Experiments · E8

Does reading a named group of output neurons give a stable ranking of alloys?

Partly. It removed a size artefact, but different random seeds still ranked the alloys differently, correlations from −0.76 to +0.61.

In the log: What a named readout fixes, and what it does not

recordedDate 2026-09-12, as written in the logunclassified0 predictions · 1 result paragraphEXPERIMENTS.md lines 234–270

Results

EXPERIMENTS.md · line 248

Result.

The full record

EXPERIMENTS.md · lines 234–270

E8 — What a named readout fixes, and what it does not

Date 2026-09-12 · Question Does reading a named population instead of the mean over all neurons recover a usable policy? · Provenance forager/mushroom.py, …/stability.py

Method. MushroomBody drives the full 166,700-neuron graph at the olfactory projection neurons, weights by contact count, signs by consensus transmitter, and scores by a signed fixed projection over the 97 MBONs. Compared against LocalFF's mean of h^2 over all neurons on the same graph, seed and 64 real RHEA descriptors. Populations resolved: KC 4,064, MBON 97, PAM 316, PPL 24, APL 2, PN 595 (family-matched, correctly excluding 213 non-olfactory wedge neurons); plastic edges 61,210; inhibitory fraction 0.356; KC sparsity ~0.10.

Result.

Part of the null is a units artifact. The softmax applies a fixed temperature to a score of arbitrary scale. Standardising scores across candidates first gives a best/worst policy ratio of 27x (mean h^2) and 21x (MBON), against a uniform 1/64 — from a policy previously confined to [0.013, 0.019].
The MBON readout breaks the norm degeneracy: r(score, ‖x‖) moves from +0.99999 to about -0.5.
But the ranking is not yet a circuit property. Across four seeds, pairwise Spearman between rankings spans -0.76 to +0.61 (mean |rho| 0.47), and one seed flips r(score,‖x‖) to +0.09.

Normalisation was not the cause: per-row, global-percentile and spectral-radius scaling all give the same result. Spectral scaling is kept because it bounds a recurrent network with a single constant and so preserves measured ratios, where a per-row budget divides a fixed sum across every input and erases them.

Interpretation. The readout was a real defect and fixing it removed the ‖x‖ degeneracy, but what replaces it is encoder noise routed through the circuit. The encoder is now the binding constraint: twenty abstract descriptors are mapped onto 595 projection neurons at random, so each seed is a different nose, where in a fly a PN is a specific glomerulus for a specific chemical class. No claim about the wiring is testable until that mapping carries meaning.