fast-brain-11: first braking brain motor to fly a full Straw Bale lap (experimental development candidate)

Published on 2026-09-29. Download from
GitHub Releases
or Hugging Face.

Straw Bale lap 1, hilltop and downhill, 4x speed: connectome activity (left) beside gameplay
(right). Brain motors, fast race-cue pilot with the obstacle stack, assisted yaw.

The first brain motor that both brakes for the obstacle stack and flies a full Straw Bale lap.
fast-brain-11-b-cw13 flew lap 1 of the Straw Bale / Field Day race in 1:42.988. fast-brain-08's first
laps took 1:46.358 and 1:46.072, but with a different pilot, so this is a whole-stack comparison. Lap 1 had one
audited ground contact on the downhill, and so did the other brain-11 run that reached it. fast-brain-08 had 3 and 2
in the first laps of its two finishes (7 and 6 per race). Audit counts are indicative only: the audit's video
false-positive check failed. brain-11's roll/pitch commands are as smooth as fast-brain-08's: 0.0025-0.0030
change per 10 ms tick on Straw, against 0.0031*. Unlike fast-brain-08 it slows down for the obstacle governor's
speed caps.

It has not finished a race, and it is not selected by its frozen gates (brain-11 gates v1: 8 of 12). Every
flight below ended in a crash, and the release is published as a development candidate with all of them.
fast-brain-08 remains the published race result (Straw Bale 5:17.805). Straw Bale and Minus Two are seen
development courses; there is no unseen-race or freestyle result.

*Stick change: the mean of |Δroll| and |Δpitch| of the roll/pitch commands per 10 ms tick after the
first 3 s of the log.

Update 2026-09-29: round-6 flights (same weights, newer pilot stack)

Minus Two, 1.5x speed: connectome activity (left) beside gameplay (right). The early brake slows fast-brain-11
before the hairpin wall, turn-first turns it, and it continues past pillar C.

Same checkpoint (44cca3c4…), flown with the round-6 pilot stack on branch
m6 at 9269a62. The round-6 additions are:

• an earlier governor brake for lagging motors (--early-brake on);
• a sighted descent that holds the descent to the ring's line of sight (--sighted-descent on);
• motor assist v4, without the approach rule that stopped the round-5 run at the first arch;
• contact support and a marker-jump check in shadow (logged, not acting).

These are development runs, and several round-6 rules fail some of their frozen gates
(flight card, "Round 6").

• Minus Two (minus-brain11cw13-r6-01): the first brain run through the hairpin and past pillar C.
  • The early brake engaged at the arch before the hairpin (18.5-20.9 s) and at the hairpin wall (21.6-23.7 s).
  • Turn-first turned the drone at 22.5 s after it slowed to 0.76 m/s short of the wall. It left the hairpin at
    23.9 s and passed pillar C at 25.7 s.
  • At 29.6 s it was passing through a low floor-standing arch as the marker switched to the next ring. The pilot
    turned right, and the arch's leg hit the drone at 4.65 m/s.
  • Stick change 0.0046.
• Straw Bale (straw-brain11cw13-r6-01): no ground contact on the downhill.
  • Every earlier brain-11 run touched the straw at about 79.4 s. Here the sighted descent withheld 0.8-1.6 m/s of
    sink while the ring was clipped below the image.
  • It passed the arch whose top bar ended r5-noassist-04.
  • Further down, still descending with the ring clipped below the image, it came down onto the next arch's top
    banner at 87.3 s (5.2 m/s).
  • Stick change 0.0034.
• Still no race finish. The remaining crashes share two causes:
  • turning toward the next ring while still inside a gate;
  • descending onto the next gate's top bar when the ring is out of view below.
• Compilers on the same PC: the postflight checks of both runs found a rustc.exe compile running on the same PC
  (0.71 core on Straw, 6.22 cores at the end of the Minus run). No controller-deadline, camera or telemetry failure was
  recorded.

To fly this stack, check out m6 at 9269a62, extract fast-brain11-inference.zip, and run the command below with
--looming-brake --obstacle-stack on --stale-evidence on --early-brake on --descent-view on --contact-support shadow --sighted-descent on --motor-assist on --marker-jump shadow. The round-6 flight records are in
fast-brain11-round6-flights.zip (sha256 829aecb4aaed864fd905a987f960064d0c81deb44550e70be84e1621244ae0fa), and the
two videos are release assets.

Every flight with this checkpoint

All runs used the original [Copy] New Drone at 6 m/s. The pilot was the fast race-cue pilot with the obstacle
stack (--looming-brake --obstacle-stack on --descent-view on), running on a hidden Anode desktop. The virtual
pad stayed inside the seat (pads_seat_only true in every preflight), and a processed-control ground check ran
for each pad.

