Paper · preprint · 2026
When does a biological circuit transfer?
Pre-registered, adversarially reviewed tests of Drosophila mushroom-body and ring-attractor models on text, chess, image, associative-recall and code-clone tasks
Abstract
We report a methods-and-negative-results study of two circuits of the Drosophila connectome, the mushroom body (MB) and the central-complex ring attractor, measured against trivial and strong baselines on text, chess, photographs, associative recall and code clones. Every experiment from EXP-002 onward was pre-registered; the first text cycle and the photograph application were not and are marked as such. Each ran with a trivial baseline, a blocking ablation and, where wiring was involved, a degree-preserving null, then adversarially reviewed and verified. The protocol caught instance leakage that let a dictionary lookup (MCC +0.374) beat every model; a Kenyon-cell layer whose removal did not change the result; a strong baseline (GRU) trained for 50 optimizer steps, whose retraining erased most pre-registered effects; and a naive null whose replacement by a degree-preserving null flipped a +13.53 pp connectome advantage to −7.11 pp. The MB is not a competitive classifier on text, chess or a synthetic conjunctive task; with a dense encoder its sparse code recognises near-duplicates, but exact search beats it wherever exact search fits (on photographs the two are indistinguishable and exact search wins the paired test), and on code clones a supervised linear model wins outright. The ring's parameter-efficiency headline fell to a 3-parameter vector integrator (0.11–0.20% loss against the Bayesian optimum versus the ring's 0.9%) and to a retrained GRU; its topology-versus-shuffled ablation (+44.49 pp) and untrained inductive bias survive. In chess, a small geometric grid beats logistic regression at full data but loses to a zero-parameter capture rule and to a dense network. The transferable rule: a strong baseline must receive more training budget than the proposed model.
What is in the record
Every number in the paper maps to a file and line of the research record (fontes.md). The record shipped with the paper includes:
- the append-only ledger of runs and the pre-registration of every experiment from EXP-002 on;
- two rounds of adversarial review and two rounds of numeric verification of this manuscript, kept in full, including what they found wrong;
- the referee report that retracted the first article of the programme, kept as written;
- the bibliography check: all 50 references resolved against their primary sources through the arXiv, Crossref and PubMed APIs on 27 September 2026.
The one number to remember
| what was compared | circuit | the cheap adversary |
|---|---|---|
| text (SWE-smith, first cycle, same split), MCC | +0.296 | dictionary on the instance id, no text read: +0.374 |
| chess puzzle moves, MCC | 0.313 | "the move is a capture", 0 parameters: 0.394 |
| code clones (BigCloneBench), F1 | 0.582 | logistic regression on TF-IDF: 0.865 |
| angular integration, loss vs optimum | ring, 9 params: 0.9% | vector integrator, 3 params: 0.11–0.20% |
The numbers come from the result files cited in the paper; the table is a summary, not a new measurement.
Cite
Patzdorf, T. (2026). When does a biological circuit transfer? Pre-registered, adversarially reviewed tests of Drosophila mushroom-body and ring-attractor models on text, chess, image, associative-recall and code-clone tasks. Preprint. https://genesisinnovation.io/mosca/paper