# Wave 3 method and evidence note

Wave 3 is a finite, deterministic cellular-automaton census. The board is 64×40, updates are synchronous, and the local rule is B3/S23 with a 3×3 Moore neighborhood. The frozen protocol evaluates the exact table, an exact table with raw ID9, a raw table with ten canonical corrections, and a raw table with all 41 corrections in the D4 support union. Eight D4 orientations, four phases, wrap and dead boundaries, and fixed one-cell additions and removals were assessed.

The clock has two separate meanings. Eligibility requires intact whole-grid motion through **absolute generation 64**. Damage is applied at absolute generations 8, 9, 10, or 11. The primary endpoint is **64 synchronous updates after damage**, so its corresponding absolute endpoint is 72, 73, 74, or 75. The early and secondary endpoints are 32 and 256 updates after damage. This distinction is part of the public evidence JSON and should be preserved in captions and tables.

Each eligible phase contributes 25 fixed row-major sites: five removals and twenty additions. A group retains the undamaged control, a sham, and all 25 one-cell interventions across 257 post-damage frames. The classifier requires a whole-grid five-cell motion signature over a 17-frame window, phase progression, exact transitions, and the transformed four-step displacement. The primary census contains 256 labelled eligibility cases, 160 eligible groups, and 4,000 one-cell damage cases.

At the 64-update wrap endpoint, the exact table has 32 moving cases out of 800; exact+ID9 has 10/200; raw+canonical-ten has 20/200; and raw+all-D4 has 4/800. The corresponding removal/addition rows and both boundary conditions are recorded in `wave-3-evidence.json`. These are labelled deterministic counts, not independent samples or probability estimates. Matched comparisons are therefore reported on identical damaged bitmaps: exact+ID9 gives 10 versus exact’s 8, canonical-ten gives 20 versus 8, and all-D4 gives 4 versus 32 in the shared wrap cohorts.

The mechanism witness is a saved-transition replay, not a new trial. In the ID9 case, the deleted cell remains dead while a neighboring birth restores the next phase; the full mutant next bitmap equals the undamaged exact control. The independent check finds conflicting required outputs for the same neighborhood ID9 (0 for the healthy context and 1 for the restoring birth), establishing an exact one-update conflict for a fixed memoryless 3×3 table. It does not establish a universal impossibility result for temporal memory, larger neighborhoods, or other injury classes.

The source ledger, active table hashes, matched mappings, validation counts, and sanitized source hashes are in [wave-3-evidence.json](wave-3-evidence.json). The full 4,000-row case ledger and multi-megabyte annotation/frame artifacts remain in the source experiment bundle and are represented here by exact SHA-256 hashes. No provider or site calls were used; the runtime record reports 160/160 groups and 4,000/4,000 damage cases complete.

This is digital rule-experiment evidence only. It does not support claims about biological cognition, learning, intention, or mechanism.