CYCLE 6 A01 Real Pipelines Spec v0.1

Date: 2026-05-05
Theorem: A01
Cycle ID: CYCLE_6_2026-05-05
Evidence mode: real_pipeline_execution

1) Claim target

Claim: Compositional pipelines with high ordinative coherence can show process-level quality advantage without requiring output-score inflation.

Primary falsifier: No process-level advantage is observed under matched-output constraints in independent contexts.

2) Execution architecture

Two independent implementations must exist:

1. P_cls: classical-validity-only pipeline.
2. P_ord: ordinative pipeline with explicit coherence gate.

Independence constraints:

1. No shared scoring function for trajectory quality.
2. Shared input contract only (same tasks/contexts), but distinct decision logic modules.
3. Separate code paths in repository and separate lock hashes.

3) Context set

Minimum:

1. CTX_A low perturbation
2. CTX_B medium perturbation
3. CTX_C high perturbation

Each context must be represented by benchmark data declared independent from cycle3_inputs.

4) Primary metrics (pre-registered)

For each context and each pipeline:

1. Delta_cum
2. Err_sem_final
3. mean_Coh_trajectory
4. trajectory_length
5. reject_rate

Matched-output indicator: abs(Err_sem_final_ord - Err_sem_final_cls) <= eps_err_match

Default threshold: eps_err_match = 0.01

Process-advantage indicator: mean_Coh_trajectory_ord - mean_Coh_trajectory_cls >= eps_coh_adv

Default threshold: eps_coh_adv = 0.03

5) Decision criteria

pass_candidate requires all:

1. At least 2/3 contexts with Delta_cum_ord < Delta_cum_cls.
2. At least 2/3 contexts with mean_Coh_trajectory_ord > mean_Coh_trajectory_cls.
3. At least one matched-output subset (Err diff <= eps_err_match) in at least 2/3 contexts.
4. In matched-output subsets, process advantage holds (Coh diff >= eps_coh_adv) in at least 2/3 contexts.
5. No gate/policy violation.

revise_needed:

1. Criteria partially satisfied but no reject trigger.

reject_candidate:

1. Any hard reject trigger from execution rules.

6) Non-promotable conditions

Cycle 6 is non-promotable if any occurs:

1. Heuristic proxy branch formula is reused as decision basis.
2. Benchmark independence proof is missing.
3. P_cls and P_ord are not implementation-distinct.
4. Lane/report artifacts are incomplete.

7) Mandatory artifacts

1. instances/A01_OCT_PROXY_PREVALIDATION_SHEET_v0_1.md
2. seals/A01_PREREG_SEAL_v0_1.json
3. claim_locks/A01_claim_lock_v0_1.md
4. threshold_locks/A01_thresholds_lock_v0_1.md
5. formula_locks/A01_formula_lock_v0_1.md
6. dataset_manifests/A01_dataset_manifest_v0_1.md
7. script_locks/A01_script_lock_v0_1.md
8. contexts/A01_contexts_v0_1.md