Could the fly's mushroom body choose which alloys deserve an expensive calculation?

PREREGISTER_acquisition.md · lines 1–106

Pre-registration: the fly as the thing that decides what to evaluate next

The scarce resource in materials discovery is the expensive evaluation, not the candidate. Whatever decides where that budget goes is what discovers the material. This tests whether the mushroom body can be that decider. Written before the benchmark exists.

The claim, stated so it can fail

Scoring candidates with the measured mushroom body - learned value from dopamine plus a familiarity bonus from the Kenyon-cell code - finds better alloys per oracle call than conventional Bayesian optimisation, and does so because of the measured wiring.

Two hypotheses, deliberately separated, because one can hold without the other:

1. Useful optimiser. Better candidates per unit budget than standard acquisition.
2. Connectome contribution. The measured wiring beats matched rewirings of itself.

What this is not

Value plus novelty is not an upper confidence bound. A Kenyon-cell familiarity score records exposure; it is not an estimate of what an observation would teach. A Gaussian process supplies a function prior, an observation model and a posterior covariance, and none of those exist here. The algebraic resemblance between value + beta * novelty and UCB carries no guarantee, and the word is not used.

Three things are kept distinct and are never used as synonyms: the Kenyon-cell code (a similarity-preserving feature map), familiarity memory (a record of what has been observed), and epistemic uncertainty (what could still be learned).

The trap this design is built around

Scoring a candidate must not teach the circuit that it has evaluated one. Habituation in the animal fires on exposure, and a pool of 8,192 candidates scanned once would leave every one of them familiar and the novelty term dead. Therefore: plasticity is frozen while scoring, transient state is reset between candidates, ranking is invariant to the order candidates are enumerated in, and familiarity updates only when an oracle observation actually returns.

Separately: a cheap evaluation must not discharge the uncertainty attached to an expensive one through a shared familiarity memory.

Protocol, frozen before any run

Arms

The fly: measured mushroom body, value from MBON output learned by dopamine reward-prediction error, plus a familiarity bonus from the Kenyon-cell code.

Conventional: Gaussian process with LogEI, UCB and Thompson sampling; random; maximum composition-distance coverage.

The controls that decide attribution:

Rewirings preserve connection blocks, cell classes, compartment membership, signs, weight distributions, learning masks and the number of adjustable parameters, and are checked for comparable sparsity and non-saturation - a shuffle that is silent or unstable is not a control. Encoder, thresholds, reward scaling and exploration schedule are frozen before evaluation and are identical for every arm, and development tuning is given comparable effort across arms.

The build gate

Continue only if simultaneous 95% confidence bounds establish both:

1. at least 10% lower A than the strongest conventional comparator chosen on development runs; and
2. a positive advantage over the rewired ensemble, accounting for graph-to-graph variation.

Ten percent is an engineering threshold chosen in advance, not a biological constant.

Failing either gate kills this implementation. An inconclusive result is not permission to add machinery. Passing only the practical gate supports a generic acquisition method with no anatomical claim; passing only the wiring gate supports an anatomical effect with no demonstrated usefulness; passing both earns a higher-fidelity replay.

What may not be claimed

Fifty starts on one cluster-expansion landscape establish robustness to initialisation on that landscape, and nothing about materials discovery in general. A single DFT energy is not an observation of short-range order or of a transition temperature; those need a stated configurational and thermal protocol, and the whole protocol is charged. Multi-fidelity allocation is a separate problem and is not attempted here: fifteen dopaminergic compartments provide capacity, not cross-fidelity correlations, and assigning compartments to fidelities would be an architectural choice needing its own controls.

MaleCNS supplies aggregate neurotransmitter predictions; synapse counts are not measured physiological gains. This is a connectome-constrained model with engineered dynamics, and is described as one.