Manual Literature Mining Protocol — Paperclip MCP Discipline

A small but load-bearing methodology page. Codifies the verification discipline for using Paperclip MCP (a full-text PubMed / bioRxiv / medRxiv / arXiv index) safely in this project — surfaced 2026-05-05 by the wiki sweep daemon (Pass 2 Connection #3 + Priority Action #1). The index is useful for deep-dive questions, but the map operator (the convenience-aggregation primitive) has produced wrong organisms, fabricated kinetic numbers, and made-up author affiliations, making automated unverified use unsafe. This page defines the safe-use workflow.

Why this protocol exists

Specific failure mode documented 2026-05-05: a Paperclip map query against a multi-paper question returned organism names not present in the source papers AND specific kinetic numbers (Km, IC50) that did not appear in any of the cited sources when grep-verified. The map primitive appears to do model-side aggregation of partial extracts, which the underlying model fills in plausibly but not faithfully. (Memory record: feedback_paperclip_map_unreliable.md, decided 2026-05-05.)

search, cat, head, grep, scan primitives DON'T exhibit this failure mode — they return raw paper content, line-anchored, verifiable. The protocol below confines automated use to those safe primitives and requires explicit verification before any quantitative claim derived from Paperclip enters the wiki.

The five-rule discipline

1. Use safe primitives only

Allowed: search, cat, head, grep, scan. Forbidden in any automated workflow: map, reduce, or any aggregation primitive that produces synthesized output rather than raw paper content.

If a finding requires aggregation across multiple papers (e.g., "summarize the IC50 range for compound X across all NLRP3-relevant studies"), do the aggregation in two stages: (1) search + grep to surface the specific lines containing IC50 values; (2) human-verify each value against cat of the source paper; (3) the synthesis happens in the wiki page, not in Paperclip's primitive call.

2. Anchor identity to meta.json

Every paper Paperclip returns has a /papers/<id>/meta.json file with: title, authors, journal, year, PMID, doi. Cite from meta.json, not from inferred metadata in the paper body. A common failure mode is misattributing a paper based on a title-like string in the body that turns out to be a citation TO another paper, not the paper's own title.

Workflow: cat /papers/<id>/meta.json | jq .title (or read directly) before quoting a paper.

3. Grep-verify all numbers before they enter the wiki

For any quantitative claim sourced from Paperclip — IC50, Km, dose, sample size, percent change, p-value, cohort n — grep for the specific number in the source paper's content.lines file BEFORE writing the number into the wiki. The verification step takes ~30 seconds and catches the map-class hallucination directly.

# Example: verify a claimed IC50
grep -i "IC50" /papers/<id>/content.lines
grep "<exact-number>\\|<exact-number with units>" /papers/<id>/content.lines

If the number doesn't appear in the source, it didn't come from the source. Do not propagate.

4. Never propagate map/reduce summaries

If you find yourself looking at output from Paperclip's map or reduce primitives (e.g., from a prior session, from a logged subagent task, from the daemon's intermediate state) — treat that output as un-validated and re-derive from the safe primitives. A map summary that LOOKS reasonable cannot be assumed reasonable.

This applies even when the map output is convenient. Convenience is the failure-mode signal — map is fast precisely because it skips verification.

5. Cite line-anchored, with the project's citation format

Paperclip provides line-anchored URLs of the form https://citations.gxl.ai/papers/<doc_id>#L<n>. Always cite at the line-anchor level, never just at the paper level. This makes verification trivial for a future reader (or for the sweep daemon's Pass 3 reviewer) and forces the citation to point at a specific claim, not a generalized invocation of "the paper."

Citation discipline (from existing wiki convention): - Inline: [1], [2] - References block at end of page:

--------
REFERENCES
[1] Authors. "Title." *Journal* vol, pages (year). doi:XX
    https://citations.gxl.ai/papers/<doc_id>#L<n>

- Never expose Paperclip's internal doc_id in prose (only in the URL).

