# Creature Lab · methods and boundaries

## A frozen graph, different motor adapters

The dataset indexes 139,255 FlyWire FAFB v783 neurons. The custom LIF model uses 2,531,358 active weighted edges after its transmitter rules, out of 2,698,236 source edges. Edges are aggregated connections, not a count of individual synapses. Anatomical weights remain fixed.

Sensory encodings and output adapters are engineered. This project does not establish biological fidelity, natural fly learning, consciousness, or superiority over another network. The original body-controller assumptions and integration equations are documented in [NEURAL_MODEL.md](../demo/NEURAL_MODEL.md).

## 01 · Body physics

The viewer replays recorded NeuroMechFly geometry and MuJoCo world transforms. Terrain is editable; generating a new recording runs the Python physics service. The neural controller uses synthetic bilateral odor/proximity cues and a calibrated descending-spike adapter. A separate A* baseline has full-map information. No pretrained FLYNN neural policy is included.

The bundled ridge recording is a six-second partial traversal. The ramp recording reaches its food target. Neither observation claims that every maze is solved. The body viewer and the grid maze use different motion models.

## 02 · Neural maze

Local wall cues stimulate the frozen spiking model. An artificial readout of 1,297 descending neurons is fit to local right-hand wall-following examples, without position, maze ID, or a planned route in the readout inputs. The saved `brain-v1` run reaches food on 10/10 layouts.

Disconnected edges and silenced descending neurons produce no motor output and reach 0/10 goals. Untrained and shuffled readout controls also reach 0/10 in their recorded diagnostic budgets. These controls establish dependence on the simulated neural signal, not biological validity. Trained trials have a 256-action budget; untrained/shuffled trials have 64, while silent-output controls stop early. These are not equal-budget policy benchmarks. The 100% held-out local cue accuracy is separate from whole-maze success.

## 03 · Fly Pac-Man

The arena is 28×31 tiles with 244 pellets, four power pellets, a ghost house and a wrapping side tunnel. Blinky pursues, Pinky targets ahead, Inky uses Blinky-relative targeting, and Clyde switches between pursuit and retreat. Chase/scatter transitions, frightened behavior, returning eyes and house release are implemented. Timing is scaled to decision steps; original arcade targeting bugs and level progression are not reproduced.

The controller uses 64 cached simulated neural responses to engineered sensory patterns: local walls, graph-distance food progress, danger, power and pellets. A numerical feature transform stabilizes learning. TD-style reward learning updates only the motor readout. During live play the model weights are frozen; displayed prediction error is telemetry, not an ongoing training update.

| Event | Reward |
| --- | ---: |
| Pellet | +1 |
| Power-pellet bonus | +3 |
| Ghost eaten | +6 |
| Clear the arena | +30 |
| Caught | −25 |
| Wall contact | −0.6 |
| Each step | −0.03 |
| Food-distance progress | +0.12 per step nearer; negative when farther |

`classic-more-v1` continues the previous readout for 20,000 training games. Validation selects the checkpoint at +5,000 games. On the same 200 fresh test seeds, the previous policy wins 60 times and the selected policy 97 times. That is a change from 30% to 48.5%, or 18.5 percentage points. The saved files retain both policies, validation history, configurations and per-seed outcomes. A historical 16/50 result is a separate cohort.

The test set measures ghost-seed variation on one layout, not new-map generalization. Showing only one successful replay is not evidence of the overall win rate; use the full evaluation summary. No biological dopamine neurons are modeled; the teaching signal is a numerical analogy.

## Published artifacts

- [Neural maze result](../lab/brain-runs/brain-v1/result.json)
- [Neural maze manifest](../lab/brain-runs/brain-v1/manifest.json)
- [Pac-Man paired evaluation](../lab/pacman-runs/classic-more-v1/result.json)
- [Selected policy](../lab/pacman-runs/classic-more-v1/policy.json)
- [Previous policy](../lab/pacman-runs/classic-more-v1/parent-policy.json)

The public copies omit the original machine nickname. Scientific values, dataset checksums, policy weights and recorded outcomes are retained. Some historical source hashes refer to the exact original worker revision rather than the later gateway/UI revision in this repository.