One-bit looming jump

Issue #8 asks for one closed loop in which an action changes the sensory input that follows it. This note freezes that task on 2026-10-11. The lock is validation/behavior.yaml (version 1). Changing the task, the body, a deciding check, or a FlyBody number requires a new version.

flymsg behavior prints the protocol and returns before loading the connectome. flymsg behavior --run executes that loop at the default parameters. Criterion v4, N1, and validation/benchmarks.yaml stay as they are. The looming case is not added again to validate.CASES.

Verdicts

Three locked questions share this page. Each section below keeps its protocol and its numbers. The limits under the one-bit task belong to that protocol.

Result (2026-10-11)

The closed loop fails. w_syn stayed 0.192 and th_jump stayed 0. Seeds 0–2. The protocol file was not edited after the numbers. The run started from commit 96cee3b with the runner still uncommitted, so the manifest records that commit as dirty. The loop itself took 90.5 s. Same MaleCNS tables as the held-out run.

Each seed has two rows. The triangle is t_jump, and the closed loop has no spike after it. With both DNp01 silenced, TTMn_R still reaches three spikes, later and inside the 300 ms drive. These times are not a takeoff latency.

Jump, feedback, and the decoder prefix hold. Silencing both giant-fiber neurons does not keep TTMn_R at or below 2 spikes. The third spike is later than in the intact loop, and it still arrives inside the 300 ms drive, so the jump rule cuts that drive. That is the finding. It is not a retune.

With the drive left on, each seed has 22 TTMn_R spikes. With the cut, the closed loop stops at the three spikes above. DNp01(GF)_R fires first, at 4.8, 4.9 and 5.1 ms, and DNp01(GF)_L at 12.4, 12.9 and 12.5 ms, both before TTMn_R. On the closed loop, no neuron is still firing from 400 ms to 600 ms. The v4 limit of 100 does not decide, and these rows do not rescue silence. The times are the model under a 100 Hz Poisson drive. They are not a takeoff latency.

gf_required_for_the_jump_bit stays unverified. The silence arm failed in the model. It does not test the animal.

Task

The task is a one-bit jump on the looming-escape circuit already in the v4 battery.

LC4_R is the stimulus of the looming case. That case scores types DNp01 and TTMn. A unilateral LC4_R drive recruits the right-hand motor neuron, so this protocol names the instance TTMn_R. The giant-fiber instances in the table are DNp01(GF)_R and DNp01(GF)_L. There is no DNp01_R row.

The endpoint is measurable inside the existing simulator. The sensory trace is the Poisson kick matrix into LC4_R. The motor trace is the spike count of the one TTMn_R neuron. The body state is two fields, escaped and t_jump_ms. escaped becomes true on the first 0.1 ms step at which TTMn_R has emitted 3 spikes. From the next step on, the LC4_R drive is zero. That cut is the action changing the sensory input.

The circuit reference is the same one the validation case already cites. LC4 and LPLC2 drive the giant fiber, and the giant fiber drives the jump-muscle motor neuron (von Reyn et al., Nat. Neurosci. 17, 962–970, 2014, doi:10.1038/nn.3741; Ache et al., Curr. Biol. 29, 1073–1081, 2019, doi:10.1016/j.cub.2019.01.079). The chemical weights already reported for this case are LC4*R → DNp01 2580/2580 (sum/max) and DNp01 → TTMn 90/70. On the traced graph the 90/70 pair is DNp01(GF)_R → TTMn_R 70 and DNp01(GF)_L → TTMn_L 20. The 70 is the maximum. It is the reported giant-fiber edge, not the residual partner. The direct LC4_R → TTMn weight is 0 (validation). Those numbers are why the case was calibratable. They were not looked up again to choose a target.

Walking, song, grooming, and proboscis extension need a map from motor neurons onto joints. This repository does not have that map. The held-out olfactory case ORN_DM1 → DM1_lPN fails specific and stability (benchmarks), so it is not the substrate. CO2 avoidance stays unscored. LPLC2, the alternate LC4 routes, MDN, and T4/T5 stay in the exploratory partition.

What the result is allowed to mean

Looming escape is in the calibration partition. It took part in choosing w_syn. A closed loop on it is a feasibility demonstration on that circuit. claim.status is calibration_feasibility and held_out is false. A pass does not show that the model generalizes, takes off, walks, or sees a loom. A failure is a finding. w_syn and th_jump stay at the defaults, N1 is not redrawn, and the held-out manifest is not edited.