Pre-commit verification gate (the rule that catches errors BEFORE the sweep, not after)

Rule 3 above ("grep-verify all numbers before they enter the wiki") covers Paperclip-sourced numbers specifically. This section generalizes the same discipline to every load-bearing quantitative claim in newly-authored wiki content — disulfide counts, residue counts, sequence lengths, IC50s, Kms, dose-response numbers, cohort sizes, percent changes, kinetic constants, evidence-tier counts. Whether the source is Paperclip, UniProt, ChEMBL, ClinicalTrials.gov, a PMC paper, or a database API response, the verification gate is the same.

The rule:

Before any commit lands a wiki page (especially wiki/<comp-NNN>-*.md interpretive pages and wiki/hypotheses/H<NN>-*.md cards) that introduces a load-bearing quantitative claim, the author must grep-verify the claim against its primary source and add a line-anchored citation inline. A claim that cannot be grep-verified is not committed; it is either re-derived, replaced with a bounded "TBD pending source verification" marker, or dropped.

Why this gate exists:

The wiki sweep daemon (Pass 1 Propagate → Pass 2 Synthesize → Pass 3 Review) is good at catching cross-page inconsistencies — exactly the failure mode that surfaced the DAF SCR1-4 disulfide-count error on 2026-05-06 (Sweep A Connection 2). But that means the discipline currently catches errors ~12–24 hours after they ship into the corpus, by which time:

• The wrong number has propagated to multiple pages (DAF SCR1-4: comp-012 → H05 stub).
• The wrong number has been ingested into the synthesizer's context for downstream reasoning (the chaperone-orthogonal triple-cassette concern used ~~1.8×~~, a fabricated coefficient with no valid replacement).
• A second-opinion synthesis pass and a hand-walkthrough are required to find and fix the propagation, instead of catching it at the source.

The sweep is a backstop, not a substitute. The pre-commit verification gate is the right moment to catch hallucinated numbers — at the moment they would enter the corpus, not after they've been laundered through the substrate.

Operational pattern:

When authoring new wiki content (especially comp-NNN interpretive pages, H-card stubs, scope pages, or any page making mechanistic / kinetic / structural claims), follow this micro-protocol per quantitative claim:

1. Identify the load-bearing numbers in the draft. Anything that downstream reasoning will depend on (cassette counts, disulfide counts, kinetic constants, evidence-tier verdicts, cohort sizes, percent-change magnitudes). Numbers used for color or rough order-of-magnitude framing are lower-stakes; load-bearing numbers feed into matrices, decision criteria, threshold gates, or other pages' calculations.

2. For each load-bearing number, name the primary source. UniProt accession + feature line, PMID + page/section, ChEMBL ID, NCT trial ID, etc. If you cannot name the primary source, the number is suspect — either find it or drop it.

3. Grep-verify the number against the primary source. For UniProt: curl -s "https://rest.uniprot.org/uniprotkb/<ACC>.txt" | grep "<feature>". For PMC: grep -i "<number>" /papers/<id>/content.lines. For ChEMBL: pull the bioactivity record and check the value field. The verification should produce the number directly from the source, not be inferred from a summary.

4. Cite line-anchored inline. Per Rule 5 above for Paperclip; for UniProt: (per UniProt P08174 DISULFID feature: Cys36-Cys81) is line-anchored enough. For PMC: (PMID 12345678 Table 2) or (PMID 12345678 §Results para 3). The citation must let a future reader (or the sweep's Pass 3 reviewer) re-verify in <30 seconds.

5. If the number cannot be verified, do not ship it. Options: re-derive from a different source, drop the claim and note the gap, or write a placeholder like [TBD: pending UniProt verification] and keep it out of the load-bearing path until verified.

What counts as load-bearing — heuristics:

• Any number that appears in a downstream comparison, table, matrix, decision criterion, or quantitative threshold ("if synergy <0.7, then..." / "29 vs. 25 total disulfides").
• Any number that, if wrong, would change the evidence-tier verdict, the experimental design, or the platform decision.
• Any number that other wiki pages will cite or reuse.
• Any number that the sweep daemon's Pass 2 synthesizer might pull into a cross-doc connection (which, in practice, is almost any quantitative claim — the synthesizer reads everything).

What counts as low-stakes — exempt from the gate:

• Order-of-magnitude framing where precision doesn't matter ("on the order of millions of dollars," "weeks not months").
• Numbers used purely for color in narrative prose, where the surrounding text would still be correct if the number were ±20% off.
• Numbers explicitly tagged as estimates or projections (with the tag making the uncertainty visible).

