Can a fly learn to trade?

The public fly runs on biological wiring. We’re investigating what it could learn from experience—and whether any advantage survives trading costs.

Verified testnet receipts → Separate, manually triggered Base and Solana engineering tests are now complete. These are not decisions from the public paper model.

Current public model and earlier measurements

A real fly connectome — 164,587 neurons from a male Drosophila CNS — was used for paper trading on an ETH-USD reference feed. The public demo currently shows archived observations, decisions and spike counts, with recorded full-state checkpoints. It is an experiment in whether biological wiring computes anything useful. It is not investment advice, and it has proved nothing yet.

1. What the fly is

The wiring is the MaleCNS v1.0 connectome: one adult male Drosophila, whole central nervous system including the ventral nerve cord. Everything around the wiring is engineered and unlearned. In the current public model there is no plasticity or training step: its weights do not update. Execution is paper only, on a simulated USD ledger.

FlyWire, the dataset behind Shiu et al. (2024) and Costi et al. (2025), is a different animal and a different dataset (female brain, no VNC, 139,255 neurons). Results on it are precedent for this network, not validation of it.

2. What has been measured

The earlier measurements below cover a synthetic input study, a bias diagnostic, a memory probe, and a live-reference paper capture whose recorder has been validated. The latest full-network input check is summarised above. Two further studies ran on 11 September and have sections of their own: the interface diagnostic (section 3) and the positive control PC0 (section 4). All are exploratory. None is a forecasting test.

2a. Synthetic input study (12 rounds per condition)

2b. Bias diagnostic (256 rounds)

2c. Memory probe (244 valid round-equivalents)

2d. Live-reference paper capture and recorder validation

As checked on 19 September 2026, the public show displays archived paper observations from an ETH-USD reference feed; it is not a currently updating trading run. The historical capture resumed on 12 September 2026 at 15:26 CEST, using malecns_lif v0.1.0 and recorder schema 1.2.0. Newer research builds have not been adopted. See the show and its capture status for recorded timestamps and freshness. These observations use simulated funds and do not establish forecasting skill.

The earlier six-observation recorder validation is retained below. The separate 12-round demonstration remains available as a historical snapshot in section 9; neither should be read as evidence that a new capture is currently running.

3. Interface diagnostic, 11 September

On 11 September a pre-declared, locally pinned diagnostic asked a narrower question than the paper does: can this interface — encoder, state mode and decoder — carry the current input at all? It ran on synthetic independent random z with no market data, on one network at one gain and tonic setting (gain 0.01, tonic 0.07). It is a draft and it was not preregistered.

The three pre-declared primaries

The first run was declared void; the re-run is the record

The first run, at 09:03–09:15Z, carries a root decision of VOID under the protocol's own pre-declared process-limit clause, §6(f): "any arm exceeds the 75-min process stop or more than 4 simulation processes run at once". In plain language: while the four diagnostic lanes were running, two unrelated simulations were reading the same connectome cache on the same machine — the positive-control assay described in section 4 and the go-live paper capture — so six simulation processes were live at once against a pre-declared ceiling of four. The clause is unqualified and the harness did not machine-check it, so the run breaches it.

The simulator is deterministic and single-threaded, the cache is read-only and verified unchanged in all seven arm metadata records, and the processes share no mutable state, so no recorded number was affected; only the wall-clock timings were inflated, and those feed nothing but two guards that never tripped. The consequence was procedural rather than numerical, and it stood: that first run issued no verdict of record, and its outputs were kept as a determinism reference — the numbers a quiescent re-run had to reproduce bit for bit.

That re-run has since been done. The same diagnostic ran again on a quiet machine under the pinned re-run protocol, in four lanes from 12:52 to 13:03 local time, with the process-limit clause machine-checked at four points and clean at every one: before launch (no other process holding the connectome cache, four lanes planned against a ceiling of four), at each of the four lane starts, at each of the seven arm starts, and after the lanes finished. All seven arms completed their full planned rounds, each lane exited zero, the analysis step exited zero and opened the held-out window once, and the P1 precondition was written to its own file before that window was opened rather than reconstructed afterwards.

