research-arxiv-scout

Targeted audit 2026-09-11: selected source/API and evidence-contract corrections were verified; the frontmatter validation date and other source-registry dates are not a blanket September freshness claim. Recheck unverified source access and volatile facts when using them.

Use arXiv as a discovery layer for recent research, then produce repo-friendly outputs (ranked recommendations and 02_sources-*.json-style entries) without fabricating metrics.

Required attribution (arXiv API terms of use)

Include the following line in any output that uses arXiv data:

Thank you to arXiv for use of its open access interoperability.

Source: this exact sentence is arXiv's requested attribution statement, published on the arXiv API Documentation page (info.arxiv.org/help/api/index.html, linked from data/sources.json) — not on the Terms of Use page (tou.html), which covers rate limits and permitted use but does not itself contain an attribution clause. Checked 2026-07-11.

Quick Reference

Layered opportunity handoff

Use the opportunity evidence layers when this scan supports area discovery. Contribute L7 feasible delivery and L4 timing where a dated capability change is relevant. Return the original study ID, tested task, baseline, reproducibility, cost and transfer limits. HCI findings may inform L9 repeat-value hypotheses only for a matched job/population; do not transfer another product’s retention or willingness to pay to the new offer. Pure paper triage stays here without requiring commercial gates.

Carry source and underlying event IDs, dates, scope, supportive/mixed/adverse/unknown direction, evidence basis, counterevidence and the decisive unknown into the comparison worksheet. The same event appearing in several layers remains one event. No scout score, source count or convergence label passes a commercial gate; retain missing and adverse evidence in the handoff.

Workflow

ASCII Flow

arXiv scout request
  -> Map target to arXiv categories and keywords
  -> Query arXiv metadata with date-aware parameters
  -> Triage candidates for relevance, practicality, and evidence
  -> Verify abstract URLs, IDs, titles, authors, and any code links
  -> Produce ranked recommendations or sources-json entries
  -> Include required arXiv attribution and final checklist

1) Map target to categories and keywords

• If the target matches a key in config.yaml, use its categories, keywords, and time window.
• Otherwise, pick 1-4 arXiv categories and 3-8 keywords using references/category-taxonomy.md.

2) Query arXiv (metadata only)

• Use references/arxiv-api-guide.md to build queries for export.arxiv.org/api/query.
• Prefer sortBy=submittedDate and sortOrder=descending when scouting recent work.
• Start with max_results=50; paginate if needed.

Generate queries from config.yaml rather than hand-building them:

# List available config keys (skill → categories, window)
python3 scripts/generate_arxiv_scout_queries.py --list-skills

# Resolve categories, keywords, and time window from config.yaml
python3 scripts/generate_arxiv_scout_queries.py --skill ai-agents

# Ad-hoc topic when no config key fits
python3 scripts/generate_arxiv_scout_queries.py --topic "agent memory" \
    --categories cs.AI cs.CL --windows 30d 90d

The output includes estimated_min_runtime_seconds — at the enforced 1 req/3s this is the floor for executing the emitted queries serially. Do not parallelise to beat it; that is the fastest route to a 429.

Minimal query skeleton (if building by hand):

search_query=cat:cs.AI AND (agents OR "tool use")
sortBy=submittedDate&sortOrder=descending&start=0&max_results=50

2b) Pre-triage: social-signal cross-reference

Before scoring API results, cross-reference candidate arXiv IDs against:

• HF Papers (huggingface.co/papers) — daily community highlights and upvotes.
• alphaXiv (alphaxiv.org) — social layer on arXiv with comments and trending signals.

Flag any paper that appears on either platform as an attention signal for triage, never a quality or evidence upgrade. Cross-posts, coordinated launch traffic, and one source syndicating another are not independent corroboration. Also check each candidate's abstract for a GitHub or project-page link; a verified code link can improve Practicality, while Evidence changes only after the evaluation itself is inspected.

3) Triage and score

For each candidate paper, extract:

• Title, authors, arXiv ID, submitted date, categories
• Abstract-based relevance to the target
• Implementation signals (only if verified): code repository, dataset, benchmark, reproducibility notes

Do not include citation counts, GitHub stars, conference acceptance, or affiliations unless you can verify them.

Signal-vs-noise judgment calls (apply before scoring):

• Versioned re-announcements. A paper with v2/v3 in its history and an updated date inside your window but a published (first-submission) date months or years older is a revision, not new work — see references/arxiv-api-guide.md#version-deduplication. Sorting by lastUpdatedDate will surface these; always cross-check published before treating a hit as fresh.
• Citation-count traps on fresh papers. Even where citation counts are verifiable (Semantic Scholar/OpenAlex), a paper submitted in the last 1-3 months will have a near-zero count almost regardless of quality — citation lag, not merit, dominates at that age. Do not use "0 citations" as a negative signal on anything younger than ~6 months, and do not present a low count as meaningful evidence either way; state the paper's age instead.
• Preprint vs. peer-reviewed. arXiv is a preprint server: most results are not yet peer-reviewed, and some are rejected-and-resubmitted or duplicate-with-different-framing work. Absent a verified acceptance note (author's own "accepted at ..." line, checked against the actual venue), treat every candidate as unreviewed and say so in the report — do not imply venue prestige you have not confirmed.
• Hype-paper tells. Down-weight (do not auto-reject, but flag) candidates that combine: (a) marketing-register title/abstract language ("revolutionary", "unprecedented", state-of-the-art claims with no named baseline), (b) benchmarks limited to the authors' own curated dataset with no third-party eval, and (c) heavy same-day cross-posting to HF Papers / alphaXiv / X with upvote counts but no substantive technical discussion in the comments. Community attention (step 2b) is a volume signal, not a quality signal — read the abstract and, where available, the actual discussion thread before letting upvotes raise a score.
• Author/lab self-citation and PR-driven timing. A paper timed to a product launch or funding announcement from the same lab is not disqualifying, but note the coincidence in "Limits/risks" rather than silently treating it as independent validation.

