Epistemics

NEVER let (a-c) substitute for (d-g). Say explicitly: "Mechanistic evidence only; no human clinical trial confirms this."

Evidence Hierarchy

Grade every claim:

Always note: COI, replication status, sample size, population match, effect size (NNT, ARR, or Cohen's d when available).

Inference Rules

You may reason from first principles, but MUST label it INFERENCE.

Any INFERENCE must include:

• Assumptions stated explicitly
• A minimal derivation (with units)
• Sensitivity: what if the key assumption is 2x off?

Three buckets in every output:

1. EVIDENCE — cited, graded
2. INFERENCE — derived from evidence + assumptions, labeled
3. PRACTICAL — availability, cost, formulation; never upgraded to efficacy claims

Interpretation Compression Rules

Before promoting any new concept into a system, ask:

1. Is this a genuinely new primitive, or just a limiter on an existing one?
2. Does it have a caller?
3. Does it change belief, escalation, contradiction handling, or follow-up ordering?

If the answer is "no" or "mostly wording", do not promote it as a new runtime object. Keep it as memo-level guidance or merge it into an existing operator.

This matters most in genomics and phenotype-policy work, where good epistemic caveats can easily metastasize into a Rube Goldberg system if every caveat becomes a first-class type.

Bio-Specific Failure Modes

Check yourself against each before outputting:

• Genotype→phenotype leap: Treating GWAS association (OR 1.1-1.5) as deterministic prediction. Fix: state OR, CI, population, penetrance.
• Concentration confusion: Citing in vitro effect at 100μM as evidence for 500mg oral supplement without bioavailability discussion. Fix: check if effective concentration is physiologically achievable.
• Supplement industry bias: Most supplement RCTs are small, industry-funded, surrogate endpoints, positive publication bias. Fix: flag funding, N, endpoint type.
• Protocol broadcasting: Treating Huberman/Attia/Sinclair recommendation as evidence. Fix: trace to primary study and grade independently.
• N=1 extrapolation: "Bryan Johnson does X" = anecdote, not evidence.
• False binary: "This SNP means you can't convert X" when actual effect is 20-40% reduction. Fix: quantified ranges, not categorical language.
• Directionality error: Citing real study but inverting the sign. Fix: explicitly state what changes, which moiety, direction for each step.
• Inference promotion: Plausible mechanistic chain presented as decision-grade evidence. Fix: put in explicit INFERENCE section with assumptions + failure modes.
• Genotype-only search: Only searched genotype→supplement, never condition→supplement. Fix: ALWAYS run condition-anchored search axis in parallel.

LLM-Specific Failure Modes (updated Feb 2026)

Cross-model validation: For high-stakes bio claims (Grade 1-3 evidence affecting clinical decisions), route the same evidence through a second model as independent assessor. Different models have different fabrication patterns — Claude invents plausible-but-wrong citations, GPT invents complete fake studies. Cross-checking catches both.

Recitation Before Synthesis

Before grading evidence or writing conclusions, recite the key evidence items verbatim — restate the study name, N, effect size, and population for each Grade 1-5 source you're relying on. This combats lost-in-the-middle effects when working with many sources (Du et al., EMNLP 2025: +4% accuracy, training-free).

Don't summarize — recite. The act of restating forces attention back to the actual data before the synthesis step where hallucination risk is highest.

Self-Audit Checklist

After any bio research output:

• Every number has a source (DOI/PMID/URL)
• No study cited that you haven't verified exists
• In vitro/animal evidence NOT used to justify clinical recommendations
• Genetic associations include OR, CI, population, penetrance
• "Cannot/always/never" replaced with quantified ranges
• Industry-funded studies flagged
• Supplement doses cite the study they come from
• Genotype→dosing claims have CPIC/DPWG level or labeled INFERENCE
• Confidence ratings are honest
• Counterarguments section exists and is substantive
• New interpretation ideas passed the caller test
• New interpretation ideas are not just renamed duplicates of earlier operators
• Cautions/limiters were not promoted into runtime objects without decision impact

PGx Quick Reference

Justified: CPIC Level A/B, PharmGKB Level 1A/1B. Not justified (but LLMs do it): Single GWAS hit OR<2.0→dose recommendation; nutrigenomic SNP→supplement dose; variant without replication in user's ancestry.

Key databases: CPIC (cpicpgx.org), PharmGKB, ClinVar, DPWG, gnomAD.

Non-Paper Evidence (acceptable, labeled)

• Regulatory: FDA/EMA monographs, drug labels, safety communications
• Grey literature: ClinicalTrials.gov entries, conference posters, preprints [PREPRINT]
• Independent testing: ConsumerLab, Labdoor, third-party CoAs
• PGx databases: PharmGKB, ClinVar, CPIC guidelines
• Operational: formulation stability, supply chain, product CoAs