Body

Two bodies were compared. The numbers for FlyBody were read on 2026-10-11 from the public repository https://github.com/TuragaLab/flybody (default branch main). FlyBody was not installed.

FlyBody's published tasks are locomotion and flight. The task list has no looming-jump environment. Vision-guided flight observes a rendered camera. It does not take LC4_R spikes. Nothing read on that day maps TTMn_R onto a tergotrochanter actuator, or a MaleCNS motor neuron onto an action coordinate.

A re-read on 2026-10-12 of FlyBody commit d015e9b (2025-07-30) found the same timesteps. The 59 walking actions in that commit's docs/getting-started.ipynb are named actuators: claw adhesion, head, abdomen, and leg joints through tarsus2. Wings and mouth are off. No name is a MaleCNS motor neuron. The Nature paper states that the reconstructed body is a female fly, and the supplementary masses are the mean of 52 females. The README, tasks/constants.py and the head of fruitfly.xml still do not state the sex. pyproject.toml there allows Python >=3.10 and pins numpy==1.26.4. Issue #19 records the re-read. It does not select FlyBody. Issue #20 left the residual TTMn_R partner unnamed. The scan below named it. The authored map is validation/motor_map.yaml. Every actuator and both TTMn neurons are unmapped. A command is refused. It does not select FlyBody.

The documented FlyBody install is Python 3.10 plus optional TensorFlow, Acme, Ray, and CUDA. flymsg requires Python 3.13 and does not depend on MuJoCo. Compatibility of the two environments was not established, and no FlyBody runtime was measured. A DMPO policy would be a learned controller. Issue #8 keeps any learned controller out of the mechanism under test.

The selected body is the one-bit jump. FlyBody stays related work. It is not a dependency of this version.

Encoding and decoding

Defaults are criterion v4 and sim.Params. w_syn is 0.19230769230769232 and th_jump is 0. Seeds are 0, 1, and 2.

LC4 has no class value in MaleCNS v1.0. Its superclass is visual_projection. The class-matched sensory control of criterion v4 is not a deciding check here.

The kernel decides a spike before it adds that step's Poisson kick (sim._kernel, threshold, then delivery, then the kick). The third TTMn_R spike therefore does not depend on the kick of its own step. Zeroing kicks after that step is the cut that later spikes can feel.

Interface

The body state is {escaped: bool, t_jump_ms: ms or null}. Each neural step is 0.1 ms. This body has no integrator and no slower step (body.timestep_ms: null).

On the step where the TTMn_R count reaches 3, set escaped and t_jump_ms to that step's time. Every later step uses a zero kick. If the count never reaches 3, escaped stays false, t_jump_ms stays null, and the drive follows the open-loop 300 ms window.

Replay is fixed by the seed. The runner stores, per seed and per arm:

• the spike-count bins and the first-spike times;
• the LC4_R kick matrix after the mask;
• escaped and t_jump_ms;
• whether DNp01 was silenced.

The same seed redraws the same open-loop Poisson sequence and applies the mask. No second random stream is drawn at the cut.

Coupling

The neural step stays 0.1 ms. A future body step is eligible only when its duration is an integer multiple of 0.1 ms and at least one neural step, and when the sensory input is held constant across the neural steps inside it. behavior.integer_multiple is that test. This version checks the frozen FlyBody numbers with it. It does not step MuJoCo.

The asset file's 0.1 ms option is not the task step. Walking and flight override it. Flight's physics step is finer than the neural step, so it fails the rule. Walking's physics step is two neural steps, and its control step is twenty. Neither is implemented. The selected body declares no second step, so there is no coupling left to integrate.

Deciding checks

All four are required. Seeds 0–2. The default parameters only. A run that has not happened cannot pass them.

decoder_match is an integrity check on the mask. The dynamics are causal and the kicks match through t_jump, so the prefix has to match. A mismatch means the runner is not this protocol.

Reported, and not decisive

The open-loop arm cannot rescue a failed closed loop. No biological latency is copied into a threshold. The 100 Hz Poisson drive is not the loom those papers used.

Wiring

No new shuffle is drawn. N1 already ran looming escape on one degree-preserving shuffle of post, seed 0. Every specific check that passes on the real graph fails on that shuffle, and the looming specific check is one of them (related). A further wiring arm would be a second draw after that result. wiring.draw is false.