The DAF SCR1-4 incident (2026-05-06) — provenance for this rule:

The 2026-05-05 Sonnet subagent that authored wiki/daf-cd55-scr14-truncated-computational.md (comp-012) asserted "3 conserved disulfide bonds per SCR domain → 12 total" in 4 places of prose narrative. The comp-012 pipeline (AlphaFold pLDDT-based protease stability) does not actually count disulfides; comp-012's own Limitations section explicitly says "Disulfide bonds not modelled." The "12" was hallucinated at write-time. The error then propagated into wiki/hypotheses/H05-daf-scr14-cp0-thesis.md (the CP0-closure thesis stub) without independent verification.

The 2026-05-05 sweep daemon caught the inconsistency the next day (Sweep A Connection 2: chaperone framework had "8 (4 SCRs × 2 disulfides each)," comp-012 + H05 had "12") and surfaced it as a Priority Action. Verification against UniProt P08174 during the 2026-05-06 walkthrough confirmed 8 DISULFID feature annotations across SCR1-4 (the canonical sushi/CCP fold: Cys1-Cys3 + Cys2-Cys4 motif, 2 per domain).

If the pre-commit verification gate had been in place when comp-012 was authored, the number would have been grep-verified against UniProt at write-time and the wrong claim would never have shipped. The sweep would still have run, but it wouldn't have needed to find this class of error — only genuine cross-doc synthesis findings.

This rule generalizes: the sweep daemon should be catching novel cross-doc connections, not catching fabricated coefficients. Numerical hygiene is upstream of synthesis.

---

Killshot tiering — pick the lowest-cost experiment that resolves the question

Sister discipline to the pre-commit verification gate: before declaring an experiment as the "killshot" for an open hypothesis, walk a cost ladder and pick the lowest-tier option that resolves the question above noise floor for the platform's purposes. The default "killshot" definition in institutional research is grant-tier ($50K–$500K mouse / cell-line / cohort study). For an open-source citizen-science project run by a CTO not a pharma lab, that default is wrong — both literally too expensive AND epistemically over-specified relative to the decision the experiment is supposed to inform.

The OE killshot tier menu:

Tier	Cost	Time	Source / route	Examples
Tier 0	$0	hours-days	Public dataset mining; lawful full-text re-read of papers comp-NNN got at abstract-level; published-data synthesis	GTEx + Human Protein Atlas sex-stratified expression mining; GWAS catalog lookup; UniProt feature-annotation queries; PubMed Central, author manuscripts, institutional-library or interlibrary-loan retrieval; Mendelian randomization via published MR-Base summary statistics
Tier 1	$200–500	days-weeks	Leverages existing self-experiment infrastructure	n=1 LabCorp / Quest panels (FEUA, hs-CRP, hormone panels); 16S microbiome OTU mining if the panel is already running; spot urinary biomarker measurements; serum-panel snapshots with pre-committed protocol
Tier 2	$0–500	weeks	Crowdsourced / community-leveraged	Recruiting men's-health forum cohorts to share LabCorp panels; Twitter/Reddit-based n>>1 cohort assembly; community-sourced patient-reported outcomes; OSF-style collaborative protocols on existing self-experimenter populations
Tier 3	$2–5K	4–12 weeks	Friendly bench (community college, undergrad thesis, sympathetic small lab)	Caco-2 / HEK293 / HepG2 cell-culture work with consumables-only budget; small-scale Western blot / qPCR / ELISA assay; tissue-culture collaboration where the academic partner has the equipment and the OE side funds reagents
Tier 4	$5–15K (academic collaborator) / $30–60K (institutional)	8 weeks–3 months	Academic-collaboration animal study / formal cell-line work / contracted assay	Mouse castration + replacement studies via academic IACUC-protocol-leveraged route; published-with co-authorship arrangement; CRO-tier work (the institutional baseline) is the last resort

The discipline: before promoting an experiment to "killshot" status, walk the ladder Tier 0 → Tier 4 and ask at each tier "could I resolve this question above noise floor without spending more?" Default to the lowest tier that answers yes. The mouse experiment is rarely the right first move for an OE-tier question; it is sometimes the right last move.

