Runs

Every run this project recorded decision by decision. Each viewer replays one ledger: the circuit's simulated activity on measured MaleCNS neuron skeletons for each decision, what was proposed, and what the evaluator said.

Generator run on viewMushroom body, 15 SeptemberE79: the fly generator on the 5,311-neuron mushroom-body circuit, re-recorded after the screen was rebuilt and rewarded on the conservative bound drive + sigma; 60 rounds of 4 proposals. The ledger holds one seed; AUC_Q and distinct finds are the means over that invocation's seeds.5,311 neurons · 240 decisions · AUC_Q 20.1 · 51.9 distinct finds · deepest in the ledger −90 meV/atomOpen the viewer →

Two replays of the current configuration with a per-decision ledger (E224 arm A, seed 0, and the stalled seed 3) are queued behind E224's last arms. Until they land, the generator run on view is fly_generator_v1, an earlier configuration on the 5,311-neuron mushroom-body circuit.

Earlier generator runs

Recorded with RunLog (stage_b.py --log-run) while the generator was being built. Each line is what the log says the run was.

Whole brain, unfinished, 16 September wholebrain_v1
Recorded 15:13 to 15:41 on the whole connectome; the ledger stops at 484 decisions and was never closed (status RUNNING). Its date and size fit E139's fly arm (150 rounds of 4 per seed), for which the log records no result. E139
164,506 neurons · 484 decisions · deepest in the ledger 5 meV/atom
Open viewer →
An early run that collapsed: whole brain, 16 September fly_wholebrain_v1
E140: the fly on the whole connectome, and the search collapsed: one distinct find in 800 proposals (AUC_Q 0.4, the mean over its two seeds). It searched the twelve-element space, so 666 of the 800 proposals in this ledger carry Co, Cu or Ni, outside the nine elements of the goal; 344 were paid nothing, almost all for having fewer than five elements, and two qualified. E140 also changed the reward (novelty multipliers) at the same time, and E144 draws no conclusion about the whole brain from it. E140
164,506 neurons · 800 decisions · AUC_Q 0.4 · 1 distinct find · deepest in the ledger −53 meV/atom
Open viewer →
Trace, run alone, 19 September det_trace_alone
Five rounds of the fly, seed 0, run alone in its process. Beside the same trace run after a CEM arm it showed the reward depending on what had run before: the same five proposals, the fifth scored 43.8 here and 44.6 there. Fixed in 247ed9d (the ordering estimate's shared random generator). E201
164,506 neurons · 20 decisions · AUC_Q 0.0 · 0 distinct finds · deepest in the ledger 0 meV/atom
Open viewer →
Trace, run after CEM, 19 September det_trace_aftercem
The same five-round trace run after a CEM arm in the same process; its fifth verdict differs from the run-alone trace (44.6 against 43.8). The pair located the reward's coupling to process history, fixed in 247ed9d. E201
164,506 neurons · 20 decisions · AUC_Q 0.0 · 0 distinct finds · deepest in the ledger −12 meV/atom
Open viewer →

The forward-forward era, 11 to 15 September

A controller on the connectome chose which of 64 fixed alloys to evaluate next: forty MACE screens, then ten Quantum ESPRESSO calculations. That approach was abandoned; the Logbook says why.

Run	Controller graph	Objective	Evaluations	Acceptance
`forager_rhea_v1`	16,384 MaleCNS neurons	ΔH_mix, 0 K, unrelaxed	40 MACE · 10 QE	205 checks, all pass	Open viewer →
`forager_rhea_wholebrain_v1`	163,972 neurons · 6,143,838 edges	ΔH_mix, 0 K, unrelaxed	40 MACE · 10 QE	205 checks, all pass	Open viewer →
`forager_rhea_wholebrain_T1500K_P5GPa_v1`	163,972 neurons · 6,143,838 edges	ΔG_mix at 1500 K, 5 GPa	40 MACE · 10 QE	207 checks, all pass	Open viewer →

Reading a viewer

The dark panel draws measured MaleCNS neuron skeletons, each coloured by the magnitude of its simulated rate when the shown decision was scored, before the evaluator answered and before that decision's lesson. The run itself used every neuron of its circuit. The panel draws the neurons of that circuit that have a measured skeleton among the 317 the repository holds, matched by MaleCNS body ID; each viewer's caption says how many that is. The colours are simulated rates of this model, not measured fluorescence or spikes.

A generator decision is one proposal. RunLog records the driving force implied by the reward the generator was paid, d = 25 ln(1/p − 1) meV/atom at 90 K; with the novelty multipliers off, as in every run from E201b on and in E224, that is the rung-0 verdict itself. A proposal the verifier paid nothing for (p = 0) shows as the transform's ceiling, 518 meV/atom. Generator runs have no crystal cell, so their ledgers show no cell render.