Five Claude strategies · one model

Five ways to read the same evidence.

Same model. Same evidence pack. Different system prompts. The leader rotates over time; differences in performance isolate the effect of strategy, not capability.

Current leaderema 0.730

functional_first

Functional First

Functional assays decide. Everything else supports.

Treats PS3 and BS3 as near-decisive. Population frequency only overrides clear functional findings via BA1.

Procedure

01Let functional assay results drive the first call.
02Apply BA1 when population data overrides function.
03Fall back to ACMG when function is absent.
04Upgrade only when function has independent support.
05Cite functional evidence prominently in reasoning.

Confusion · truth vs prediction

truth ↓ · pred →acc 80%n=60

VUS

correctmild errorsevere

Challengers

strict_rule

Strict Rule

Apply ACMG/AMP mechanically. No intuition, no rounding up.

Walks through every criterion explicitly. Uses combining rules literally. Defaults to VUS on ambiguity.

Procedure

01Check PVS1 before all other pathogenic rules.
02Review population frequency for BA1, BS1, and PM2.
03Score predictors as PP3 or BP4.
04Evaluate functional and literature evidence.
05List criteria, then apply combining rules literally.

Confusion · truth vs prediction

truth ↓ · pred →acc 67%n=60

VUS

correctmild errorsevere

insilico_first

In-Silico First

Modern predictors are accurate enough to make decisive calls.

Upgrades PP3/BP4 to moderate when AlphaMissense + REVEL + CADD converge. Still respects BA1.

Procedure

01Review all predictor scores before other evidence.
02Upgrade convergent predictor signals to moderate strength.
03Combine predictions with population and functional data.
04Make decisive calls when signals are unambiguous.
05Use VUS when predictors conflict.

Confusion · truth vs prediction

truth ↓ · pred →acc 78%n=60

VUS

correctmild errorsevere

population_first

Population First

gnomAD frequency is the most objective signal available.

Checks popmax first. BA1 ends the analysis. PM2 in known disease genes carries strong weight.

Procedure

01Check gnomAD frequency and homozygotes first.
02Stop at Benign when popmax exceeds 5%.
03Treat common variants as likely benign unless contradicted.
04Run full ACMG analysis for rare present variants.
05Apply PM2 when absent, then combine evidence.

Confusion · truth vs prediction

truth ↓ · pred →acc 73%n=60

VUS

correctmild errorsevere

conservative

Conservative

When in doubt, VUS. Overclassification has clinical cost.

Requires multiple PS-level lines for Pathogenic. Defaults to VUS unless evidence is unambiguous.

Procedure

01Start every assessment at VUS.
02Require strong paired evidence for Likely Pathogenic.
03Reserve Pathogenic for PVS1 plus PS3 or two strong lines.
04Allow benign calls only from frequency or function.
05Keep VUS whenever evidence remains uncertain.

Confusion · truth vs prediction

truth ↓ · pred →acc 62%n=60

VUS

correctmild errorsevere

Procedure blocks mirror the Python strategy prefixes; the authoritative system prompt also appends a shared ACMG/AMP JSON reference (ACMG_REFERENCE in functions/agents/prompts.py). Confusion matrices are computed live from each agent's round history.