Research

Research

One-shot research with source grading. For autonomous cycles, knowledge compilation, training-data diff, or parallel dispatch, use /research-ops.

Query

Current Environment

!echo "Date: $(date +%Y-%m-%d) | Project: $(basename $PWD) | MCP servers: $(claude mcp list 2>/dev/null | wc -l | tr -d ' ')"

Research with the rigor of an investigative journalist, not a search engine. Every claim needs provenance. Inference is fine — but say it's inference, not fact.

This skill produces a research memo, not code. If findings imply something should be built, write the memo with your findings, then offer a plan. Do not pivot to implementation within a researcher session.

Plan persistence: If research produces an actionable plan, write it to .claude/plans/ or docs/ — not just in-context. In-context plans are lost on compaction (285 lost update_plan calls observed in Codex sessions).

Session awareness: !cat ~/.claude/active-agents.json 2>/dev/null | python3 -c "import sys,json,time; entries=json.load(sys.stdin); active=[e for e in entries if time.time()-e.get('started_at',0)<7200]; print(f'{len(active)} active sessions') if len(active)>=3 else None" 2>/dev/null If 3+ sessions active: keep questions shorter, batch ambiguous items.

Companion refs: references/epistemics for bio/medical/scientific claims. references/source-grading for investigation/OSINT (Admiralty grades). eval when findings imply a model/engine/config/quality comparison should be measured — research surfaces the question and the methodology literature; /eval designs the benchmark (pre-registration, power, blind judges). Hand off; don't run the benchmark inside a research session. Project routing: Check .claude/rules/research-depth.md if it exists. Domain gotchas: Read ${CLAUDE_SKILL_DIR}/DOMAINS.md when the domain applies. No arguments? If invoked without a question, check schemas/open_questions.md or .claude/rules/research-depth.md for open research items, or ask the user.

Tool Routing

Default web backend: Exa, then Brave (operator pref 2026-06-09). Perplexity is fallback-only (quota-flaky, NOT an independent index) — reach for it only for the biomedical exact-number niche below. Exa + Brave are the durable independently-indexed pair (but they agree ~95% at verdict level, κ=0.708 — triangulating both rarely adds signal). Cost/benchmark basis for every routing call here: references/tool-routing.md.

Quick routing by need (tool ← need; full tool table + per-engine precision in the reference):

• Factual lookup (need the number): Exa web_search_advanced_exa → Brave. perplexity_ask ONLY for biomedical exact numbers (bp/OR/gene sizes; 5-0 vs Exa on the genomics eval) after Exa/Brave miss.
• Factual verification (have the number): verify_claim (Exa /answer, ~$0.005, cached 7d).
• Academic papers: search_papers (S2) → fetch_paper + read_paper before citing.
• Recent papers (<6mo): web_search_advanced_exa category:"publication" + date filter (S2 has no date filter; the enum is publication — "research paper" is rejected by the tool schema).
• Recent preprints: search_preprints (bioRxiv/medRxiv, free, date-range).
• Citation stance: search_literature (scite). Patents/grants/FDA 510(k)/MHRA: scite Pro (no fleet equivalent).
• Entity enrichment: web_search_advanced_exa type:"deep" + outputSchema (per-field citations).
• Database lookups (UniProt/gnomAD/ClinVar): Exa/Brave websearch, NOT S2 (S2 returns papers about the DB).
• News/events: brave_news_search (24h-7d); Exa + date filter for older.
• Deep "why": perplexity_reason ($0.05-0.15). Surveys: perplexity_research (pass reasoning_effort:medium, never default high). Autonomous deep: deep_research ($2-5, async). Multi-step causal chains: parallel_task (retrieval≠synthesis — start with Exa for exact-source memos).

Parallel vs Exa, and the genomics empirical routing (Exa wins exact-source targeting; Parallel core wins broad multi-hop): full write-ups in references/tool-routing.md §Parallel-vs-Exa and §EBF3.

Paper pipeline (Standard+ academic queries) — run this, not just Exa snippets: search_papers -> save_paper (seed papers) -> fetch_paper -> get_paper -> prepare_evidence -> ask_papers(use_rcs=True) This is the highest-quality evidence path. RCS scoring produces significantly better synthesis than websearch snippets (PaperQA2 ablation: p<0.001). 3 well-read papers beat 20 snippet-scanned papers.