Why this matters specifically for OE:

1. Budget reality. OE is a CTO + AI-substrate research operation, not pharma. A daemon sweep costs ~$0.65; freaking out over a $50 sweep cost is the appropriate calibration. Proposing $30K+ experiments without walking the upstream cost ladder is a category error.
2. Tier 0 + Tier 1 frequently resolve or sharply reroute the question. COMP-016's bounded scan did not identify direct androgen-driven intestinal ABCG2 suppression, and COMP-017 corrected the Hoque magnitude, Liu and Slepnev attributions, and exposure interpretation. That was enough to reject the asserted direct-suppression mechanism at the reviewed-source tier. It did not resolve healthy-human intestinal sex stratification because no direct GTEx/HPA values were extracted. Tier 0 can therefore kill an overclaim while leaving a narrower measurement question open; a later bench experiment is justified only for what the public evidence could not decide.
3. Existing self-experiment infrastructure makes Tier 1 nearly free. OE has running n=1 self-experiment protocols (self-experiment-protocol.md) with established lab-panel workflows. Adding a new measurement to that workflow costs $50–100 per data point, not $30K.
4. Crowdsourced cohorts produce real n>>1 evidence at $0 marginal cost. The men's-health, gout, and microbiome communities on Twitter / Reddit / Hone routinely share lab panels publicly. Treating that as a usable data source (with appropriate methodological caveats) is the open-source-platform thesis applied to evidence gathering.
5. The "killshot" framing should be: cheapest experiment that resolves the question, not biggest experiment that proves the answer beyond doubt. Falsification-card discipline (per linter-design.md) is about attempting to falsify, not about overwhelming the question with budget. A $300 experiment that crosses a pre-committed threshold kills (or saves) a hypothesis as decisively as a $30K one.

This discipline composes with the pre-commit verification gate. Both are "walk the checklist before shipping." The verification gate catches numerical hallucinations; the killshot tiering catches budget over-specification. Different failure modes; same shape of fix (named protocol, applied at the right moment in the workflow).

For falsification-card stub authoring (per hypotheses/README.md), the killshot menu in the stub should be tier-explicit: each killshot listed with its cost tier, so the stub-to-full-card promotion can pick the right starting tier rather than defaulting to institutional.

---

Specific OE questions worth Paperclip search-and-verify time

These are questions whose primary-literature grounding would meaningfully advance the platform but require non-trivial scanning. Listed here so they're discoverable as a queue (anyone reading this page can pick one up):

1. Does Candida utilis uricase have published exact-construct processing or cassette data? — informs a future sequence-bound construct comparison; retired COMP-011 supplies no KEX2, mutation-benefit, clinical-sequence, or cassette-risk conclusion. Search: Candida utilis uricase Aspergillus + grep exact sequence, signal peptide, junction, processing, secretion.
2. Aspergillus oryzae heterologous expression precedents for exact single- and dual-protein configurations, particularly solid-state formats. Retired COMP-010 supplies no production-capacity conclusion. Search: Aspergillus oryzae solid-state heterologous protein + grep dual cassette, multi-cassette, two genes.
3. ABCG2-probiotic intersections — does any published probiotic strain (engineered or natural) modulate intestinal ABCG2 expression in vivo? Informs the LBP track's butyrate-via-F. prausnitzii thesis. Search: ABCG2 probiotic + grep intestinal expression, fold-change.
4. GRAS-host complement-regulator expression — has anyone published heterologous expression of soluble complement regulators (sCR1, Factor H, DAF/CD55) in any GRAS organism? Informs H05, §1.25, and future exact-configuration work; retired COMP-006 and COMP-012 supply no survival result. Search: complement regulator heterologous expression yeast Aspergillus + grep specific protein names.
5. Si Miao San and related TCM gout formulas — modern Chinese clinical evidence. Per the global-multilingual default (CLAUDE.md §"Global-multilingual research by default"), search Chinese-language sources directly via CNKI / WanFang / ChiCTR, not just PubMed-indexed translations. Informs the TCM × rigor track (tcm-modern-rigor-intersection.md) Phase 2 P2-1.
6. NSlD-ΔP10 strain availability outside the Maruyama lab — verify the operations/ward-1995-lab-access.md finding that the strain isn't in any public repository by direct catalog query at JCM, NBRC, CGMCC, CBS-KNAW, ATCC, FGSC.