It reproduced the archived records bit for bit apart from wall-clock timings. Every one of the 1,852 recorded rounds and footers across the seven arms is identical to the void run's once the elapsed-time, memory and timestamp fields are stripped, and the hash over each arm's whole stripped record stream matches in both directories. Of the 1,113 comparable values in the analysis output — every leaf that is not a clock, a memory figure, a code hash or a guard field — 1,112 are identical; the single difference is a prose footnote that lost a citation. The verdicts are therefore unchanged, and they are now the verdicts of record: P1 FAIL, P2 FAIL, P3 INCONCLUSIVE.

The re-run's own fingerprints, for checking: results.json 584aa11f…, results.md 5cbe8621…, protocol.md 24ed1e1f…, meta.json 627423eb…. The scope is the one it started with and has not widened: synthetic independent random input, one network, one gain and tonic setting, a draft that was not preregistered.

Secondaries under Benjamini–Hochberg

Six secondaries carried p-values, and a Benjamini–Hochberg step-up at q = 0.10 was computed over all six. One survives it.

How this was put publicly

On 11 September 2026 we ran a pre-declared, locally pinned, not-preregistered diagnostic of the fly interface on synthetic random inputs (no market data), one network, one gain/tonic setting. All three primary tests came out negative or inconclusive: the sign-and-offset decoder fix did not pass under persistent state (balanced accuracy 0.57 on the rounds where it acted, holding 61 %), the trained readout did not pass the input gate under persistent state (0.55, CI 0.47–0.64, threshold 0.75), and the count-versus-wiring question stayed inconclusive. The harness control passed: when the network is reset to the same fresh state before every round, the current input is recoverable (trained readout 1.0 on 34 mirrored pairs; the flipped legacy decoder 0.93 on the 54 % of rounds where it does not hold). We are considering resetting state before each decision and flipping the decoder sign; that configuration has not been tested as it would be deployed, it removes any across-bar memory, and it says nothing about markets, forecasting, or whether the wiring matters. The fly has proved nothing yet.

Addendum recorded with that paragraph: the diagnostic was later declared void on a pre-declared process-limit clause because two unrelated simulations shared the machine during the run; the numbers are unaffected by construction and will be re-run on a quiescent machine before anything is quoted as a result. That quiescent re-run has since been run and reproduced the same numbers and the same three verdicts, so the paragraph above reads the same on the record of the re-run.

The amendment is a proposal, not a decision

The proposal on the table is to reset the network state before every decision, flip the decoder sign so that a right-descending-neuron excess reads as sell (the action is taken from −m), and add an offset fitted on the first window. The configuration actually measured is the fresh-reset arm A1 with decoder D2: balanced accuracy 0.929 on the 54 % of labelled test rounds where it acts (37 of 68 rounds; hold 0.46), buy 19 / sell 18, sell share 0.49, p = 0.001, offset m0 = +0.0192. Those 68 labelled rounds are 34 mirrored (z, −z) pairs from one fresh start state, so the honest confidence half-width is ±0.146, not ±0.103.

Status: proposal, not adopted. The quiescent re-run condition is the one that has been met; three conditions are still open — the engineering check on the real cache, a pinned gated protocol for the amended configuration (written and pinned, not yet run), and external anchoring. The registry entry for the decision records amendment_adopted: false and market_data_read: false. Nothing in the protocol or the configuration was edited.

Five conditions were set before it could be adopted. The first and fourth are met; the other three are open:

1. Met. PC0 re-run under the new mode — reset per decision, a declared start state with its hash, the flipped sign. Run on the same seeds under the amended package: identical result (section 4).
2. Open. The engineering check passing on the amended interface: byte-identical decisions under the new mode on the real cache, the recorder showing the declared start state before every bar, and the process guard clean.
3. Open. A new pinned, gated protocol for the exact amended configuration, run on a quiescent machine, passing on its own pre-declared bars.
4. Met. The quiescent re-run of this diagnostic reproducing the archived raw records bit for bit — done under the pinned re-run protocol with the guard clean at every check, 1,852 of 1,852 records identical once timings are stripped.
5. Open. An external anchor of the amendment note and the configuration hash — without it, "decided before market data" is only a local file timestamp.

