Historical: Planned capacity study

Stage 1 supplies the library, conformance controls, sparse recordings and decoding. Its small software checks do not estimate model performance. Stage 2 will determine which probes supply useful benchmark coordinates and at what difficulty.

The experiments/capacity-probes/v1/ bundles are ordinary ExperimentSpec values with evidence_state=planned. They start with one cell per family, including original delayed cue. They reuse fixed 200-node Falandays and SORN profiles and the exact saved Dense and Structured CTRNN development nominees. No new genome search is part of these proposals.

The archived instructions retain bundle validation and plan generation. These commands belong to the earlier implementation described in version history. Validation checks a configuration without generating scientific results.

Bundle	Proposed operation	Independent units
calibration	Input-gain sweeps, 16 blocks per gain	Fresh world and reservoir per block
native_pilot	Paired native comparison, 32 blocks per cell	Fresh world and reservoir per block
decoding	Profiles with 256/128/256 trial splits	Four wiring blocks per model and cell

Gain 1 in the native and decoding proposals is provisional. The proposal calls for a timed preflight, development-seed calibration and a new pilot protocol after selection. The archived pilot is explicitly uncalibrated and diagnostic. No confirmation seed set is included.

The archived generator covers all 64 difficulty cells. The proposal requires a justified subset and reports connections, internal states, integration work and runtime alongside node counts and performance.

Stage 2 should compare reference and negative controls, measure variance and sensitivity, then select a compact set using the archived coverage criteria. Pair worlds for causal controls and use applicable plasticity ablations for reversal. Confirmation and admission require a new frozen protocol and later human evidence review. Existing benchmark versions and the three visual tasks remain intact.

Explore when plasticity matters

The archived tooling generates native freeze-time curves with declared cells, scales, repetitions, cutoffs and freezing mechanisms. These are exploratory measurements, not sealed confirmation.

The default covers one cell from each family, with ten independent wiring blocks. Freezing after n completed adaptive ticks applies before tick n+1. Weight-only freezing retains target adaptation; complete freezing stops both mechanisms. Each episode starts afresh. This measures within-episode timing, not accumulated learning across episodes. The reversal branch fixes the reversal at round 48 so cutoffs align across blocks.

The archived reporting tool produces raw curves, adjacent paired slopes, standard errors, normalised outcome tables, complete records and a ZIP beside the report directory. The steepest observed interval suggests a midpoint for another exploratory run. It does not establish a unique knee; flat or chance-level curves may have no useful transition. The analysis verifies record checksums and paired comparisons before writing the page. Sending or publishing the ZIP is separate from generating it.