Query-framing for non-Western compound classes (added 2026-05-19)

The 2026-05-19 traditional-name re-scan (method receipt) showed that mechanism-name-only seeding had produced query-bound gaps in the TCM and fungal-natural-product work. Species, material, formula, and traditional-pathology frames surfaced additional primary-study leads that the original database and mechanism queries did not. Neither pass was a coverage census, and the second pass supplies no cross-material rank.

Discipline: for any compound-class scan touching natural products, TCM, Kampo, Ayurveda, or other non-Western traditional medicine subfields, seed queries from multiple framings, not just mechanism-name:

1. Mechanism-name query (the default Western frame): <target> + <compound-class> (e.g., "ABCG2 flavonoid", "XO inhibitor anthraquinone"). Catches the Western-curated database overlap.
2. Species-name query: scientific binomial + traditional name in original language (e.g., Phellinus igniarius AND 桑黄; Wolfiporia cocos AND 茯苓). Catches the species-anchored evidence regardless of which target the paper frames around.
3. Traditional-formula-name query: classical formula composition (e.g., Si Miao San AND component herbs; Bai Hu Jia Gui Zhi Tang AND mangiferin). Catches formula-level RCT evidence + the cardinal-herb attribution chain.
4. Traditional-pathology-term query: the original-language pathology framing (e.g., 痛风, 痹证, 湿热痹, 消渴). Catches indication-anchored evidence where the paper doesn't index against a Western target column.

When to use which: if a Pass 1 or comp-NNN scan returns no match for a fungal or botanical chokepoint after a mechanism-name-only query, treat this as a candidate for query-framing re-scan, not a confirmed empty chokepoint. Additional leads recovered by a second frame demonstrate search sensitivity; they do not make the second pass comprehensive.

Pre-commit gate interaction: query-framing failures and the pre-commit grep-verify gate (§"Pre-commit verification gate" above) target different error modes. The pre-commit gate catches fabricated or wrong numbers inside a single paper's claims. Query-framing catches whole subfields of evidence the search never surfaced. A well-verified claim from a query-framing-incomplete scan can still mislead — the absence in the scan output isn't an absence in the literature. Apply both disciplines independently.

Cross-references: - scripts/SWEEP-ARCHITECTURE.md — the current full-synthesis and source-rehydration architecture - logs/lit-scan-query-framing-retrospective-audit-2026-05-19.md — query-framing retrospective; neither pass was a literature-wide recall census - CLAUDE.md §"Global-multilingual research by default" — the upstream discipline this operationalizes

Translation protocol (two-model independent cross-check + inline disagreement annotations)

Canonical home for the translation discipline — pointed to from CLAUDE.md, AGENTS.md, and the lit-scan skill. The translate_source_two_model() helper in experiments/lib/agentic_lit_synthesis.py implements it.

When ingesting non-English source material, translate with two independent models (different vendors, ideally different training pipelines) and produce an annotated translation that surfaces disagreements rather than collapsing them. Same multi-vendor cross-check discipline the wiki sweep daemon already uses (Pass 1 Propagate + Pass 2 Synthesize + Pass 3 Review per wiki/etc/open-source-platform.md §"Multi-model synthesis as guard against epistemic homogenization"). Translation has the same homogenization risk; the same heterogeneity guard applies.

