CYCLE 6 Benchmark Independence Protocol v0.1

Date: 2026-05-05
Cycle: CYCLE_6_2026-05-05
Theorem: A01

1) Independence definition

A Cycle 6 benchmark is independent from Cycle 3 iff all are true:

1. No shared record identifiers.
2. No shared normalized content hashes.
3. No shared pre-labeled trajectories reused as-is.
4. Context generation process is documented and new.

2) Required evidence

1. A01_dataset_manifest_v0_1.md with source URLs or generation recipe.
2. A01_cycle3_overlap_report_v0_1.md containing:
   • record ID overlap count
   • hash overlap count
   • overlap percentage
3. Signed auditor check on overlap report.

Acceptance threshold:

1. Record overlap count = 0
2. Hash overlap count = 0
3. Overlap percentage = 0.0%

3) Hashing method

Canonicalization rules:

1. UTF-8 encoding
2. LF newline normalization
3. Trim trailing whitespace per line
4. SHA-256 digest

4) Context stratification rule

Cycle 6 contexts must not be decorative rescaling of a single distribution. Each context declaration must include:

1. Distinguishing features
2. Expected difficulty
3. Perturbation profile
4. Dataset slice path

5) Failure policy

If independence proof fails:

1. Cycle execution status -> blocked
2. A01 cannot be promoted
3. New benchmark generation is mandatory before rerun