When fetch_paper fails (Sci-Hub + Unpaywall both miss): the green-OA / Cloudflare retrieval ladder (confirm-the-wall → find green OA Unpaywall missed via PubMed sidebar / EuropePMC fullTextUrlList → bypass Cloudflare with crawling_exa on the DSpace bitstream → institutional access for genuinely-paywalled) lives in references/tool-routing.md §fetch_paper-failure-recovery.

Critical rules:

• fetch_paper then get_paper/read_paper BEFORE citing. Abstracts are not primary sources.
• Treat the fetch_paper quality card as part of the evidence, not optional metadata. If the card says vetoed, do not cite the paper as ordinary support without naming the veto.
• When citing a paper with quality metadata, surface components directly: RCT, n=200, human, double-blind, placebo, government-funded or VETOED — candidate gene study.
• Never trust PMIDs or PDB IDs from websearch without S2/database verification. Websearch confabulates citation details.
• Sequential exploration: 3 broad queries -> scan results -> 3 narrower queries refining signal. Don't shotgun — query at position 3 in a burst cannot incorporate what query 1 returned.
• Search/discovery tools are map tools. Primary documents, official databases, and fetched full text are evidence tools. Do not blur them.
• For "newest", "latest", or date-sensitive queries, state the date anchor explicitly and bias toward the last 12-18 months unless the field is known to move slowly.
• If one backend is repeatedly failing (403, auth block, junk results), switch early. Do not spend multiple rounds proving the same backend is unreliable.

Full tool table: ${CLAUDE_SKILL_DIR}/references/tool-routing.md (optional depth).

Search Heuristics

• Start broad, narrow later. First 1-2 queries per axis should be short and general (3-5 words). Agents default to overly-specific initial queries that miss relevant results. "protein folding disease" before "alpha-synuclein aggregation Lewy body mechanism." Narrow progressively based on what the broad pass surfaces.
• Exa is semantic, not keyword. Describe the concept — "gene-diet interaction abolishing cardiovascular risk" beats "9p21 diet interaction."
• Use type: "deep" + additionalQueries (2-3 domain-specific variations) for high-value queries.
• Recency by field velocity: Fast fields (AI, markets) = 30-day filter. Medium (biotech, policy) = 6 months. Stable (physics, law) = no filter. If results reference superseded tools or outdated benchmarks, tighten the date.
• Newest-seam research: run one broad recency pass, then narrow immediately to one seam. Repeated broad recency searches mostly generate overlap, not new evidence.
• Training data: Trust for foundations, verify for numbers. Tag [TRAINING-DATA]. The dangerous zone: the more specific and numeric a memory feels, the more likely it's reconstructed. Verify or hedge.
• Scan 8-10 results at summary level, then read the 2-3 with signal. First results are often SEO noise.

Effort Classification

User can override with --quick or --deep. Announce the tier before starting.

Adversarial Mode (--adversarial)

Activated by --adversarial flag OR when the question asks to challenge, debunk, stress-test, or "call BS on" a field, hypothesis, intervention, or result. Always Deep tier.

What it does differently: Inverts the default stance. Standard research asks "what does the evidence say?" Adversarial research asks "what's the strongest case that this is wrong, noise, or built on sand?" — then evaluates whether that case holds.

Reference: ~/Projects/agent-infra/research/adversarial-case-library.md — 45+ curated exemplars of strong adversarial thinking across 9 domains. Read it before starting adversarial research — it has query patterns, URL sources, and quality criteria. The six markers of strong adversarial work: (1) evidence not rhetoric, (2) target genuinely believed by serious people, (3) structural not anecdotal critique, (4) often from insiders, (5) unanswered, (6) constructive null (explains what the data actually show).

Phase modifications:

• Phase 1 (Ground Truth): Inventory what the field claims and WHY people believe it. Steel-man the target before attacking.
• Phase 2 (Search): Mandatory axes: replication status, strongest critic, alternative explanation, meta-analysis of the meta-analyses. Creative Exa queries modeled on case library patterns — describe the concept adversarially ("the strongest case that X is noise," "someone who proved X was wrong with data," "celebrated Y finding that failed to replicate"). Use additionalQueries with 2-3 diverse framings per search.
• Phase 3 (Verify): Flip direction — verify the CRITIQUE, not the original claim. Is the debunking itself solid? Check for: overclaiming by critics, selective counter-evidence, contrarian bias.
• Phase 4 (Synthesize): Output is a case brief, not a balanced review. Structure: Target Claim -> Why It Was Believed -> The Evidence Against -> Strength of the Case Against -> What Survives -> Verdict.

