Approximations: what the model simplifies, and why it stays
Status on 2026-10-14. Every simplification of the model in one place, each with its status and what would remove it. Nothing here is new: each row restates a page or a lock, which stays the authority. The roadmap holds the verdicts.
Status words:
- measured: the value is taken from a cited measurement;
- scanned: the quantity is not measured, so a lock runs its whole plausible range and reports whether the verdict depends on it;
- selected: chosen among stated variants by model selection, then replicated on fresh seeds;
- compensation: a structural simplification the lock or page names;
- blocked: removing it needs a measurement, a published map or hardware that does not exist yet (issue linked).
None of these can be removed by writing code alone. A value invented to fill a gap would be a new, unsourced approximation. Sources that could lift some of them are collected in leads.
Neurons and dynamics
| Simplification | Status | Removed by | Source |
|---|---|---|---|
| Point leaky integrate-and-fire neurons, no morphology or compartments | compensation | Per-type passive and active properties measured at scale; none exist for most of the 165,122 neurons | model |
| One parameter set for every neuron (τ_m 20 ms, τ_syn 5 ms, v_th −45 mV, t_ref 2.2 ms) | measured (Shiu et al. 2024) | Per-type values; the engine can take them (sim.Membrane, except t_ref); only T4/T5 v4 (run: robust neither) sets one: τ_syn of Mi1 and Tm1 from Behnia et al.'s delays |
model |
| One synaptic delay, 1.8 ms, for every connection | measured (Shiu et al. 2024) | Per-connection delays from conduction distances and measured velocities | model |
| Every neuron spikes by default; graded (non-spiking) neurons only inside a lock (photoreceptors and L1–L3 in T4/T5 v2–v4; Mi1, Tm3, Mi4, Mi9 and C3 in T4/T5 v3–v4) | compensation; curves scanned | A Drosophila voltage-to-release curve (#97), blocked: evidence | model |
| Fixed step of 0.1 ms; discrete events act at step boundaries | measured step (Shiu et al.); converged (SF2 v1) | Not a biological gap. Convergence v1 (seeds 0–4): rates and holds verdicts hold at 0.05 and 0.025 ms, and weak-pathway latencies and recruitment counts move with the step. Path-metrics v1 (seeds 5000–5004): at 0.1 ms recruitment is inside the seed range, and latency is not converged for sugar → MN9 | validation, path metrics |
| Input arriving in a neuron's own spike step is lost to the reset (about 1% at 100 Hz) | compensation (brian2 parity) | A model change, then a new validation run | model |
Synapses
| Simplification | Status | Removed by | Source |
|---|---|---|---|
Weight = synapse count × w_syn; no per-synapse strength |
compensation | Measured unitary strengths per connection type; the engine can scale weights per postsynaptic neuron (post_scale, density rule), none used |
model |
w_syn 0.192 mV: Shiu's 0.275 divided by 1.43, the lower quartile of the MaleCNS/FAFB density ratio |
selected (29/29, replicated on fresh seeds) | A measured MaleCNS-specific scale | model |
| Sign from the predicted transmitter; glutamate and histamine inhibitory everywhere | measured receptor classes, applied per type | Receptor identity per postsynaptic type | model |
| 4,143 neurons with sign 0 (unclear transmitter or modulatory), 2.61% of synapses | compensation | Better transmitter predictions; re-signing from FlyWire types would cover 0.34% of synapses and was judged not worth it | model |
| No short-term facilitation or depression; one optional plasticity rule (KC → MBON11) | compensation | Measured dynamics per synapse type | model |
| Neuromodulators have no fast effect (sign 0) | compensation | Per-type gains and time constants (#85), blocked: evidence | model |
No electrical synapses by default; optional sim.Gap inside a lock |
compensation; coefficients measured (KC v5–v8) | A gap-junction list on MaleCNS bodyIds and junctional conductances (#85), blocked: evidence | model |
