incl. the long spec-gaming paper — coverage profile at dogfood-run-2026-08-01.md: 2 uncited-relevant findings surfaced via fetched prior art (spec-gaming NC34 ↔ SWE-ABS; psv NC2 ↔ AI-Powered Citation Auditing), and the known self-cert-paradox collision came back clear-under-search under the arXiv-only oracle — demonstrating the two-layer weakness in production (a zero-collision result is "no collision in what was searched," NOT "novel"). The offline dogfood.py gate covers the canonical fixture only (rule-1 skip path when no network). Remaining deferral: a different-family extraction leg (partial mitigation of the extractor blind-spot). The name prior-art-search is PROPOSED, pending human LOCK (proposal §8 Q1).

A skill change is not done while any self-check fails (workspace rule 1). Run before commit:

python3 skills/prior-art-search/infra/test/disconnect_check.py        # structure + loading-chain
python3 skills/prior-art-search/infra/test/plugin_layout.py           # plugin.json + agents well-formed
python3 skills/prior-art-search/infra/test/findings_shape_check.py    # every emit path → collision-findings + collision-record valid
python3 skills/prior-art-search/infra/test/coverage_policy_check.py   # counts sum M=C+N+U+I; M=0/I>0 escalate; no novel_confirmed; fetched-quote invariant
python3 skills/prior-art-search/infra/test/fetched_quote_gate.py      # every collision/uncited-relevant finding cites fetched prior-art text; quote-less → inconclusive (the L1/rule-3 invariant)
python3 skills/prior-art-search/infra/test/lint_self.py               # dogfood: lints this skill's own infra (ruff)
python3 skills/prior-art-search/infra/test/dogfood.py                 # runs the pipeline on a fixture doc with a planted collision (skips gracefully w/o network; recorded log substitutes)

Reference Files

• proposal.md — authoritative design (problem, non-goals, owned contract, layering, §8 decisions).
• iteration-plan.md — Phase-A execution blueprint (NC-I0–NC-I6, DoD, DAG, risks).
• design-decisions.md — ADR log (the non-obvious decisions).
• install.md — provisioning: prior-art-search's credential surface (open prior art searched without credential; web-search API = separately-named HTTP surface, NOT csr's LLM-token namespace) + the two-layer oracle weakness + SSRF posture. prior-art-search is env-armed (no arm command — its gates are self-gates).