CYCLE 5 - D03 Locked Taxonomy Spec v0.1

Date: 2026-05-01
Theorem: D03
Claim type: taxonomic
Evidence class target: L1 empirical external (paired pre/post)

Claim under test

"Informational loss of forgetful functors is classifiable into preservative vs degenerative regimes."

Core correction from Cycle 3-4

Cycle 3-4 showed post-fail criterion drift risk.
Cycle 5 enforces preregistered locked taxonomy with explicit no-rescue clause.

Mandatory data structure

Each sample must be an explicit pair:

1. D_pre
2. D_post = U(D_pre) with declared forgetful functor U
3. functor family label
4. context label

Single-instance sentence proxies without explicit U mapping are invalid for D03.

Primary metrics (pre-registered)

1. Delta_Coh = Coh(D_post) - Coh(D_pre)
2. Delta_Phi = Phi(D_post) - Phi(D_pre)
3. Loss_index = 0.5 * abs(Delta_Coh) + 0.5 * abs(Delta_Phi)
4. Weight lock: (w_Coh, w_Phi) = (0.5, 0.5) is fixed for Cycle 5 and cannot be changed on same data.

Class partition (pre-registered)

1. preservative: Loss_index <= t_low
2. degenerative: Loss_index >= t_high
3. ambiguous: t_low < Loss_index < t_high

All thresholds fixed before first run.

Confirmation criterion (pre-registered)

D03 is pass_candidate only if all are true:

1. Both classes (preservative, degenerative) are non-empty in at least 2/3 contexts.
2. Ambiguous share remains below max_ambiguous_rate in at least 2/3 contexts.
3. Coh and Phi both participate in classification (no single-axis collapse).

Reject conditions (pre-registered)

D03 is reject_candidate if one or more hold:

1. No explicit U mapping and no pre/post pairs.
2. Classification changes criterion/threshold on same dataset after first fail.
3. One class is structurally unreachable under preregistered formula/range.
4. Coh is absent from effective decision logic.

Review conditions (pre-registered)

D03 is revise_needed if:

1. both classes appear but stability is fragile across contexts, or
2. ambiguous share is borderline around threshold, or
3. independence checks reveal family imbalance risk.

Anti-pattern lock

1. No post-hoc rescue on same data.
2. No variable switch (drop to loss, or equivalent) after fail without new preregistration + new cycle.
3. Any rescue attempt triggers automatic reject_candidate for that cycle.