Freeze-plasticity onset

Exploratory protocol

No completed result packet for this experiment is committed under experiments/results/. This page documents the runnable method, not a measured onset or mechanism.

Question

If Falandays plasticity is frozen at tick $T$, how does later task behavior and population activity depend on $T$? The intervention can distinguish a useful learned state from a configuration that still requires online adaptation, provided that the time grid and score have enough resolution.

Current protocol

julia --project=. experiments/run.jl freeze_onset

The registered default uses:

• tasks :tracking, :wall, and :pong;
• freeze ticks [1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377];
• seeds 0:5;
• a 600-tick post-freeze window;
• :freeze_plasticity as the live intervention;
• a matched full-learning run for each task, seed, and horizon.

For each freeze tick, freeze_sweep aggregates normalized task score, raw score, and population rate. onset_tick defines a descriptive knee as the first sampled tick at which the mean frozen-score curve reaches halfway from its earliest value to the median of the last three sampled values, provided that span exceeds 0.03.

Method limits

• The default grid does not sample ticks 6 or 7. It therefore cannot support a precise claim that onset occurs at either tick.
• The knee is computed from the seed-mean curve; the current output does not estimate uncertainty in onset across seeds.
• Normalized scores remain task-specific bookkeeping quantities. A ceiling-saturated or low-resolution task score can hide a real activity transition.
• Population rate can diagnose silence or saturation, but neither state alone identifies why behavior changed.
• Different freeze ticks imply different total rollout lengths (T + window). The matched full-learning control handles this within a cell, but comparisons across $T$ should retain that fact.

Evidence required for a result

A promoted study should predeclare the target task and onset resolution, sample every tick in the suspected transition region, preserve per-seed paired rows, report an interval for the onset or explicitly call it unresolved, and include resolved config and complete provenance. Claims about long-term persistence after freezing require a separately declared long post-freeze horizon.

Source: experiments/freeze_onset.jl and experiments/harness.jl.