2026-09-11 record
Checks the downloaded file’s SHA-256. A byte match is not a rerun.
RESEARCH / THE SCIENTIFIC BASELINE
Look beneath the glass.
The wiring is real. The model makes assumptions. These files let you examine the difference.
ADAM and ATOM share one token but run independent model states. The neural-response gate follows the Protocol; market value is not evidence of learning or biological mating.
NEURONS IN FILTERED MODEL—MaleCNS v1.0 anatomy
SIGNED MODEL CONNECTIONS—MNonzero directed connections
INITIAL ASSAYS—6 input conditions × 3 seeds · 200 ms each
CURRENT EXPERIMENT / AUDITORY BRIDGE
One input crossed the gate. One did not.
Artificial vpoEN drive: ADAM 91.7 Hz and ATOM 83.3 Hz peak pC1_6a activity above both controls. Artificial vpoIN drive: 0 Hz in both seeds. Six resting-state full-network trials per input kind; 14 × 20 ms windows per trial.
The buy mapping passes the frozen 25 Hz / three-window engineering rule. The sell mapping does not. Zero output does not demonstrate suppression. Historical transaction rehearsals created no birth; these results do not establish learning or physiological mating.
Verify the reduced circuit locally170 cells / a separate experiment ↗
RUN IT ON YOUR MACHINE
A result you can check.
170 actual MaleCNS cells. 840 retained signed connections. A small, frozen LIF experiment runs locally in your browser.
This is a new reduced circuit, not a rerun of the full 165,122-neuron assay. Two separate checks keep that distinction visible.
Reference · 2026-09-11
Recomputes all 42 windows across Connected, No input and Blocked; compares integer results against an independently run Python reference.
Selection, assumptions & downloadable evidence
The cut retains 32 L1 and 32 L2 cells by ascending body ID, the 96 strongest direct downstream targets, and 10 descending cells. External connections are omitted. Direct artificial 100 Hz input replaces the full visual encoder. The fixed seed and all parameters are public; matching confirms computation of this cut, not whole-brain or biological validity.
Checksums are compared with references published by this project. Independent replication and external review remain welcome.
REPRODUCIBILITY / SIX MATCHED TRIAL SUMMARIES
A response reproduced. Movement still unproven.
The visual-interface assay was run again with the same MaleCNS graph, model, interface and seeds. All six summary measurements matched the earlier run. Connected vision propagated beyond the retinal input cells; neither control did. None produced locomotion.
| Condition | Seed 17 · peak cells | Seed 29 · peak cells | Distance |
|---|---|---|---|
| Vision connected | 245 | 257 | 0.000 mm |
| No visual input | 0 | 0 | 0.000 mm |
| Retinal output blocked | 0 | 0 | 0.000 mm |
Peak cells means active neurons beyond the retinal input populations in one 20 ms window. This checks input transmission; it does not establish pursuit, learning or mating. Next: calibrate the dynamics and independently validate the motor decoder before increasing behavioral complexity.
Reproduce and inspect provenance
Run the existing assay from the project checkout after installing the documented MaleCNS data and Python dependencies. The command writes a separate run and does not alter the live founder worker.
python science/habitat_assay.py --out work/replication
Record generated 2026-09-11 13:39:01 UTC · run e98cdf49-6001-484e-9f4e-6870ae6ecbb0
Graph SHA-256: afaefbda9d28aee886ccace0fb0b7f45fbc65279f5ccdf02dd4c439f162f5bdbEXPERIMENT BULLETINS
The result, including the zero.
Dated records. No new date without a new experiment.
Input propagated.
No locomotion.
Connected peaks: 245 and 257 cells beyond retina. Both controls: 0. All six trials: 0.000 mm.
Open the matched rerun ↗The association screen
did not pass.
DM1 paired and delayed reinforcement both scored 87.5%. Neither odor met all criteria across the 88-episode study.
Open the negative result ↗EXPERIMENT 01 / IMPLEMENTED INTERFACES
Does a visual input reach a motor output?
A synthetic target is projected onto measured L1/L2 hex-column coordinates. The full filtered MaleCNS network integrates that input, then a target-blind decoder maps DNa02, DNa01, MDN and DNp09 measurements to a body pose. This is an artificial sensory and kinematic interface, not a reconstructed retina or validated animal.
