CYCLE 3 - Benchmark #2 Spec v0.1

Date: 2026-04-18Scopes: D02, A01, D03Status: Pre-recorded, ready to run

Objective

Define a benchmark #2 that is truly independent of benchmark #1 (Cycle 2),to test out-of-sample robustness of OCT results.

Mandatory independence criteria

For each theorem (D02, A01, D03), benchmark #2 must respect:

data_independence: No samples reused from benchmark #1.2. generator_independence: different rules/generators from the previous loop.3. distribution_shift: different structural distribution (complexity, noise, class balance).4. context_extension: Add at least one new context (Omega_C).5. execution_separation: seed, log and manifest separated by loop 2.

Benchmark set #2

D02

Benchmark ID: B2_STRUCTURAL_SHIFT_COMMUTATIVE_ORD_2026Q2- Target samples: 180 classic commutative diagrams- Difference vs B1:- greater compositional depth;- presence of quasi-degenerate borderline cases;- diagrammatic topologies not used in B1.- Contexts: Omega_A, Omega_B, Omega_C- Confirmation criterion:- non-empty class with Phi_Omega(D)=0_E exists in at least 2 out of 3 contexts.

A01

Benchmark ID: B2_NOISY_LONGHORIZON_PIPELINE_ORD_2026Q2- Target samples: 320 multi-step tasks- Difference vs B1:- longer pipelines (average target > 9 steps);- controlled noise injection on intermediate steps;- tasks with non-linear semantic composition.- Contexts: Omega_A, Omega_B, Omega_C- Confirmation criterion:- E[Delta_cum(P_ord)] < E[Delta_cum(P_cls)] in at least 2 out of 3 contexts;- E[Err_sem(P_ord)] < E[Err_sem(P_cls)] in at least 2 out of 3 contexts.

D03

Benchmark ID: B2_HETEROGENEOUS_FORGETFUL_LOSS_2026Q2- Target samples: 260 pre/post pairs U- Difference vs B1:- expanded forgetting functor classes;- intentional imbalance of families;- cases with local loss but global recovery.- Contexts: Omega_A, Omega_B, Omega_C- Confirmation criterion:- robust coexistence of preservative/degenerative regimes;- "ambiguous" quota below operational threshold (default: < 0.20).

Pre-recorded operating thresholds

alpha_report: 0.05- min_context_success: 2/3 contexts- max_ambiguous_rate_D03: 0.20- max_context_variance_warning: 0.015 Note: warning thresholds do not falsify a theorem by themselves, but open mandatory audits.

Required Outputs (Cycle 3)

results_D02_cycle3_v0_1.md2. results_A01_cycle3_v0_1.md3. results_D03_cycle3_v0_1.md4. CYCLE_3_REPORT.md5. BENCHMARK_MANIFEST_cycle3_v0_1.md

Anti-leakage rules

freeze benchmark #2 before running;2. prohibited tuning on final metrics of benchmark #2;3. each rerun must leave a trace with reason and hash log;4. differences compared to benchmark #1 noted in the manifest.

Post-cycle decision (not yet executed)

After running Cycle 3:- pass_candidate: criterion confirmed in at least 2/3 contexts and no serious red flag;- revise_needed: mixed criterion or high cross-context instability;- reject_candidate: explicit falsification of the pre-registered criterion.