What it does not establish, on the record of the amendment and the v2 package's own "what is NOT claimed" list:

• No warmed start state was tested; every reset round restarted from a fresh state. The one warmed-start fixed-decoder condition on file failed its input gate (test R² −0.101), so the A1 numbers cannot be carried onto a warmed design.
• The sign is regime-dependent: −1 in the fresh-reset regime, +1 under persistent state — and undetermined there, since the bootstrap interval on the underlying correlation contains zero (r = +0.086, n = 238, t = +1.33, CI95 [−0.047, +0.219]). The persistent-state sign is unstable across data sets.
• Reset per decision removes all memory across bars, so the amended system is a feed-forward map of one scalar. That changes the question the paper asks rather than answering it.
• The offset is a single constant fitted on one window of one arm of one network at one operating point (gain 0.01, tonic 0.07, 60 steps) on synthetic input. It is not market-calibrated and must not be refitted on market windows without a new preregistered protocol.
• On the toy fixture the change flips every action the previous version took. That is the mechanical consequence of the sign, not evidence that the new decoder decides better.
• No degree-preserving rewire null was run, so nothing here speaks to whether the MaleCNS wiring matters. No market bar was read.
• Reset per decision is a production change, not a configuration change: the live adapter currently resets only when no prior state exists, and the grid check does not yet compare state policy, decoder sign or margin offset, so it would pass the new interface without noticing it.

4. Positive control (PC0)

A model that cannot reproduce a reflex which has already been validated in silico against behaviour is not worth pointing at a price feed. PC0 is that check, declared on 11 September before any run on the real graph and hashed so the runner refuses to start if the rule block has changed. It reproduces the contrast in Shiu et al. (2024) — real wiring versus shuffled weights — not the absolute firing rates, on our own dataset with our own simulator unchanged.

The pathway is sugar-sensing gustatory receptor neurons driving the proboscis motor neuron MN9. On this graph the populations resolve to 12 sugar-GRN candidate cells, all of type LB3c and all on the left side, and 2 MN9 cells, one per side. The drive is a seeded supra-threshold kick of +4.0 units per stimulated cell at a nominal 100 Hz for 1,000 steps from a fresh reset, with tonic current 0 everywhere — Shiu's zero basal assumption, not the 0.07 per step that the market adapter gives every non-photoreceptor. Thirty trials per cell of the sweep, master seed 20260911.

What that buys is narrow, and the protocol says so in advance: it means only that under this drive the real wiring separates from the shuffled-weight control on this one pathway, so the interface is not degenerate. It does not mean the MN9 rates are quantitatively right — Shiu et al. disclaim absolute rates themselves — nor that the graph carries anything market-relevant, nor that the selected gain is correct beyond this pathway, nor that the model handles inhibition-dominated or neuromodulated circuits, nor that a FlyWire-based validation transfers to MaleCNS.

Three things logged against this result

• LB3c as the sugar GRN type is an unverified lead. At the time of this run it had not been checked against the MaleCNS gustatory-typing companion paper. We have since acquired that paper; this population validation remains open. The label table gives 23 LB3c cells in total, sensory superclass, 12 left and 11 right; this run stimulated the 12 on the left.
• The selected gain is not the gain in production. The protocol freezes the selected gain (0.006) for the market study's gain axis, but production and the interface diagnostic both ran at gain 0.01. The discrepancy is logged for the configuration owner, and 0.006 has to be added to the declared gain grid; it has not been silently resolved.
• The null arms reported after the fix. The protocol's own null families at the selected gain 0.006, 30 trials each, reported rather than gated: the block- and degree-preserving rewire (NB:1, 232,273,579 accepted swaps, all 101 self-edges kept fixed) fired MN9 in 0 of 30 trials; the within-block weight permutation (NP:1, 25,562,837 weights moved) fired in 0 of 30. With the global weight shuffle also at 0 of 30 and the real wiring at 30 of 30, the reflex on this graph needs the specific wiring, not only its block structure or its weight distribution. The first attempt had crashed in the null generator because the real graph contains 101 self-edges and the generator assumed none; the generator was corrected, the run resumed in place reusing the 90 recorded trials verbatim, and the frozen rule recomputed the same verdict. Generating one rewired null of this size took 36 minutes.