Input and senses
| Simplification | Status | Removed by | Source |
|---|---|---|---|
| Default stimulus: Poisson spikes into a named type, one kick of 48.1 mV per input spike | measured scale (Shiu et al.) | Sensory transduction models for each organ | model |
| Odors: DoOR responses are normalised; the rate for a response of 1 (100 Hz) is stated by each lock | compensation | Absolute ORN firing rates per odorant and receptor | model |
| Eye: linear photoreceptors without adaptation; only columns with a viewing direction are driven | compensation (T4/T5 v4 lock) | Measured Drosophila photoreceptor transduction (#84), blocked: evidence | eye |
| Sensory cell bodies and transduction are outside the volume | compensation | Not in the connectome | roadmap |
Body
| Simplification | Status | Removed by | Source |
|---|---|---|---|
| M1/M2: one-degree-of-freedom body, engineered decoder and transducer gains | compensation (labelled engineering) | Measured LC4 tuning (#34) and a motor map, blocked: evidence | embodiment |
No general motor map: every row of validation/motor_map.yaml is unmapped |
blocked | Independent motor-neuron → muscle evidence (#19, #35) | embodiment |
| NeuroMechFly jump: open loop, TTM only, twitch shape not measured, female body with a male connectome | compensation (jump3d lock) | The femur-tibia extensor force and the other legs' drive, both unmeasured | embodiment |
Numerics of electrical synapses
| Simplification | Status | Removed by | Source |
|---|---|---|---|
| Registered protocols KC v5 and v7 keep the boundary-only refractory reset (v7 qualified on 50 ms of one odor and seed); KC v8 and new configurations use the continuous clamp (dense or Taylor solver) | compensation (protocols kept) | A new prospective lock for any rerun under continuous clamps (KC v8 for v7) | run integrity |
| The native executor refuses continuous-clamp electrical fixtures with refractory cells (KC v8 runs on the CPU oracle); GPU and console are not qualified | blocked | Physical hardware (#39, #40) | runtime |
Per-lock compensations
Each lock lists its own; the file is the authority.
| Lock | Count | Main items |
|---|---|---|
kc_sparseness.yaml v8 |
6 | eLNs by morphology (lLN2T, lLN1_bc variant); somatic coefficients; no rectification; rectangular spikelet; Taylor truncation |
mb_plasticity.yaml v1 |
7 | Eligibility window shape, all-or-none pulses, both sides, no recovery, odors as ORN types |
t4t5.yaml v4 |
13 | Threshold-linear release, medulla anchored at its settled dark potential, scanned rates and lamina gain, Behnia delays via τ_syn, T5 inputs spiking |
gf_looming.yaml v2 |
7 | Engineered LC4/LPLC2 forms with scanned gains, both sides driven, inferred disc sizes, genotype controls of unstated sex, passive giant fibers, follower band as the direct path, what was seen before the lock |
apl_claws.yaml v1 |
9 | Point KCs for claws, contrasts not amplitudes (no GCaMP), every KC and a normalised contrast, clipping by threshold, DoOR drive at 100 Hz for 1 s, male model, ALPN for NP225, TNT as synapse removal, seeds for flies |
song.yaml v1 |
8 | Temporal-only JO input, spike counts for calcium, zero model baseline, synapse-weighted neuropils, free JO rate, unmapped tracts, magnitude sign, dependent neuropils |
jump3d.yaml v1 |
6 | Open loop, held posture, fixed moment arm, twitch shape, TTM only, female body |
What this means for the results
The validation battery passes 29/29 with the defaults; the held-out ORN_DM1 test, the DNp01 silence check, Kenyon cell sparseness v1–v5 and T4/T5 v1–v4 fail (v1 neither; v2–v4 robust neither); Kenyon cell sparseness v7 is within only with the variant eLN set, and v8 replicates it under the continuous clamp. The model is a connectome- constrained hypothesis generator, not a validated fly: biological validation (M4, #36) is blocked on an independent held-out benchmark.