Skip to content

0007 — Idealization parity is statistical, asserted against a frozen tolerance

  • Status: accepted
  • Date: 2026-06-24
  • Deciders: bioedca
  • PRD anchor: §7.4 (idealization parity), §11.2, §8 NFR-VALID(b)
  • Milestone: M0.5 (freeze) → M2, M6 (inherit)

Context and problem statement

Tether's one-click vbFRET (and later consensus/ebFRET) must agree with standalone tMAVEN. But tMAVEN's variational inference self-reseeds, so runs are not bit-identical even on the same input. How do we define and enforce "parity" without chasing nondeterminism?

Decision drivers

  • Variational Bayes HMM results vary run-to-run; exact equality is impossible.
  • A trustworthy, non-flaky gate must compare against a measured spread, not a guessed threshold.
  • M2 (one-click) and M6 (consensus/ebFRET) both depend on this gate.

Considered options

  • A. Statistical parity vs a frozen tolerance — measure the cross-seed spread from ≥20 standalone-tMAVEN runs on committed SMDs, freeze four numbers (state count agreement, per-state mean ΔE, Viterbi per-frame agreement, relative ΔELBO) into §11.2 + schema/parity_tolerance.json, and assert against them.
  • B. Bit-exact comparison (impossible — self-reseeding).
  • C. Pin tMAVEN's RNG seed (brittle; not how the standalone runs).

Decision outcome

Chosen option: A. M0.5 S4 ratifies and freezes the tolerance; the sidecar/parity CI job asserts against the frozen numbers and never recomputes them. M2 and M6 inherit the tolerance by reference and cannot sign off until it is frozen (a hard gate).

Consequences

  • Good: a stable, non-flaky parity gate grounded in measured variance.
  • Trade-off: the freeze is a prerequisite milestone gate (M0.5 S4) blocking M2/M6.
  • Follow-up: provisional defaults (≥90% / ≤0.02 / ≥95% / ≤0.01) are replaced by measured values at M0.5 S4 — done (the measured cross-seed spread was negligible, confirming the defaults; see ADR-0009).

More information

PRD §7.4, §11.2, §8 NFR-VALID(b); PLAN §1.1 hard gates, M0.5 S4. The metric definitions, freeze policy, and measured result are recorded in ADR-0009.