Output template (replaces standard memo):

## Adversarial Review: [Target]

**Target claim:** [what's being challenged]
**Prior belief:** [why serious people believed this, steel-manned]
**Date:** YYYY-MM-DD

### The Case Against
[Evidence marshaled against the target, with sources and effect sizes]

### Strength Assessment
| Criterion | Rating | Notes |
|-----------|--------|-------|
| Evidence quality | HIGH/MED/LOW | [specific] |
| Target was genuinely believed | YES/NO | [by whom] |
| Structural, not anecdotal | YES/NO | [mechanism identified?] |
| Critique is unanswered | YES/NO | [rebuttals exist?] |
| Constructive null | YES/NO | [alternative explanation?] |

### What Survives
[Parts of the original claim that withstand scrutiny]

### Verdict
[DEMOLISHED / MORTALLY WOUNDED / WEAKENED / SURVIVES WITH CAVEATS / CRITIQUE FAILS]

### Sources & Search Log
[Full provenance per standard Deep tier]

Turn budget: After exceeding your tier's tool call budget, force synthesis. Hard ceiling: 40 tool calls (Deep) or 20 (Standard). Reserve remaining capacity for writing. A partial synthesis with sources beats an exhaustive search with no output.

Phase 0 — Dedup Check

Before starting a full research pass, check if this question was already researched recently: !git log --since=24h --name-only --diff-filter=A -- 'research/*.md' 'docs/research/*.md' 'artifacts/*.md' 2>/dev/null | grep -v '^$' | head -20

If a memo on the same topic was written in the last 24 hours, read it first. If it answers the question, skip to Phase 4 (synthesize with any delta). If it partially answers, narrow your scope to the gap.

This prevents the dominant waste pattern: re-running full inventory + external search for a question that was answered hours ago.

Gap and novelty audits: mechanism-consumer completeness

For “what have we not researched/mined yet?”, “what is missing?”, or “why did the scout miss this?”, do not treat a keyword, paper title, lever name, or broad backlog row as coverage. Build a compact coverage matrix across the requested surfaces (for example architecture, objective, loss, optimization, interpretability, data, ontology, and language) and grade each direction as DONE, QUEUE, FAILED, CONDITIONAL, or GAP from live evidence.

A cell counts as covered only when all five are identifiable:

1. input/output contract, including uncertainty or abstention states;
2. mechanism or state transition, not merely a topic/name;
3. live consumer and repository destination;
4. cheapest discriminating gate plus held-out principal metric; and
5. relation to prior work: consumes, supersedes, failed-at, or reopens-if.

If a named local direction lacks one of these, classify it as an underspecified queue/gap and search using the missing mechanism language (“partial AST realizability”, “ontology revision after empty version space”, “multiple-hypothesis data association”), not only the repo's nouns. Every genuine gap must be routed to an existing consumer or a new dependent obligation; “interesting” without a consumer is not a finding. In the closeout, state why the previous discovery process missed each survivor and make the smallest durable process repair that would have surfaced the whole class.

Phase 1 — Ground Truth

Before external search, check what exists locally:

1. Entity pages first — !grep -ril "keyword" docs/entities/ 2>/dev/null | head -5 — these are curated, authoritative, and current. If an entity page covers the topic, read it before doing any external search. Entity pages outrank everything else.
2. Research memos — !grep -ril "keyword" docs/research/ 2>/dev/null | head -10 — prior deep dives on this topic.
3. Personal knowledge — selve MCP search for personal context (conversations, git history, notes).
4. Research corpus — list_corpus for previously saved papers.
5. Training data — what you know (label [TRAINING-DATA]).

Source hierarchy for interpreting results: Entity pages (curated) > research memos (sourced) > git commits (implemented) > AI conversations (raw, often stale). When old AI conversations contradict entity pages, the entity page wins.

Flag contradictions with later findings. Quick tier: If ground truth answers the question, stop here.

Phase 2 — Search

Name 2+ independent search axes before searching. Axes must come from different categories:

If all your axes are from the same category, you have one axis with multiple queries, not genuine diversity.

Perspective-guided divergence (Standard+): Choose 3-5 perspectives (practitioner, critic, adjacent-domain, historian, data analyst, regulator, end user). Generate 3-5 questions from EACH perspective — they must be genuinely different. Merge -> select 5-8 as search queries from 2+ perspective categories. This defeats the Artificial Hivemind structurally (STORM: +25% organization, +10% breadth).

