Grading and honesty

Every claim carries a label: code output (reproducible), consistency, observation (cited) or interpretation (relabelled 2026-09-29 from [V] verified / [F] forced), and [O] open. Every clinical magnitude is [O]. The discipline is mechanical: an L3-honesty gate, a fail-closed forbidden-claim scanner, and an irreproducibility ledger of nine open classes. Reproduce all gates with repro/run_all.py — 11/11, drift 0.

Labels are code output (reproducible), consistency, observation, interpretation, and [O] open. The discipline is enforced mechanically by an L3-honesty gate (fail-closed), a forbidden-claim scanner (fail-closed), and an irreproducibility ledger that aggregates nine open classes with reasons. The 27 reads reproduce bit-for-bit (drift 0, offline; corr(γ,GC) = 0.99898 is a consistency property of the SantaLucia table, not a finding).

Three grades

• code output reproducible (formerly [V]) — bit-reproducible from public promoter sequence (the 27 γ reads; drift 0, offline; corr(γ,GC) = 0.99898 over the full set is table-intrinsic, a consistency check).
• code output / interpretation (formerly [F]) — the ordering by |h_sp| is a code output (monotone in γ); the lever placement and direction are interpretations anchored to cited channel biology and agents (observations), not forced by the reads.
• [O] open — needs external laboratory / clinical data; asserted nowhere. Every clinical magnitude is [O].

The discipline is mechanical

An L3-honesty gate verifies that for every upstream-sensitiser target (NGF, NTRK1, CALCA, CALCB, CALCRL, RAMP1) the receptor/network mechanism is graded [O] cited-biology-never-derived. A forbidden-claim scanner fails the build closed if any dosing, synthesis, efficacy, or safety claim appears in the asserted text. An irreproducibility ledger lists all nine open classes with their reasons.

Reproduce every gate: repro/run_all.py (11/11 checks, drift 0). The full open ledger is IRREPRODUCIBILITY_LEDGER.md in the package.