Experiments · E224
Not yet. Three of eight arms have finished and the result is not written; one seed stalled for 150 rounds.
In the log: has all eight arms; E225 and E226 pre-registered and queued (2026-09-22 06:2x)
book/assets/diagrams/exp/E224.svg).Arms F and G exist now (forager/search/generate.py): ElitistDiverse is Elitist carrying
the fly's three diversity devices — the 0.05 L1 proposal filter lifted out of FlyGenerator
into a shared _pick_novel so both use one filter, 24 walkers' worth of elite, and a uniform
Dirichlet restart after PATIENCE = 6 observations without a new best; RidgeSwarm is
FlyGenerator with the connectome head replaced by a ridge readout on the same input features
(E12's ridge on the concentrations under E224's broadcast encoder), refit on every observe.
Registered as elitist-diverse and ridge-swarm; 8 tests. So E224 runs A–H as pre-registered
and its A−G line — "what the wiring adds beyond a linear readout of the same inputs" — is the
decisive one for the abstract regardless of A−B.
E224 seed 3 — a cold-start stall, diagnosed post hoc (20:2x; a diagnostic, not a decision arm).
Qualifiers per 50-round window: real wiring s3 0 / 0 / 1 / 34, random head s3 0 / 0 / 0 / 1,
rewire s3 113 / 191 / 194 / 198, real s0 127 / 162 / 187 / 188. Same seed, same per-walker gains:
with the real and random heads the swarm finds nothing qualifying for 150 rounds; the rewired head
escapes at once; the late winners are 2–4-element compositions (s0 spreads over 1–7). The dumps keep
only winners, so the stalled rounds cannot be reconstructed. Replay (runs/e224diag_chain.sh,
queued behind E224 so one search runs at a time): arm A seed 3, E224's exact environment, with
--log-run (per-decision ledger). Questions, in order: (1) does the stall reproduce with the same
seed (if not, it is device non-determinism and the paired design needs repeats); (2) where the walkers
sit during rounds 1–150 (number of elements, distance to corners) and what fraction of proposals the
0.05 L1 filter rejects; (3) whether the readout's score correlates with reward before round 150.
This entry is written in 2 separate places in the log, shown here in log order.
Arms F and G exist now (forager/search/generate.py): ElitistDiverse is Elitist carrying
the fly's three diversity devices — the 0.05 L1 proposal filter lifted out of FlyGenerator
into a shared _pick_novel so both use one filter, 24 walkers' worth of elite, and a uniform
Dirichlet restart after PATIENCE = 6 observations without a new best; RidgeSwarm is
FlyGenerator with the connectome head replaced by a ridge readout on the same input features
(E12's ridge on the concentrations under E224's broadcast encoder), refit on every observe.
Registered as elitist-diverse and ridge-swarm; 8 tests. So E224 runs A–H as pre-registered
and its A−G line — "what the wiring adds beyond a linear readout of the same inputs" — is the
decisive one for the abstract regardless of A−B.
E224 seed 3 — a cold-start stall, diagnosed post hoc (20:2x; a diagnostic, not a decision arm).
Qualifiers per 50-round window: real wiring s3 0 / 0 / 1 / 34, random head s3 0 / 0 / 0 / 1,
rewire s3 113 / 191 / 194 / 198, real s0 127 / 162 / 187 / 188. Same seed, same per-walker gains:
with the real and random heads the swarm finds nothing qualifying for 150 rounds; the rewired head
escapes at once; the late winners are 2–4-element compositions (s0 spreads over 1–7). The dumps keep
only winners, so the stalled rounds cannot be reconstructed. Replay (runs/e224diag_chain.sh,
queued behind E224 so one search runs at a time): arm A seed 3, E224's exact environment, with
--log-run (per-decision ledger). Questions, in order: (1) does the stall reproduce with the same
seed (if not, it is device non-determinism and the paired design needs repeats); (2) where the walkers
sit during rounds 1–150 (number of elements, distance to corners) and what fraction of proposals the
0.05 L1 filter rejects; (3) whether the readout's score correlates with reward before round 150.
Built with PRISMWebsite and visualizations made using Claude