Analogical forcing (Deep, optional): For one axis, reframe through an unrelated domain's lens — "How would a supply chain engineer think about this?" Forces different search terms and literatures.

Search budgets: See effort table above — queries and tool call ceilings are per-tier.

Standard+ academic queries — run the paper pipeline: search_papers -> save_paper -> fetch_paper -> prepare_evidence -> ask_papers(use_rcs=True)

Phase 3 — Verify

Hypotheses (Standard+): Form 2-3 falsifiable claims: "If X is true, we should see Y in the data/literature. If X is false, we should see Z."

Disconfirmation (Standard+, structurally required):

• Search for "X does not work", "X failed", "X criticism", "no association between X and Y"
• Check single lab/group vs independent replication
• If no contradictory evidence after genuine effort: "no contradictory evidence found" (not "none exists")

Citation stance check: For claims that assert literature consensus or direction ("evidence supports...", "studies show...", "the literature suggests..."), and for any fetched paper you plan to lean on heavily, run search_literature (scite) and append [SCITE: S:X C:Y M:Z]. These consensus claims are the highest-risk for hallucination — scite's contrasting citations catch things keyword disconfirmation misses. If scite returns 0 results, note [SCITE: NO COVERAGE] — don't treat absence as confirmation. (~$0/call, user-scope MCP, available everywhere.)

Claim-level checks:

• Numbers: from a source, or generated? If generated -> [ESTIMATED]
• Names/authors: from a source you accessed, or memory? If memory -> verify or [UNVERIFIED]
• Papers: does this actually exist? If unconfirmed, DO NOT cite it
• For high-stakes claims (valuations, statistics): use verify_claim (~$0.005/call)

YOU WILL FABRICATE under pressure to be precise. The pattern: real concept + invented specifics (author name, fold-change, sample size). Catch yourself. Vague truth > precise fiction.

Citation rate-gate (mechanical, run at finalize). The per-claim checks above are judgment; this is the deterministic backstop. Before publishing a memo with arXiv IDs / DOIs, run:

uv run --no-project python3 ${CLAUDE_SKILL_DIR}/scripts/verify_citations.py <memo.md>   # --json for the machine form

It resolves every arXiv ID + DOI against Crossref / arXiv / DBLP (keyless, zero-dep) and emits a dispatch-brief.

• hallucinated > 0 is BLOCKING — a citation resolving to nothing is fabricated; fix or remove it before finalizing (exit code is non-zero).
• Advisory (report in the source log, act if cheap): resolved_rate < 80%, arxiv_only_ratio > 60%, and DBLP venue-upgrade candidates (an arXiv-only cite that has a published version). Advisory thresholds stay advisory until a measured miss promotes one (constitution P3).
• unreachable ≠ hallucinated — transient API failures are reported and excluded from the rate, never block. Re-run.

Limitation: the gate proves a cite exists and classifies its venue; it does NOT prove a DOI resolves to the claimed paper (title-mismatch) — that stays with the in-session verify_claim / read_paper checks above. ADR: agent-infra decisions/2026-06-19-autoresearch-grounding-imports.md.

Phase 4 — Synthesize

Stop conditions (VMAO-derived, arxiv:2603.11445 — +35% completeness from formal verification):

Evaluate after every search round, not just at the end. After each search round (2-3 queries), assess: "How many of my named search axes have produced at least one useful result? What's missing?" A "search axis" = one of the axes you named in Phase 2. "Nothing new" = the last 2 tool calls returned results already covered by existing findings.

When to stop searching: if the new material is collapsing into existing local concepts, the last two rounds only produce narrower restatements, or the best surviving idea is just one bounded caveat, synthesize and stop. Do not keep searching to make the output look richer.

Recite evidence before concluding. List concrete data points from sources, then derive the conclusion. "Study A: 26% improvement (n=500). Study B: no effect (n=200). Weighing by sample size..." This surfaces contradictions that narrative synthesis buries.

Mark every quotation; reword the rest. Organize the write-up around where sources agree and differ, in your own indirect speech, one or two sentences per source. A source's exact wording appears only inside quotation marks with its citation; unmarked reproduction of source passages is a provenance defect (Fable 5.1 reproduces source text unmarked more often than Fable 5 — prompting guide, 2026-09).