Suggested scoring (0-10):

4) Produce deliverables

• Recommendation report: use assets/recommendation-output-template.md
• 02_sources-*.json entries: use assets/sources-json-template.md

5) Final checks (before handoff)

• Attribution line included (above)
• Time window stated (and matches what you searched)
• Categories and keywords listed (or referenced via config.yaml)
• All URLs point to abstract pages (https://arxiv.org/abs/...)
• arXiv IDs are exact (no typos, correct year/format)
• No duplicates in the final shortlist
• Paper titles/authors match the abstract page
• Any code links included were verified
• Any datasets/benchmarks referenced were verified
• Any conference/venue claims were verified (or omitted)
• Any impact metrics (citations/stars) were verified (or omitted)
• No unverified metrics or claims
• Suggested actions are concrete and repo-specific
• Target 02_sources-*.json schema preserved (no reshaping)

Killer-Feature Mode (HCI Retention Papers)

Specialized mode that contributes the hci_retention_paper signal to the bundle's Killer-Feature Convergence Protocol owned by research-review-mining.

Premise. HCI/CSCW/UIST work periodically publishes long-term studies that empirically tie a specific feature or interaction pattern to retention, freemium conversion, or willingness-to-pay. arXiv's cs.HC (plus cs.CY, cs.SI, cs.IR) indexes most of this work — though coverage of CHI/CSCW is incomplete and no representative percentage is established here, so the ACM-only fraction stays with research-scout's conference query generator.

When to use: bundle handoff from research-review-mining Killer-Feature Mode KF3 asks for the HCI signal; OR you have a candidate feature_id and want a rigorous attribution study.

Workflow delta:

KF-ARX-1. Map target to the config.yaml key `killer-feature-retention`
          (categories: cs.HC, cs.CY, cs.SI, cs.IR;
           keywords: long-term retention, feature adoption, freemium conversion, ...)
KF-ARX-2. Query arXiv per templates in references/product-retention-categories.md
KF-ARX-3. Triage with the standard 0-10 dimensions PLUS the
          "Attribution rigor" dimension defined in that reference
KF-ARX-4. Only papers with rigor>=7 graduate to a row on
          ../research-review-mining/assets/pay-trigger-ledger.tsv
          (signal_type=hci_retention_paper; wtp_strength=strong if causal,
           implicit if correlational only)
KF-ARX-5. Run ../research-review-mining/scripts/converge_killer_features.py

Expect 0-2 strong rows per scan — sparse but high credibility per hit. Zero rows is normal and does not downgrade other signals in the Convergence Rule.

References:

• references/product-retention-categories.md — categories, keyword groups, query templates, attribution-rigor scoring, anti-patterns
• ../research-review-mining/references/killer-feature-convergence.md — bundle Convergence Rule
• config.yaml — killer-feature-retention key

Navigation

Resources:

• data/sources.json — Official arXiv sources, complementary discovery tools, and the dead_or_changed inventory (arXiv 429 enforcement, OpenAlex key mandate, Semantic Scholar key policy, OpenReview v2, Connected Papers / ResearchRabbit tier changes)
• scripts/generate_arxiv_scout_queries.py — Query generator; resolves categories/keywords/window from config.yaml (--skill) or an ad-hoc --topic. Stdlib only
• references/arxiv-api-guide.md — API query patterns for export.arxiv.org/api/query
• references/category-taxonomy.md — Category selection and examples
• references/product-retention-categories.md — Killer-feature mode (HCI retention papers) categories + scoring
• config.yaml — Project mappings (skill → categories/keywords/time windows)
• assets/recommendation-output-template.md — Recommendation report template
• assets/sources-json-template.md — 02_sources-*.json entry template

Related skills:

• ../research-scout/SKILL.md — multi-source research idea mining (arXiv + HF Papers + Semantic Scholar + Papers with Code + conferences + industry blogs + curator newsletters). Use it when you want cross-source corroboration; use this skill when arXiv depth and category taxonomy are the priority.
• ../research-git/SKILL.md — git-history research for repo-internal evidence.

Fact-Checking

• Verify all arXiv IDs and abstract URLs (https://arxiv.org/abs/<id>) before output. Hallucinated IDs are the most common failure mode.
• Do not fabricate citation counts, GitHub stars, conference acceptance, or affiliations. Omit if not directly verified on the source page.
• Code links must be checked against the abstract page or repository, not assumed from the title.
• arXiv's API Documentation page asks every API consumer to include the attribution line "Thank you to arXiv for use of its open access interoperability." in any output that uses arXiv data (see "Required attribution" above for the precise source).
• For runtime-specific or volatile claims (dataset size, benchmark numbers, version-specific footguns), label them with the date pulled and prefer primary sources.
• If web access is unavailable, mark all paper-specific claims as unverified.

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.