Six independent trials cover two random seeds and three conditions: connected vision, no visual input, and blocked retinal output. Each trial lasts 280 ms of model time, including onset, a moving target and stimulus withdrawal. Neural state persists within a trial and resets between trials. Seeds are stochastic realizations of one anatomical specimen.
The first run recorded propagation beyond the retinal input populations, abolished by the output-blocking control. All six trajectories had zero locomotion. These results establish neither pursuit nor learning. The short assay does not replace physiological calibration, and the female receptive circuitry needed for male–female mating is absent from the MaleCNS baseline.
Input rates, decoder scales and the trial schedule are fixed in the downloadable record. The decoder has no access to the target. No learning update, wallet access or deployment command runs in this assay. The original continuous founder worker remains separate.
Recorded visual-interface assay
EXPERIMENT 01 / SENSORY INTERFACES
A stimulus. A circuit. An observed response.
Independent assay states · same MaleCNS anatomy · no learning or mating enabled
WORLD & BODY / mmAwaiting measurement
Position is measured from the motor decoder. A stationary point means no locomotion was produced.
INPUT → NEURAL OUTPUT20 ms / window
- Visual input events
- —
- Active beyond L1/L2
- —
- DNa02 left / right
- — Hz
- Decoded body speed
- — mm/s
- Target bearing at input
- —°
L1/L2 INPUT MAP · — mapped cells
Synthetic angular projection · not a reconstructed retina
Synthetic angular projection · not a reconstructed retina
MATCHED CONTROL CONDITIONS
| Condition | Peak beyond retina | Distance |
|---|
OBSERVATION
Read the method and limitations ↗Collecting the comparison
The experiment is recording input, propagation and motor output. Results remain provisional.
Window log 0 measured windows · select a row to inspect
| Model time | Input spikes | Active neurons | Beyond retina | Speed mm/s | Inspect |
|---|
Archived digital-evolution prototype
A separate rate-model food-finding prototype tested parameter inheritance and generated algorithmic offspring. It is paused. Its offspring are research artifacts, excluded from the current token-bearing lineage. Its results do not validate mating or authorize token deployment.
Archived colony & birth records ↓Loading experiment records…
OFFLINE LEARNING STUDY
Learning needs a control.
A separate conditioning pilot tested whether the model learns a specific association. These are offline experiments, not active founder training.
Download study results ↓88 STUDY EPISODES COMPLETEDScreen not passed
- DM1 test accuracy · paired
- 87.5%
- DM1 test accuracy · delayed
- 87.5%
Paired and delayed reinforcement produced equal DM1 test-choice accuracy. Neither DM1 nor DA1 passed the full study criteria. These percentages are not founder intelligence or breeding scores.
EVIDENCE & LIMITATIONS
What the records establish.
The initial archive contains 18 independent 200 ms assays: six artificial input conditions × three seeds. The visual-interface experiment is a separate set of six 280 ms trials, repeated with matching summary results. Both use the filtered MaleCNS model with 165,122 neurons and 10,228,000 nonzero signed connections.
The Cell paper supplies anatomical evidence. Transmitter signs, neural dynamics, sensory encoding and motor decoding add model assumptions. The transaction bridge now runs separate full-network auditory assays through vpoEN and vpoIN, with matched no-input and blocked controls; see the Protocol for the artificial mapping and response gate. The colony separately shows saved windows from the continuous founder worker; receiving new windows does not mean learning is enabled.
The culture dish uses ambient flight and grooming animations. Individual portraits illustrate received motor and stimulus signals; neither view is a validated behavioral reconstruction. Measured neural panels update only with received records and retain their timestamps when the stream is unavailable.
Only seed 17 has spatial samples in the initial assay archive. The research plot labels that seed explicitly. Each founder’s neural state panel uses its own numerical observations; no shared spatial recording is presented as live neural activity.
Data: HHMI Janelia FlyEM and collaborators, CC-BY 4.0. Simulator adapted from fruitflydev/flycoinrh, MIT. Derived JSON records retain checksums, body IDs and parameters.