The runs used two stacks:

• Round 4b (m4b b4d9e2c): wall pilot v5, descent view v2, vertical guard v4.
• Round 5 (m5 3b2fe2f): adds --stale-evidence on --contact-support on, with wall pilot v6, descent
  view v3 with contact support v3, and stale evidence v2.

The motor assist was off unless stated. Times are from the start of the flight log.

Four further launches never flew: straw-brain11cw13-r4b-noassist-01, straw-brain11cw13-r5-noassist-02 and
-03, and minus-brain11cw13-r5-01. The preflight refused them because compilers were running on the same PC.
The archive includes their preflight records.

The postflight checks of straw-brain11cw13-r4b-noassist-02, straw-brain11cw13-r5-noassist-04 and
minus-brain11cw13-r5-02 found a rustc.exe compile using 0.6-0.7 of a CPU core at the end of the flight. No
controller-deadline, camera or telemetry failures were recorded.

What the crashes point to (open problems, not fixed in this release):

• Gate structures after a checkpoint switch: the pilot turns toward the next ring while the drone may still
  be inside or beside a tall arch (r5-01), and a false ring-marker reading turned it into an arch leg (r4b-02).
• Descending onto a gate's top bar: with the ring clipped at the bottom of the image, the descent line
  carried the drone onto the next arch's bar (r5-04).
• Straw downhill contact: both runs that reached the downhill touched it at the same place after a ~24-degree
  descent, and both recovered. No camera cue measures the straw there.
• Minus hairpin: a brain that follows about 0.3 s late needs about 1 s of braking. The governor started
  braking only ~0.75 s before the wall, although looming had read a TTC of ~1 s earlier.

What it is

Only readout.weight and readout.bias rows 0-2 (throttle, roll, pitch) changed relative to the parent, motor10
candidate05, which is also the parent of fast-brain-06 to -10. The largest changes are 0.290 (weights) and 0.0102
(biases). The yaw readout, connectome wiring, transmitter signs and the 38 other tensors are bit-identical. The PD
teacher is used only offline, and scene currents are blanked.

Training. DAgger distillation of the fast PD in the measured-drone surrogate:

...

fast-brain-10b and fast-brain-09b: experimental braking brain motors (development candidates)

Published on 2026-09-27. Download from
GitHub Releases
or Hugging Face.

Minus Two hairpin, real time: connectome activity (left) beside gameplay (right).
fast-brain-09b brakes, the round-4 turn-first rule stops it short of the wall and
turns it toward the next ring; accelerating out of the turn it then loses height and
hits the floor.

Brain motors that brake for the obstacle stack, not yet a race result.
fast-brain-08 (the published Straw Bale finisher) does not slow down when the
obstacle governor asks for 3-4.5 m/s, so it cannot make the Minus Two hairpins. These
two readouts were distilled with synthetic governor speed caps so that they brake:

• fast-brain-10b is as smooth as fast-brain-08 in flight (roll/pitch command
  change 0.0031 per 10 ms tick*; fast-brain-08's published race value is also 0.0031). Live
  it braked for the governor's caps at both arches before the hairpin (to 2.8-2.9 m/s
  under 2.7-2.9 m/s caps, somewhat later than fast-brain-09b). At the hairpin itself,
  while the vertical guard's false 1 m/s terrain climb overrode the pilot's descent
  (0.85 to 1.9 m, over the ring), it held 4.9-5.3 m/s under 3.6-4.0 m/s caps and hit
  the wall.
• fast-brain-09b brakes live: it slowed to 2.2-2.5 m/s at the arch before the
  hairpin in both flights. In round 3 it slowed to the ~3 m/s cap before grazing the
  hairpin wall; in round 4 it was braking (5.1 to 3.5 m/s in 0.5 s, still 0.6-1.6 m/s
  above the 2.8-3.6 m/s caps) when the new turn-first rule engaged and stopped it at
  0.5 m/s short of the wall. It is about twice as rough (0.0056-0.0063 per tick) and
  loses height when it accelerates hard from low speed.

*Stick change: the mean of |Δroll| and |Δpitch| of the roll/pitch commands per 10 ms
tick, after the first 3 s of the log (launch excluded); over these short flights the
launch would otherwise dominate.

Neither checkpoint finished any race, and neither passed all of its frozen gates
(fast-brain-10b fails 1 of 7, fast-brain-09b 2 of 5; see below), so neither is
selected. They are published as development candidates with their full evidence;
fast-brain-08 remains the published race result (Straw Bale 5:17.805). Neither has
flown Straw Bale or Pine Valley. Minus Two is a seen development course; there is no
unseen-race or freestyle result.