Refuse if evidence is insufficient. If you have only [TRAINING-DATA] or [UNVERIFIED] tags with no retrieved sources directly addressing the question, output:

## Insufficient Evidence
**Question:** ... | **Searched:** ... | **Not found:** ... | **Partial findings:** ... | **Next steps:** ...

Confident synthesis from noise is worse than an informative refusal.

Source assessment: For each key source — peer-reviewed vs preprint vs blog? Sample size, methodology, COI? Confirms or contradicts prior work? "We don't know yet" is valid.

Claim-Level Source Gating (Mandatory)

Every empirical quantitative claim (effect sizes, trial results, epidemiological numbers, "X reduces Y by Z%") in the output MUST have:

• A resolved DOI or PMID
• The specific finding from that source that supports the claim

This does NOT apply to:

• Deductive claims — math, inference from known quantities, logical derivation -> tag [INFERENCE], valid without DOI
• Computed values — dose/kg from body weight, unit conversions -> tag [DATA], valid
• Established facts — well-known biology (e.g. "CYP2D6 metabolizes codeine") -> no source needed

Empirical claims without source provenance are tagged [UNSOURCED]. Unsourced empirical claims should not drive protocol decisions or dosage recommendations without explicit acknowledgment of the gap.

Paper quality assessment (auto-runs on fetch_paper, returns quality card):

Hard vetoes (block citation):

• RETRACTED — never cite retracted papers (Crossref + PubMed fallback)
• CANDIDATE_GENE — pre-GWAS single-gene association studies (~98% non-replication). Exempts PGx (CYP*, HLA, UGT) and mechanism studies.

Informational flags (surface to reader, don't auto-block):

• NON_HUMAN_ONLY — animal/in-vitro biological study. Transfer depends on context: conserved pathways (energy metabolism, DNA repair) transfer; PK, behavior, immune do not. State the organism.
• CASE_REPORT_ONLY — case reports establish existence, not magnitude.
• Blinding (open-label inflates effects), control type (waitlist inflates), single-center, industry-funded, no pre-registration, data "on request" — all informational.

When citing papers with quality metadata, include component display: "Paper X: RCT, n=200, human, double-blind, placebo, gov-funded" or "Paper Y: VETOED — candidate gene study"

Don't build composite scores or traffic lights from these components. Display them as a list. The citing agent (you) makes the judgment call based on context.

Quantitative-Claim Construction Gate

Source gating covers where a number came from; this gate covers how it was built. Before finalizing any memo where a number does argumentative work: every load-bearing figure carries its ledger, unit, base, window justification, gross-or-net status, and measurement-vs-model-output label; when several constructions are defensible, the range across constructions is the headline. Full 32-item checklist with historical anchors: ${CLAUDE_SKILL_DIR}/references/quant-bias-checklist.md.

Quick Tier

Answer inline with source citation. No formal report.

Standard Tier

## [Topic] — Research Memo

**Question:** [what was asked]
**Tier:** Standard | **Date:** YYYY-MM-DD
**Ground truth:** [what was already known]

### Claims Table
| # | Claim | Evidence | Confidence | Source | Status |
|---|-------|----------|------------|--------|--------|
| 1 | ... | RCT / dataset | HIGH | [DOI/URL] | VERIFIED |
| 2 | ... | Inference | LOW | [URL] | INFERENCE |
| 3 | ... | None found | — | — | UNSOURCED |

### Key Findings
[With source quality assessment and paper quality metadata where available]

### What's Uncertain
[Unresolved questions]

Deep Tier

Standard tier plus: Disconfirmation results, Verification log (claims verified via tool vs training data), Search log (queries, tools, hits/misses).

Provenance Tags

Tag every claim:

• [SOURCE: url] — Retrieved document
• [DATABASE: name] — Reference database query (ClinVar, gnomAD, DuckDB)
• [DATA] — Our own analysis, reproducible
• [INFERENCE] — Logically derived (state the chain)
• [TRAINING-DATA] — From model training, not retrieved
• [PREPRINT] / [FRONTIER] — Unreplicated work
• [UNVERIFIED] — Plausible but not verified
• [SCITE: S:X C:Y M:Z] — Citation stance from scite
• [ESTIMATED] — Number generated, not sourced

Never present inference as sourced fact. Never present training data as retrieved evidence.

