e7

Documentation

Overview

Command e7 measures evaluation hypothesis E7 (§12.3): the proportion of sampled verdicts reconstructable from their evidence alone, with no query back to the store.

The test is the strict form R5 implies: for each sampled evaluated verdict, the p-value is recomputed from nothing but the verdict's own evidence card — the sufficient statistics of §6.4, §7.7 and §8.3 — and reconstruction succeeds when the recomputed value matches the verdict's to within 1e-9 relative. Detectors whose evidence cannot drive the recomputation fail the sample, which is the point of measuring rather than asserting.

Reconstruction routes, per detector:

• novelty: p̂ from (4) via n_v, N, K, α; the full tail mass of (5) needs every value's count, which the card deliberately does not carry, so reconstruction validates (4) exactly and accepts the verdict when the observed value is modal (P = 1) or unseen (P = p̂(∅)), the two cases (5) pins to card values. Others count as partially reconstructable and are reported separately.
• timing: f̂(φ) from (7) via the moments c_h, s_h, W and κ on the card, then (9) on the G-point grid stated by the card.
• volume: the upper tail of (11) via a, b, ρ and k_obs.
• cooccurrence: λ from (14) or (15) via k_i, k_j, m (and D_r, D_s, m_rs when a partition priced the pair), then the Šidák step of (16) via T.