literature

name	literature
description	RigorQuant literature lane: walled, grad-student-style citation-graph traversal per research line, an independent literature adversary for validity + freshness, and the verified-negative membrane that keeps the novel lane un-anchored. Load for known/novel intake, deep literature review, or when the orchestrator needs 'what is settled / impossible / open / current'.

Literature lane — operating procedure

Owning decision: docs/architecture.md Decision 14, which holds the locked constraints (C1-C7), the membrane's crossing edges, and the named residual holes. This skill is the field procedure; the schemas and rq_check.py are the enforcement.

The membrane (non-negotiable)

Only VERIFIED NEGATIVES cross to the novel lane. A verified negative is a mathematically proven impossibility (category 'impossible'), independently re-retrieved by the lit-adversary, judged load-bearing for an active sub-problem. It is transmitted provenance-stripped ("closed path, not a premise and not a clue") — no author, title, or source id. 'open' and 'settled' never cross. If the study's conclusion would REST on a negative, escalate to the math lane before relying on it.

Step 0 — line hypotheses

The orchestrator forms 2–4 crude lines per sub-problem (seed queries, seed papers, sub-questions) from prior knowledge + one shallow web_search. No line is told it is favored.

Step 1 — walled line agents (parallel)

Launch subagent_lit_line per line, blank context, each in its own interim/lit// dir. Per line: resolve the seed (arxiv + academic-paper-search), read abstract → intro/related work → load-bearing theorem/method → references, then follow backward (references) and forward (citations via Semantic Scholar). Deduplicate by arXiv id / DOI / title; a revisit marks a hub, never a re-read. Write interim/lit//dossier.json (schema: schemas/dossier.schema.json). Never read another line's dossier. Batch independent web queries: on DSH >= 0.1.1-rc.1, one web_search call accepts a queries array (up to the configured searchMaxQueries, default 4) run in parallel and merged — batch rather than serialize; a single-query schema accepts one query per call.

Step 2 — literature adversary (independent)

For each line, send the lit-adversary (subagent_lit_adversary) the CLAIMS LIST only — never the dossier prose. It re-retrieves each load-bearing source itself and returns one verdict per claim: verified-current | verified-stale | unverifiable | false-claim, with source id, version, access date, retrieval method. It certifies "the literature says X and X is current", never "X is true".

Step 3 — synthesize

The orchestrator merges verdicts into literature/known-results.json (schema: schemas/known-results.schema.json), keyed by sub-problem id, each entry settled | impossible | superseded | open with sources. Also write, all four under literature/ and all four validator-enforced:

• negative-exports.json — the provenance-stripped constraints sent to the novel lane (schema: schemas/negative-exports.schema.json). Every export must trace to an impossible entry with a verified-current source, and that entry must carry negative_export: true: the map and the exports file record ONE fact.
• completeness.json — the per-line checklist (schema: schemas/completeness.schema.json), one line per swept research line. A sub-problem with a non-open record and no completeness line is refused.
• refs-seed.bib — the verified bibliography seed. Entries whose question is still open may not appear: the seed carries results, not reading lists.
• study.json literature.phase — set concluded only when the map holds verified state; not-run/running with verified records is refused. The sweep is mandatory at intake: the only way past it is phase: "skipped" with the user's verbatim assertion in skip_reason.

Transmission to the novel lane is by tool, not by tone: call subagent_novel (web, skill, and delegation denied in the composition) and pass the constraint text only. If a sub-problem's ANSWER is the impossibility, mark it status: "impossible" in study.json and record the math lane's acceptance in the entry's escalation path — the literature lane certifies that the literature says X, never that X is true.

Completeness checklist (anti-premature-termination)

A line may not conclude until the orchestrator records that it swept: forward citations; related-work sections of every hub; relevant surveys; adjacent fields named in the hypothesis; retractions (Crossref/Retraction Watch) and newer arXiv versions; author pages for load-bearing papers; and the frontier with a reason to stop (or none). "The model finished early" is a failing condition.

Termination and budget

Done when the adversary certifies every leaf is current, or a foundational sink / citation closure is reached. The budget (max_lines / max_depth / max_papers_per_line / max_rounds) is the default finish target, not a floor: a line concludes at the budget with the strongest completed dossier and its remaining completeness-checklist items recorded as open. Exceeding the budget requires an explicit user escalation recorded in study.json (literature.budget) — a silent overrun is a defect. Checkpoint map + frontiers + checklist after every round; a resumed session continues the walk. When the orchestrator cannot decide (new rounds / conclude / unclear), it is a HARD STOP: checkpoint and ask the human.

Retrieval order (executable)

Resolve the best retrievable version through the tiers below, in order, and record which tier won as the source's retrieval_method:

python3 <academic-paper-search skill dir>/scripts/resolve_tiers.py \
    --doi 10.2307/2975974 [--arxiv 1706.03762] [--title "..."] [--author-page URL]

It emits the ordered plan (author-page → open-repo → preprint → openalex → user-mirror). The mirror tier is EMPTY unless the user sets DSH_LIT_MIRRORS; no mirror endpoint is hardcoded in this repository, and the legal basis for using one is the user's to record. A paper reachable only as an abstract is unverifiable — a lower confidence tier, never verified-current, never load-bearing.

Provenance (validator-enforced)

rq_check.py refuses: a 'known' mark without a verified-current record; a sub-problem routed away as impossible without a verified impossible record AND its math-lane escalation; a declared escalation path that does not exist; a negative export that does not trace to an 'impossible' verified entry; a map whose negative_export flags disagree with the exports file; a missing completeness checklist, a missing or empty mandatory sweep, or a mapped sub-problem no line swept; a literature.phase that claims less than the record on disk; and any refs-seed.bib or PASS-time refs.bib entry that does not trace to a verified-current, non-open record. Dossiers in interim/lit/ are advisory — never counted as verified records, though a malformed one is still a defect.