Prediction

gf_required_for_the_jump_bit is unverified. The statement is that silencing both giant-fiber neurons keeps TTMn_R below 3 spikes while LC4_R is driven at 100 Hz for 300 ms. The silence arm tested that statement in the model and failed: TTMn_R still reaches 3 spikes. The label stays unverified because the arm does not test the animal. von Reyn et al. 2014 is the behavioral citation already used to choose the looming case. King and Wyman 1980 describes the anatomy of the pathway. It is not a silencing measurement. Until an independent animal measurement is brought in under a new version, the prediction stays unverified and tests_animal: false.

The equality of the closed-loop prefix and the open-loop prefix is the decoder_match check. It is a property of this decoder. It is not a biological prediction.

Limits

• One bit on one calibration circuit. Scope stays there.
• An animation of the bit is not acceptance. Acceptance is the four checks above. The run fails silence.
• No full-body fidelity is claimed, and no subjective experience is claimed.
• No learned controller is added, and its contribution is not estimated because it is absent.
• The Poisson rate is not a loom waveform.
• The model has no gap junction on the giant fiber.
• FlyBody's flight physics step is ineligible under the coupling rule. Its walking steps are eligible only as a documented later option.
• The loop took 90.5 s. The closed-loop count of neurons still firing from 400 ms to 600 ms is 0. That count does not decide.
• A failure does not retune w_syn or th_jump, does not redraw N1, and does not edit the held-out manifest.

Residual partner scan

Issue #20 leaves the partner that still drives TTMn_R with both DNp01 silenced unnamed. This section freezes the measurement that would name it. The lock is validation/silence_scan.yaml (now version 3; version 1 is the file at v0.8.16, version 2 at v0.8.27). The candidate rule, the screen, the draw count, the p and the verdict are code in src/flymsg/silence_scan.py, tested on synthetic graphs. flymsg silence-scan prints the protocol and does not load the connectome; --run executes it with --workers and --checkpoint. The runner was committed before the scan ran. Version 1 of the one-bit protocol is not edited.

Result of version 1 (2026-10-12)

Verdict: none. The run started from 6709f13 (v0.8.17, clean) and took 2 h 46 min. Its manifest records 7d939b2: until v0.8.27 a manifest took the git state when the run ended, and the checkout had moved meanwhile. The code that ran is 6709f13.

The base arm (both DNp01 silenced) gave TTMn_R 7, 5 and 6 spikes on seeds 0–2 and 5, 6 and 7 on seeds 10–12. Thirteen units met the candidate rule. Two passed the screen:

The other eleven stayed between 4.7 and 6.7 spikes. With m = 2 the threshold is 0.025, so neither unit contributes and neither names_partner, although each one, silenced with DNp01, removed every TTMn_R spike in all six seeds. Report: results/silence-scan/v1/ (archived from runs/).

Post-hoc, exploratory (scripts/silence_scan_posthoc.py, after the verdict). The v1 null pool excluded only the tested unit. Of the 1,000 draws for DNp02, 35 reached 0 spikes, and 35 contained DNp06_R (bodyId 10584); 48 contained DNp06_L. Of the 1,000 draws for DNp06, 40 reached 0 spikes, and 40 contained DNp02_R (10117); DNp02_L was not a responder. The counts coincide; per-draw values were not kept, so the identity of the draws is not shown. The reading is that the null for each unit was contaminated by the other, and that the right-hand neuron of each matters. It does not change the verdict.

Version 2 (fixed 2026-10-12, after the v1 verdict, before its run)

Version 2 changes one rule and the seeds: the null pool also excludes every other unit that passed the screen, and confirmation runs on fresh seeds 20–22 (not 10–12). Candidate rule, screen, draw count, threshold and verdicts are those of version 1. The lock records what it supersedes and why. Being written after the v1 result is disclosed here; the v2 confirmation had not run when the lock was committed. A test checks that no null draw for one screen pass can silence another.

Result of version 2 (2026-10-12)

Verdict: names_partner: DNp02, DNp06. The run started from 764af9e (v0.8.27, clean, in a separate worktree) and took 2 h 33 min; its manifest records the same commit.

The base arm gave TTMn_R 7, 5 and 6 spikes on screen seeds 0–2 and 6, 6 and 6 on confirm seeds 20–22. The same thirteen candidates as in version 1; the same two passed the screen:

With the other screen pass out of the pool, no draw reached the observed 0 spikes, so p is the smallest the 1,000 draws allow, below the 0.025 threshold. Report: results/silence-scan/v2/ (archived from runs/).