PC0 was run twice: once under the current interface (above) and once under the amended package (fresh reset per decision, flipped decoder sign, offset), the latter on the same seeds. The amended-package run reproduced the same numbers exactly: 0 of 30 at gain 0.004, 30 of 30 at 0.006 at 43.73 Hz, global shuffle 0 of 30, selected gain 0.006. This was expected rather than informative: the assay drives sensory neurons and reads a motor neuron with a fresh reset every trial, so the decoder and the state policy are never exercised. It closes the first of the amendment's five conditions; the two null families were not re-run and the current-interface null results stand.

5. What has not been established

These are active research questions and engineering checks. Our first priority is the input-to-decision path: neutral bias, readout calibration and the effect of the chosen dynamics. Memory tests need an interface that first passes an input-sensitivity check. Market evaluation and matched wiring controls will answer separate questions about economic usefulness and the contribution of biological connectivity. A successful repair does not guarantee a positive scientific or trading result.

• Full-network forecasting skill remains unestablished. The separate learning-rule market screen was negative. It did not test the full-network model, and no net economic advantage has been established. Where input does move actions, it shows only that the fly reacts to recent returns, not that it anticipates future ones.
• Memory remains unresolved. The persistent-state input gate failed, so every lag probe is uninterpretable. The correct statement is "not measurable at this interface and sample size", not "no memory" and not "memory".
• Biological fidelity has not been validated. Real connectivity does not mean biological fidelity. With the eyes off, all 1,314 descending neurons fire the same count within a round; that is a warning about the modelled regime.
• The all-buy readout is a candidate engineering bug under test. Under neutral input the right-DN margin (+0.03 to +0.10) sits above the 0.03 threshold, so the readout is near-constant buy. The selected right optic-sensory input population is 36.56% larger than the left; all selected inputs are histamine-labelled. Unequal input drive is a confound to test, not a demonstrated cause. DN side labels use somaSide fallback; functional left/right trading roles are not established. In that earlier input-study protocol, the one-round response from reset moved the margin opposite to the decoder's sign convention and disappeared after 120 warm steps; the separate 13 September eyes-on check retains a response after 240 warm steps. This is a decoder property, not a market signal, and it must be resolved as a bug.
• Sign and tonic conventions differ from the published reference model. Shiu et al. (2024) use 0 Hz basal firing and treat dopamine, octopamine and serotonin as excitatory, with one free parameter calibrated against a measured motor readout. This model adds a tonic current of 0.07, sets those three transmitters to 0, and uses an uncalibrated gain of 0.01. The 3,017 neurons whose transmitter is "unclear" are set to +1. This is a third convention, different from both Shiu et al. (2024) and Costi et al. (2025), and a sign-convention sensitivity arm is required before any primary result.
• Topology and dynamics are not separated. The contributions of the wiring and of the chosen dynamics have not been isolated; no rewired null, no matched input population and no topology control has been run on the interface.
• No wallet should be funded on the current mapping. That is the standing recommendation of the input study.

6. What the paper will ask

Primary contrast (protocol v0.2, H1). With encoder, readout, hyperparameter grid, simulation length and per-bar compute held identical, does a reservoir whose recurrent weights are the real MaleCNS v1.0 connectome reach higher out-of-sample directional forecasting skill on a frozen test window than 79 matched nulls in which topology is rewired while each neuron's degree, sign and outgoing weight multiset are preserved?

Two cautions from the internal audit stand in front of that study. First, the confirmatory design carries state across bars, and the only persistent-state evidence (section 2c) says the input gate fails, so as written the study would likely stop at its own gate. Second, its primary arm costs 459–569 CPU-hours depending on the per-round time assumed, and the 29–33 h wall-clock figure assumes unverified linear scaling on 16 processes.

The earlier study plan proposed a separate interface gate before the primary study: four arms (production and E-bal encoder × reset and persistent state), 400 rounds each, asking whether any encoder lets the DN counts carry the current input when state persists, at about 37 minutes of CPU. That proposed gate is not a completed result. The current development tracks are listed at the top of this page; the study design still needs review of null construction and multiple comparisons.