Operational pattern:

1. Two independent translations. Pick two models from different vendors:
2. Model A: Claude (Anthropic) OR Gemini (Google)
3. Model B: DeepSeek OR Qwen (both Chinese-vendor; native-language depth) OR GPT-5 (OpenAI) For Chinese-source material, at least one model should be a Chinese-vendor model (DeepSeek or Qwen) — the native-language training depth catches idiomatic and classical-TCM-terminology nuances Western-trained models miss. For Japanese-source material, prefer a model with strong Japanese (Claude is competent; Gemini is reasonable; for deep Kampo terminology a Japanese-vendor option if available; otherwise two competent Western models is acceptable). The Model-A cost rule: when a Claude subagent runs the scan, it IS Model A — the cross-check only pays OpenRouter for Model B. Never run both halves through OpenRouter (feedback_subagent_as_model_a.md).
4. Sentence-level comparison. Compare the two translations at sentence granularity (or paragraph if sentences are too short to differ meaningfully).
5. Where models AGREE → confident translation. Use that text directly.
6. Where models DISAGREE → inline annotation. Use a clear, scannable convention:

   The compound shows {Model A: "significant" | Model B: "notable"} reduction in IL-1β secretion at 10 μM.
   

   Or for substantive disagreements (different mechanism implications, evidence tier, magnitude):

   {Model A: "decreased serum urate by 1.2 mg/dL"} {Model B: "decreased serum uric acid by approximately 71 μmol/L"}
   [TRANSLATION NOTE: Models agree on direction and rough magnitude (1.2 mg/dL ≈ 71 μmol/L), differ on unit choice in the source. Verify against original-language paper if precision matters for downstream calculation.]
   

7. For load-bearing claims, escalate. If a translation disagreement affects an evidence-tier judgment, a dose calculation, a mechanism mapping, or a chokepoint assignment — flag it explicitly with [TRANSLATION-DISAGREEMENT] so future readers know the underlying source has interpretive ambiguity. Do not resolve the disagreement silently by picking one translation.
8. Specific high-risk categories where disagreement should always be flagged: scientific terminology with mechanism implications ("inhibits" vs "modulates" vs "suppresses"), evidence-tier hedging ("shows" vs "suggests" vs "may"), dosing units and routes, classical TCM / Kampo / Ayurvedic terminology vs modernized equivalents, statistical-significance language, sample-size and study-design descriptions.

Why two models, why independent, why surface disagreements: translation is interpretation. A single model's interpretation has its training-distribution bias, vendor bias, and idiomatic-fluency strengths/weaknesses baked in. Two models from different vendors share less bias. The disagreements are EXACTLY where translation nuance lives — silently picking one model's choice loses information the original-language paper had. Surfacing the disagreement preserves the precision the source intended.

Cost note: translation runs add a variable, provider-dependent marginal cost only to explicit lit-scan or source-ingestion flows, not to the ordinary sweep daemon. Record the actual cost in the compact method receipt rather than relying on a standing per-paper estimate.

Why this rule exists: added 2026-05-05 with the global-multilingual default. Brian's framing: "if we bring in non-english papers, i think we need to have a protocol for translation that involves 2 completely independent models and we can have inline annotations where the models may disagree on nuance because in science nuance and precision matter." The discipline matches the heterogeneity-guard logic established for the sweep daemon.

When Paperclip is the wrong tool

• For ChEMBL bioactivity data — use the ChEMBL MCP directly. Paperclip's index is paper-level; ChEMBL is target-and-compound-level with curated quantitative bioactivity. See chembl-cross-check.md for the cross-check discipline.
• For patents — Paperclip indexes academic literature, not patent literature. Use Google Patents, Espacenet, Lens.org, JPlatPat, CNIPA directly via WebFetch.
• For non-English-language clinical trial registries — ChiCTR (Chinese), JPRN (Japanese), KISS (Korean) registries are not in Paperclip's index. Direct WebFetch of those registries is the right tool.
• For real-time updates — Paperclip's index has an indexing lag. For the very most recent findings (within the last 1–4 weeks), the journal websites or PubMed direct may have content Paperclip doesn't yet have.

Cross-references

• paperclip-deep-dive.md — full Paperclip MCP capability + limitation audit (the upstream documentation this page operationalizes for OE)
• bio-ai-tools.md — broader AI-tool landscape for biology research
• chembl-cross-check.md — sister discipline for ChEMBL bioactivity verification
• open-source-platform.md §"Multi-model synthesis as guard against epistemic homogenization" — the broader epistemic-rigor framework this protocol fits within
• CLAUDE.md §"Global-multilingual research by default" + §"Translation protocol" — the multilingual + cross-vendor disciplines that combine with this protocol for non-English Paperclip-adjacent work