Reading. In the model, with both DNp01 silenced, TTMn_R is driven through DNp02 and DNp06: silencing either one removes every spike on all nine seeds of both versions. Each is necessary for the residual drive here; neither alone is shown sufficient, and combinations were not searched. The seed 0 replays on the site add one exploratory observation: with DNp02 silenced as well, DNp06_R still fires 27 times and TTMn_R none; with DNp06 silenced, DNp02_R still fires 59 times and TTMn_R none. In this model TTMn_R needs both inputs together, not either one alone. Compare DNp01 against DNp01 + DNp02 and DNp01 against DNp01 + DNp06. Three limits hold. Version 2 was written after the version 1 result was seen, although before its own run; a replication on further fresh seeds would remove that doubt. The result is about this LIF on MaleCNS, not the fly. gf_required_for_the_jump_bit stays unverified, and version 1 of the one-bit protocol is not edited.

Version 3 (replication, fixed 2026-10-12, before its run)

Version 2 chose its null rule after the version 1 result was seen. Version 3 repeats version 2 with every rule unchanged and only the confirm seeds moved to 30–32, which no version has used; the runner refuses seeds 10–12 and 20–22. A test checks against the v0.8.27 file that nothing else changed. The screen runs on seeds 0–2 again, so it selects the same units by construction.

Failure rule, fixed here before the run: if DNp02 or DNp06 does not reach names_partner, the v2 naming is reported as not replicated, #20 is reopened with both results, and neither run is dropped. If both reach it, the naming is reported as replicated on fresh seeds and stays model-side.

One description is added and decides nothing: in each confirm arm, the spikes of the other screen pass per seed (DNp06 when DNp02 is silenced, and the reverse). It turns the seed 0 observation that TTMn_R needs both into three seeds, with no test attached.

Result of version 3 (2026-10-12)

Verdict: names_partner: DNp02, DNp06. The v2 naming replicates on fresh seeds. The run started from 1bbf60e (the lock on main, clean, in a separate worktree) and took 2 h 23 min; its manifest records the same commit.

The base arm gave TTMn_R 7, 5 and 6 spikes on screen seeds 0–2 and 5, 4 and 5 on confirm seeds 30–32. The screen selected the same thirteen candidates and the same two passes:

The failure rule did not apply. The descriptive column, which decides nothing, extends the seed 0 observation to three seeds: with either unit silenced the other keeps firing and TTMn_R stays at 0, so in this model TTMn_R needs both inputs together. Across versions 2 and 3, silencing either unit with both DNp01 removes every TTMn_R spike on twelve seeds. The runner keeps only the null mean and p, not the value of each draw, so the null distribution is not archived. Report: results/silence-scan/v3/.

The limits of version 2 that remain: the result is about this LIF on MaleCNS, not the fly; gf_required_for_the_jump_bit stays unverified; version 1 of the one-bit protocol is not edited. What would test it on the animal is in #36.

Protocol (all versions)

Blind. The 2026-10-12 listing of TTMn_R inputs was read before this lock, so it chooses nothing. No unit is named here. The runner computes the set from the rule, and the set is reported before the confirmatory stage starts.

Arm. The behavior v1 silence arm: LC4_R at 100 Hz, both DNp01 silenced, w_syn 0.192, th_jump 0. The drive is not cut. The statistic is the number of TTMn_R spikes during the 300 ms drive, averaged over seeds. Runs stop at 300 ms because nothing after the drive is read.

Candidates. A unit is a cell type, and a neuron without a type is its own unit. A unit is a candidate when its neurons together make at least 5 synapses onto TTMn_R (the v4 control weight) and at least one of them fires during the drive in 2 of 3 screen seeds of the base arm. A silent unit cannot change the run, so leaving it out does not change the answer. The whole unit is silenced, except LC4_R, TTMn_R and DNp01. Inhibitory units stay in. There is no cap. Every route into TTMn_R ends on a direct chemical input, so the scan covers every single-type route the model has.

Screen (seeds 0–2, no null, does not decide). A unit passes when the mean TTMn_R count with DNp01 and the unit silenced is at or below half of the base-arm mean. The half is a choice made here, not a measured value. If the base arm already keeps TTMn_R at or below 2 spikes in every seed, nothing is named on those seeds.