7. What the literature says

• Damicelli, Hilgetag & Goulas (2022). Across macaque, marmoset and human macroscale connectomes and two memory tasks, connectome reservoirs performed as well as, not better than, random reservoirs, and the variant preserving empirical weight ranking was significantly worse; the closest prior null for this project.
• Costi, Hadjiivanov, Dold, Hale & Izzo (2025). FlyWire subgraphs as echo-state reservoirs had higher training error but lower test error than a random reservoir at ridge β = 10⁻⁶, an overfitting-resistance effect that grows with size, is matched or beaten by the control at β = 10⁻³, and disappears or reverses at spectral radius 0.75, 0.5 and 0.25; any advantage here must be held against that regularisation explanation.
• Shiu et al. (2024). A whole-brain FlyWire LIF model with 0 Hz basal firing and one calibrated parameter (W_syn = 0.275 mV) drove the MN9 motor neuron in 100/100 simulations with the real wiring versus 1/100 with shuffled weights and matched 91 % of 164 predictions; its sign rule (GABA and glutamate inhibitory; dopamine, octopamine and serotonin excitatory) differs from the one used here.

Nothing read supports an expectation of a market edge, and two of the three papers give concrete mechanisms by which a positive result would be an artefact.

References (APA 7)

• Berg, S., Beckett, I. R., Costa, M., Schlegel, P., Januszewski, M., Marin, E. C., … Jefferis, G. S. X. E. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell, 189(18), 5504–5526.e15. https://doi.org/10.1016/j.cell.2026.08.015
• Costi, L., Hadjiivanov, A., Dold, D., Hale, Z. F., & Izzo, D. (2025). The Drosophila connectome as a computational reservoir for time-series prediction. Biomimetics, 10(5), 341. https://doi.org/10.3390/biomimetics10050341
• Damicelli, F., Hilgetag, C. C., & Goulas, A. (2022). Brain connectivity meets reservoir computing. PLOS Computational Biology, 18(11), e1010639. https://doi.org/10.1371/journal.pcbi.1010639
• Shiu, P. K., Sterne, G. R., Spiller, N., Franconville, R., Sandoval, A., Zhou, J., … Scott, K. (2024). A Drosophila computational brain model reveals sensorimotor processing. Nature, 634(8032), 210–219. https://doi.org/10.1038/s41586-024-07763-9

8. Transmitter-sign provenance

Synapse signs come from the source dataset's neurotransmitter classifier (ResNet50 following Eckstein et al., 2024; argmax over per-synapse scores; 80/20 neuron-level split). The held-out neuron-level confusion matrix is printed only as a figure, Figure S1 panel K of Berg et al. (2026); the values below were transcribed from that raster and are confirmed against the Cell 2026 version of record, re-read cell by cell (neuron-level, held-out test set).

9. How to check us

Receipts

Decisions are published as inspectable receipts. The public state is: inspectable receipts, simulated USD ledger, experimental fixed-weight LIF connectome, no real-money execution.

Archived paper mirror

../recorded/live/ — archived observations; not currently updating. Check the capture status for recorded timestamps.

Recorded snapshot

../recorded/live-demo-12/ — the recorded snapshot of the stopped 12-round paper capture.

Trial registry

A paper trial registry now exists and is hash-chained: 23 entries, each carrying the protocol, code, data and window-boundary references of the evaluation it records, plus one entry for the amendment decision itself. The chain was verified end to end, and again by an independent reader of the same file. The entry hashes are not published on this page yet.

Hash chain

Each research event is chained; the validation run above verified its chain from genesis. The chain tip is not externally timestamped.

Protocol hash

Not yet anchored. No hash of the protocol has been externally anchored and it is not registered anywhere; the site's integrity model detects casual edits only and is explicitly not tamper-proof.

Replay

Decisions replay bit-exactly from the archived pre-state and stored input on a fresh adapter loaded from the immutable cache (6 of 6 in the validation run; 0 mismatches in the input-study harness).

Re-analysis

The memory probe's analysis re-runs offline from its raw per-round file with no simulation and no network; the bias diagnostic ships its code, protocol and cache hashes so the run can be repeated against the same reference.

Validation-run public export

The historical six-observation run's JSON rows validated against fly_research_event.v1.1.json; each row's payload hash recomputes; header paths redacted.

What would change this page: a passed persistent-state input gate, a resolved cause for the buy bias, an anchored protocol hash, a public registration, and only then a read of the frozen test window.