Skip to main content

Module independence

Module independence 

Source
Expand description

Measured validator dependence (PSP-9 system 8, Paper III Lemma 5.1 and Corollary 5.2).

“Correlation among redundant validators is measured, never assumed away.” This module computes, from matched labeled verdicts, each validator’s miss rate q_i, the pairwise miss correlation κ_ij, the empirical pairwise joint miss, and the conservative conjunctive-ensemble bound ρ_eff = min_{i<j} u_ij over pairwise joint-miss upper confidence bounds.

Statistics discipline (Gates R and T):

  • Matched strata. q_i, q_j, κ_ij, and the joint miss for a pair are all computed over the same jointly evaluated candidate set.
  • Degenerate marginals (q ∈ {0, 1}) use the empirical joint miss directly, as Lemma 5.1 requires; they are never assigned κ = 0.
  • Exact only at n = 2. For n > 2 validators the bound is deliberately conservative; it is never tightened by inferring higher-order joints from pairwise correlations, and n is reported beside it. A three-validator bound tighter than its best pair would be a defect, not a better certificate.
  • Point estimates are diagnostics. A certified ρ_eff uses upper confidence bounds and a minimum matched sample count; below it the status is insufficient evidence, not a number.
  • No energy attenuation. Correlation controls the reported risk bound; it never rewrites the fixed quadratic energy weights. The former attenuate_weight helper is removed for that reason.

Structs§

IndependenceStats
Independence statistics computed from the verdict ledger.
PairStats
Matched-stratum statistics for one validator pair.
VerdictRecord
One validator’s verdict on one candidate: did it miss an unsafe state?

Enums§

EnsembleCertification
Certification status of the ensemble bound.

Constants§

DEFAULT_MIN_PAIR_SAMPLES
Minimum matched samples per pair before a certificate may be issued.

Functions§

compute
Compute independence statistics from a verdict ledger.
compute_with_floor
Compute with an explicit matched-sample floor for certification.
miss_std
Standard deviation of a Bernoulli miss indicator: σ = sqrt(q(1-q)).
pairwise_joint_miss
Lemma 5.1’s exact pairwise joint miss: q_i q_j + κ σ_i σ_j.