Confirmation (fresh seeds 10–12). Each of the m units that pass the screen is tested against the T2/T3 null: as many of the base arm's responders on these seeds, matched on superclass and transmitter sign, never LC4_R, TTMn_R, DNp01 or the tested unit. DNp01 stays silenced in every draw. Draws per unit are max(1000, 20 m), so the smallest p, 1/(1 + draws), is below 0.05/m. p = (1 + draws whose mean count is at or below the observed mean) / (1 + draws). A unit whose matched pool is short is reported without a p and cannot pass.

names_partner is the behavior v1 silence bar applied with the unit silenced as well. A unit that contributes without reaching it is a contribution, not the partner. none is a finding. Combinations of units are not searched in this version, and neither are gap junctions, a retune, a redraw of N1 or an edit of the held-out manifest. gf_required_for_the_jump_bit stays unverified: the scan is model-side.

Cost, measured 2026-10-12 under load 13. A 300 ms run takes 3.8 s on one worker. The screen is three runs per candidate plus three base runs. Each confirmed unit needs 3 × max(1000, 20 m) runs, about 3.2 h on one worker at 1000 draws. The run will use --checkpoint.

Residual partner follow-up

Scan v3 names DNp02 and DNp06 on the right side and shows that each is necessary. Issue #76 asks two more questions of the model, locked in validation/partner_followup.yaml before any run. The runner is src/flymsg/partner_followup.py, tested on synthetic graphs. flymsg partner-followup prints the protocol; --run executes it on seeds 40–42.

Seen before the lock: the mirrored arms (exploratory)

scripts/silence_mirror.py ran the scan arms on both sides, seeds 0–2, after the v3 verdict. The right side reproduces the scan base arm. TTMn spikes during the 300 ms drive:

The scan silenced each type on both sides; here the effect follows the pair on the side of the stimulus. The left side is weaker: TTMn_L receives 2,424 synapses against 3,608 for TTMn_R, and with DNp01 silenced only 1–3 spikes remain, close to the 2-spike bar.

Version 1 (fixed 2026-10-12, before its run)

Left. The scan's silence arm mirrored (LC4_L at 100 Hz, both DNp01 silenced, TTMn_L), testing only DNp02 and DNp06 with the scan's matched-responder null (1,000 draws, lower tail, α 0.05 / 2). Verdict: replicates if both reach names_partner, partial if one does, not_replicated if neither, no_residual if the base arm already stays at 2 spikes or fewer, and undetermined if a null is not defined. Because the base arm is weak, no_residual is a likely outcome and is stated here as one.

Drive. No LC4 and both DNp01 silenced. DNp02_R, DNp06_R and the pair are driven directly at 200 Hz for 300 ms, about the rate each fired during the looming drive in the v3 partner counts. An arm reaches the jump bit with at least 3 TTMn_R spikes in every seed. The pair is compared with 1,000 pairs of descending cholinergic neurons (DNp01, DNp02, DNp06 excluded) driven the same way, upper tail. Verdicts, in order: single_sufficient if either alone reaches the bar; insufficient if the pair does not; not_specific if the pair does but p is above 0.05; conjunction if both single arms stay at 2 spikes or fewer; otherwise pair_sufficient.

Every null draw is kept (null_draws.csv). The run starts from a commit on main. Results stay model-side; gf_required_for_the_jump_bit stays unverified.

Result of version 1 (2026-10-13)

The run started from 39b53a5 on main (clean, in a separate worktree) and took 3 h 30 min under load.

Left: replicates. The mirrored base arm gave TTMn_L 3, 1 and 1 spikes on seeds 40–42, a weak residual as expected. With DNp02 or DNp06 silenced as well, TTMn_L stays at 0 on all three seeds: p 0.001 (null mean 2.21) and p 0.003 (null mean 2.15), both below 0.025. The naming holds on the other side.

Drive: single_sufficient: DNp06_R. Driven at 200 Hz without LC4 and with DNp01 silenced, TTMn_R fires 0, 0, 0 with DNp02_R alone, 3, 4, 6 with DNp06_R alone, and 9, 8, 11 with the pair; 1,000 matched descending-neuron pairs give a mean of 0.013 (p 0.001). The conjunction hypothesis is rejected: DNp06 alone reaches the jump bit at the driven rate, and DNp02 adds to it. Read with the scan: during looming DNp06_R fires about 150 Hz (45–50 spikes in 300 ms), below the driven 200 Hz, so silencing DNp02 removes the drive it needs; that is why each is necessary in the looming arm. The pair is specific: random descending pairs almost never drive TTMn_R.

Outputs: results/partner-followup/v1/ (with every null draw).