---
name: strict-research-literature-evidence-gate
description: "Use when building a research reading list or evidence base: separate claim support, novelty collisions, protocol sources, frontier watch, and discard so strict quality filtering does not hide exact prior art."
metadata:
  id: "strict-research-literature-evidence-gate"
  type: "research"
  title: "Gate research literature by evidence role"
  summary_zh: "文献先按证据用途分轨:主张证据只留顶会顶刊/奠基作,但 exact novelty collision 即使弱也不能漏"
  tags: "research,paper-reading,quality-gate"
  created: "2026-07-15"
  last_verified: "2026-07-15"
  status: "valid"
  context_cost: "medium"
  produces: "literature_evidence_lanes"
  verify_command: "For every candidate source, fill six fields: decision-impact, evidence-role, formal-identity, artifact, independence/conflict, and final-lane."
  verify_expected: "Claim-support sources are top peer-reviewed venue/journal papers or necessary field foundations; exact preprint collisions remain visible only in novelty-collision/watch; first-party lab material is explicitly bounded and never treated as independent corroboration."
  verify_failure_next: "Downgrade sources whose venue, method, or artifact cannot be verified to frontier-watch or discard; retain metadata for an exact collision but remove it from claim-support."
---

## Trigger

- Building or cleaning a literature map, related-work section, reading queue,
  benchmark rationale, or paper-facing evidence base.
- Search results mix peer-reviewed papers, arXiv preprints, workshops, vendor
  blogs, system cards, technical reports, and weakly related papers.
- A strict top-venue filter is desirable, but it may hide an exact prior-art
  collision and create a false novelty claim.

## Fix

Do not rank all sources on one keep/drop scale. First assign each source one
evidence role; the role determines what the source is allowed to prove.

| Lane | Admission rule | Allowed use | Hard boundary |
| --- | --- | --- | --- |
| Claim support | Directly relevant full paper at a leading peer-reviewed venue/journal, or a necessary field-defining foundation | Support scientific claims, comparisons, and the main narrative | Venue prestige cannot rescue a method or experiment that does not support the cited claim |
| Novelty collision | Exact or near-exact overlap with the proposed contribution, regardless of venue status | Test novelty and delimit the contribution | Keep weak/unreviewed work as metadata; do not promote it into scientific evidence |
| Protocol source | Original benchmark, dataset, method, evaluation specification, or authoritative implementation with an inspectable artifact | Define the protocol and reproduce it faithfully | Supports the protocol it specifies, not unrelated general conclusions |
| Frontier watch | Highly relevant recent preprint or official lab research paper, system card, or evaluation report | Track fast-moving results and form hypotheses | Label unreviewed or first-party evidence; seek independent confirmation before claim support |
| Discard | No decision impact, weak relation, unverifiable identity, redundant contribution, or suspicious evidence chain | None | Do not download or create a standalone note merely to make the search look comprehensive |

Apply this gate before deep reading or bulk PDF download:

1. **Decision impact** — state which decision changes if the source is true:
   claim, novelty, protocol, or frontier monitoring. No changed decision means
   discard.
2. **Formal identity** — verify title, authors, year, and publication status
   against official proceedings/journal pages and durable scholarly records
   such as DOI, DBLP, ACL Anthology, or an accepted OpenReview record. arXiv is
   a distribution channel, not proof of venue acceptance.
3. **Claim fit** — inspect the actual method, setting, baselines, and results.
   Topic similarity and citation count do not show that a paper supports the
   sentence being written.
4. **Artifact check** — for empirical or protocol-defining work, look for the
   official code, data, benchmark, evaluation script, or sufficiently precise
   specification. Missing artifacts reduce reproducibility and may force a
   downgrade; they are not an automatic rejection for theory or historical
   foundations.
5. **Independence/conflict** — distinguish independent corroboration from
   repeated reports by the same authors, lab, model vendor, benchmark owner,
   or shared evaluation stack. Record the conflict instead of silently
   counting dependent sources as separate confirmation.
6. **Final lane before storage** — only claim-support and required
   protocol-source papers enter the core reading queue by default. Keep
   novelty collisions and frontier-watch items metadata-first; download them
   only when their details can change a decision.

Use a strict but explicit exception for foundational work: retain an older or
non-current-venue source when it originated a method, benchmark, or concept
that the present argument must trace. Durable adoption and lineage justify the
exception; raw citation count alone does not.

For frontier AI lab material:

- Product documentation, release notes, and blogs support product facts only.
- Official research papers, system cards, and evaluation reports may enter
  protocol-source or frontier-watch when they expose methods or measurements.
- Treat them as first-party evidence. Claims of safety, efficacy, superiority,
  or general scientific validity still need independent peer-reviewed evidence
  or a faithful external reproduction.

Minimal audit row:

| Field | Required answer |
| --- | --- |
| Decision impact | What research decision can this source change? |
| Evidence role | Claim support / novelty collision / protocol source / frontier watch / discard |
| Formal identity | Official venue/status record, or explicitly unreviewed |
| Artifact | Official code/data/protocol link, not found, or not applicable |
| Independence/conflict | Independent, first-party, shared authors, shared benchmark, or other dependency |
| Final lane | The assigned lane and one-sentence reason |

## Reuse Rule

- Load before literature discovery, reading-list cleanup, paper-facing claim
  construction, or novelty review.
- After admission, use paper-reading-three-pass to decide how deeply to read;
  use novelty-is-the-delta-not-the-territory for exact-delta judgment; use
  research-metrics-from-current-sota when selecting evaluation metrics.
- Re-run the gate when publication status, artifacts, independent replications,
  or the research claim changes. A source may move lanes; its evidence role is
  not permanent.