Precedence: Admiralty grades ([A1]-[F6] per references/source-grading) are the standard for investigation/OSINT — strictly more granular. Don't duplicate by tagging the same claim with both systems.

• Synthesis mode default: Summarized training data instead of fetching primary sources — THE failure mode
• Confirmation bias: "X validation" queries instead of "X criticism" or "X failed"
• Precision fabrication: Invented numbers or author names under pressure to be precise
• Telephone game: Cited primary via review without reading the primary
• Single-axis search: All queries from same starting point, same intellectual tradition
• Named-entity scoping: Researched only the people/tools/things named in the prompt when the task is governed by a whole DISCIPLINE + product-category you never went and got. If research feeds an application on a standing surface (editor, tool, pipeline, UI), name the governing field first, pull its canonical heuristics + category conventions, and apply them — the named entities are a subset, not the field. (#f 2026-06-18: researched Victor/Matuschak to improve an editor, skipped HCI/UX entirely.)
• Source hoarding: Saved papers but never fetched/read them
• Scope creep without pushback: User asks 15 things — say "this session can handle N well; which are priority?"
• Websearch citations as primary: Trusting PMIDs/PDB IDs from websearch without S2 verification
• Academic tools for database lookups: Using S2 for gnomAD frequencies — query the database directly
• Exa monoculture: Using only Exa when S2/scite/Brave would reach different sources
• Skipping paper pipeline: Using Exa snippets when search_papers -> fetch_paper -> ask_papers would give higher-quality evidence

Skill upkeep: if this research arc hit a defect or friction in THIS skill (a dead tool route, a wrong cost, a broken pointer), log it for the next run — ~/Projects/skills/hooks/append-skill-memento.sh research '<one-line issue>'.

$ARGUMENTS

Known Issues

• [2026-07-09] fetch_paper hangs indefinitely on direct arXiv DOI; parallel fetch also serial-locks after first item, requiring termination and direct arXiv PDF fallback

• [2026-07-10] mcp__research__fetch_paper can hang beyond 60s on an arXiv PDF; terminate the transport, then use the primary-source browser or direct PDF tooling instead.

• [2026-07-14] 2026-07-14: mcp__research__search_papers backend=s2 hung over 45s on exact new arXiv id 2607.07508; direct primary arXiv PDF worked.

• [2026-08-12] fetch_paper hung >180s on DOI 10.1097/FPC.0b013e328012b8e4; an existing metadata-only DOI also returned immediately without fetching paper.pdf

• [2026-08-17] Exa web_search_advanced_exa date filter unreliable for category=publication (null dates, 2024/25 leakage) — for dated arXiv sweeps use the arXiv Atom API with submittedDate ranges (2026-08-17 design-agents lane)

• [2026-08-17] verify_citations.py regex misses bare arXiv IDs like "2607.20767" without an "arXiv:" prefix — reports vacuous PASS; curl-loop the abs pages as backstop until fixed (2026-08-17)

• [2026-08-18] verify_citations.py truncates old-style Elsevier DOIs containing parens (10.1016/S0160-2896(03)00053-9 -> "10.1016/S0160-2896(03)" flagged HALLUCINATED); regex needs paren-aware DOI matching — manually curl-resolve paren DOIs until fixed (2026-08-18)

• [2026-08-24] fetch_paper 0/8 on 2026 open-access DOIs (Nature/BMC/MDPI/Cambridge: Sci-Hub has no 2026 content, Unpaywall missed; 3 of the 8 hung >120s). Working route for anything with a PMCID: EuropePMC https://www.ebi.ac.uk/europepmc/webservices/rest/<PMCID>/fullTextXML (6/6, no wall; S2 search_papers returns the PMCID). Publisher HTML via urllib → 3 KB cookie walls; bioRxiv 429; Sage/T&F 403; crawling_exa OK on nature.com/mdpi/BMC HTML but times out on Springer PDFs and OSF downloads (plain curl -L gets the OSF PDF). search_preprints returned [] on 3 well-formed queries the same day (2026-08-24, iq-sex-differences)

• [2026-09-01] 2026-09-01: arXiv export API silently IGNORES parenthesized OR-groups in search_query — (abs:A OR abs:B) AND abs:C returns unfiltered noise; a noisy result set is a BROKEN QUERY, not a negative. Use simple two-term ANDs or intersect client-side.

• [2026-09-02] references/source-grading is named by SKILL.md but absent from the installed research references directory (2026-09-02 Eileen Gu claim audit)