Historical: Tracking plasticity and branching diagnostics
This registered experiment asks how stopping plasticity at different times changes the
declared Tracking outcome and estimated branching dynamics. It uses the canonical
falandays_tracking preset: the paper-matched Falandays node configuration with 200 nodes
on the standard Tracking task.
The resolved node settings are leak = 0.25, lrate_wmat = 1.0,
lrate_targ = 0.01, threshold_mult = 2.0, targ_min = 1.0,
input_weight = 0.75, weight_init_std = 1.0, link_p = 0.1, excitatory
weight initialisation, Bernoulli topology, no activation rectification, and no mask repair.
The retained resolved.toml repeats these values and the resolved task settings for every
operation record.
| Condition | Plasticity rule |
|---|---|
freeze_tick_1 | Disable recurrent and target updates before the first step |
freeze_tick_1800 | Run steps 1–1799 with plasticity, then freeze before step 1800 |
continuous_plasticity | Keep both update mechanisms active for all 7200 steps |
Each condition uses the same root seed and 60 paired, independently constructed blocks.
Every block contains one trial. The declared seed streams are topology and world, the
reset policy is full, and there is no warm-up interval.
The primary task result is the full-run track_score, reported with its raw and normalised
values. The registered tracking_plasticity_diagnostics analysis also produces:
- 400-tick branching windows every 100 ticks, using the MR estimator with
kmax = 20; - fit quality, a legacy through-origin estimate, population rate, heading error, object visibility, and spectral radius aligned to those windows;
- branching ratio and change in branching ratio binned by absolute heading error.
Spectral radius is recomputed every 100 ticks. The record retains trial scores, seed ledgers, scalar analysis statistics, and aggregate series summaries. It does not retain raw tick traces. These diagnostics can describe estimator behaviour; they do not by themselves establish that the network is critical.
The archived instructions retain the protocol validation command. See version history for its implementation boundary.
The evidence state is planned, and no result record is attached yet. After a run from a
clean, committed source revision, its standard records can be listed here and used to draw
the score, time-series, and heading-error charts directly from the retained CSV tables.
Read the archived protocol bundle.