Work · Replayable and auditable simulation
A simulated organism that can be replayed, resumed and audited
A simulated quadruped in the MuJoCo physics engine, driven by an integer-arithmetic controller (the physics simulator itself computes in floating point), running four hours without interruption — where every step can be replayed identically, every state resumed bit for bit after interruption, and every organ disconnected to establish that it is necessary. The result is not the walking; it is the verifiability.
What actually sets this work apart
The question deserves a direct answer. The simulation in itself is not new. A quadruped walking in a physics engine, a neural network driving a body, a model inspired by the fly connectome — all of this exists, and many teams do it. If the claim were “look, a simulated creature that walks”, there would be no discovery in it.
What is genuinely uncommon is not the steps — it is the properties. A simulated body replayable bit for bit, whose state saves and resumes identically, driven by an integer-arithmetic controller (the MuJoCo physics simulator itself computes in floating point), whose every organ can be causally audited by disconnecting it, all within a measured peak of 144 MB on this run: that combination of properties is unusual. A floating-point simulation with non-deterministic parallelism is generally not bit-reproducible from one machine to another; that is a known property of such numerical stacks, not a measurement made here.
The accurate formulation, without overstating: this is not a fly, but a simulated quadruped whose controller uses a FlyWire-derived topology; nor is it “an intelligent animal” — the tasks and the learning remain deliberately simple at this stage, and no cognition is claimed. This is a demonstration of properties: showing that a simulated body, driven in closed loop, can be made deterministic, replayable and auditable. No world first is claimed — it is an unusual combination, to be set against the state of the art. That is what the project seeks to establish: computation whose behaviour can be established after the fact.
The four-hour monitor
A dashboard for the four hours has been reconstructed from the run data: it animates the chain of checkpoints (“Replay the 4 hours”); further down, the causal audit shows what happens when an organ is cut.
The graph and the body during execution
A second visualisation brings together the two sides of the system: the network that computes, and the body that acts. As a topology backdrop, it displays the entire FlyWire v783 graph — 139,255 neurons at their original coordinates, which can be rotated; the controller of this run, for its part, executes at every step a 50,000-neuron subgraph drawn from this graph (see the scope of the result). Bit-for-bit replay at the full scale of the graph is separate work, presented on the connectome page. The silhouette advances according to the distance actually travelled during the four hours, and the strip along the bottom marks the twenty-four checkpoints, the seven verified replays and the three causal ablations, at their exact moments. Between the two lies the world it moved through: the ten situations the programme actually cycled through over four hours — exploring in the dark, seeking a source, escaping a looming mass, recovering after a push — each on its own track, with what it yielded and the distance it produced.
A separate campaign (12–13 July 2026)
Two other views come from a separate campaign, carried out after the four-hour run. Their generalisation figures are not reproduced here. What they show solidly: the graph structure and the spikes recorded along the model’s olfactory pathway.
The replay evidence
This is the property that connects this work to everything else in the project. On 24 occasions during the four hours, the complete state of the organism was saved, then checked. On resumption after interruption, the fingerprint of the reloaded state was identical to the expected one:
expected fingerprint = reloaded fingerprint = c792e9b0…
Interrupting the organism and resuming it later produces no divergence: it resumes the same movement, the same internal dynamics, bit for bit. 24 checkpoints, 24 exact matches.
Causal audit, organ by organ
How can it be established that an organ genuinely serves a purpose? Disconnect it, and see whether behaviour collapses. This was done here for six distinct organs, at the end of the four-hour run (three control ablations were also made during execution, at 1 h, 2 h and 3 h):
| Organ disconnected | Reward | Effect |
|---|---|---|
| Intact organism | 0.278 | reference |
| Vision cut | 0.049 | collapse |
| Sensory projection pathway cut | 0.049 | collapse |
| Associative subgraph cut | 0.052 | collapse |
| Reinforcement signal cut | 0.052 | collapse |
| Integrator subgraph cut | 0.052 | collapse |
| Motor output shuffled | 0.057 | collapse |
Each organ drops the reward by a factor of about five or more (4.9 to 5.6) when removed. This is not a correlation observed after the fact: it is a causal ablation, measured by running the organism. Sensory input, associative subgraph, reinforcement signal, motor output: each is necessary to the measured behaviour. The six ablations do, however, give a very similar floor (0.049 to 0.057), which establishes the necessity of each part rather than its distinct contribution. That is what “auditable” means concretely.
Four hours of integrity
source: REPORT + long_life_checkpoint_report.json
Scope of the result
- This is a simulation scaffold: no cognition is claimed. Stability, continuity and auditability are demonstrated; the tasks remain simple and the rewards low.
- The topology is derived from FlyWire — a source of structure, not a biological connectome nor a reproduction of the fly’s nervous system. The simulated body is a quadruped, not a fly.
- During the 4-hour run, the quadruped controller executed at every step a 50,000-neuron subgraph (about 1.57 million connections) sampled from the full fly connectome (139,255 neurons, 3.73 million connections). The full graph was only used in short stability checks (7 times, a few dozen steps each). This subgraph is not the 50,000-neuron graph of paper D (2,121,443 connections).
- Subgraphs are named by their functional role in the model, not by a biological identity: the typing that would authorise such an attribution is not established. See the connectome page.
- No world first is claimed: the combination of properties is unusual, and is to be set against the state of the art.
- Result of a single run, on a single seed (seed 7), not reproduced with other seeds at this scale.
- Bit-for-bit resumption is verified within this run (24 checkpoints, 7 replays). Reproducibility from one machine to another is not claimed here: the physics part, in floating point, would depend on it.
- The execution is purely simulated, and this is guaranteed by the code: the runtime forbids any hardware access (
real_mode: falseis a safety invariant, not a degraded mode — the opposite value raises an error and halts execution). The verdict and its 16 gates cover this nominal, intended regime. - The region mappings for descending neurons (DN) and the ventral nerve cord (VNC) are approximations that are not biologically certified (
dn_certified: false,vnc_certified: false): any anatomical reading of these two regions is beyond the scope of this run. - The 27 million plasticity events are updates issued; their effect on the weights is not measured in this run pack.
Execution data
The sealed run
The complete execution package contains the checkpoint reports, the time series (5.7 MB), the learning, ablation and drift reports, the saved states, a fingerprint manifest and the four-hour monitor.
Extract from the end-of-run report (verdict name shortened): the other field names are those of the internal repository.
verdict : PASS_FORT (16 checks out of 16)
checks : 16/16 PASS
duration: 14,411.91 s (~240 min)
resume : fingerprint c792e9b0… — identical (bit-exact)
NaN/exc.: 0 / 0The complete time series of the run accompanies the data package.