What they are

Both keep fast-brain-08's parent (motor10 candidate05), contract (6 m/s nominal,
scaled speed 2.4, vertical goal = request x 0.4 s) and blanked scene currents. Only
readout.weight and readout.bias rows 0-2 (throttle, roll, pitch) changed; the yaw
readout, connectome wiring, transmitter signs and all 38 other tensors are
bit-identical to the parent. The PD teacher is used only offline.

The gates were frozen before any candidate was scored (brain-09: 22 candidates;
brain-10: 6). The frozen brain-10 rule allows the best-ranked failing candidate
(fast-brain-10b) to fly only as a disclosed development deviation; fast-brain-09b
flew under the same disclosure. Training ran on the CPU: a benchmark DAgger iteration
(10 drones, 20 s of flight, then a refit) took 39.5 s on two CPU threads and 36.8 s
with CUDA, because most of each tick is per-drone Python (pilot, teacher, simulator),
not the connectome; each full collection took about 45 min on the CPU.

Live evidence (Minus Two / Turn Signals, 6 m/s)

Original [Copy] New Drone, fast race-cue pilot with the obstacle stack
(--looming-brake --obstacle-stack on), hidden Anode desktop, virtual pad kept inside
the seat (pads_seat_only true in every preflight and postflight), processed-control
ground check per pad.

Times are from the start of the flight log. The round-4 stack's rules each failed at
least one of their own frozen offline gates, so these flights are development
evidence (flight card).
In the same round the fast PD passed pillar C for the first time at 6 m/s (after
pillar A and the hairpin), then touched the garage floor for 1.5 s and struck the
ceiling after a false terrain climb. Fixes for the false climb, contact detection and a
brain that both brakes and holds height are in progress.

Limits

• No race finish with either checkpoint; one development course (Minus Two) flown.
• fast-brain-10b did not follow the caps at the hairpin (+1.0 to +1.7 m/s, mean +1.5)
  during a simultaneous false guard climb. This is one episode, so it is untested
  whether it misses caps in general or only while climbing; its surrogate cap excess is
  +0.35 m/s.
• fast-brain-09b brakes but is rough and loses height when it accelerates hard from low
  speed near the floor.
• The readout moved much further from its parent than fast-brain-08's did; how that
  transfers across courses is untested.
• The obstacle stack they were flown with lives on branch m4 and is not merged into
  main; its rules are off by default.

Download and use

Extract fast-brain10-inference.zip into the repository root of branch
m4 (the round-4 flights ran at
9a41acb). It contains both checkpoints, their configs, weight audits, training
source, logs and the surrogate evaluations; fast-brain-09b's collection run; the
frozen gate files and scored results; the GateNet detector, stick mapping and
measured dynamics profile the runs used; the four flight sidecars with their
preflights; the ground checks; and per-file hashes (fast-brain10-manifest.json).
It also contains the 16-course gate harness the gate files pin by hash
(tools/brain_gates/harness16.py; the gate files record the absolute path it had
when the gates were scored). Training examples (~1.1 GB each) and raw flight CSVs are
kept locally, not included.

Prerequisites for the obstacle stack: its gap cue needs a CUDA GPU, the
transformers package and Depth-Anything-V2-Small
(depth-anything/Depth-Anything-V2-Small-hf,
revision 5426e4f0f36572d16453bbda7a8389317b1bef99, Apache-2.0; model.safetensors
sha256 3152477ce0d8d6978d76b995120de97cb5b928701fd0f817769f59e249a16b70). Put its
config.json, preprocessor_config.json, model.safetensors and model card
README.md in runs/dense-depth-probe-20260923/model, or set
HALTERE_RELATIVE_DEPTH_MODEL_DIR; nothing is downloaded automatically. The flight
limits below are the ones every published run used; without them the runner's
defaults stop a flight after 20 m.

.venv/Scripts/python.exe -m haltere.liftoff.visual_brain runs/fast-brain-10b/candidate.pt `
  --mapping runs/pine-route-collection-01/liftoff-original-drone.yaml --device cpu --vision-device cuda `
  --pilot-assistance race-cue --pilot-profile fast --assist-speed 6 --motor-controller brain `
  --dynamics-profile runs/measured-dynamics-low-speed-20260923/profile.json `
  --looming-brake --obstacle-s...

Reviewed on 2026-09-25. Download from
GitHub Releases
or Hugging Face.

Straw Bale lap 1, hilltop through the descending checkpoints, at 4x speed:
connectome activity (left) beside gameplay (right). Brain motors, fast race-cue
pilot with the downhill fix, assisted yaw.

The fastest brain-motor races so far, without the side-to-side weave on the
downhill. Under the fast race-cue pilot at a declared 6 m/s, fast-brain-08
finished the full three-lap Straw Bale / Field Day race twice: 5:17.898 and
5:17.805. Neither the runtime impact guard nor the offline contact check found
an impact, but the pilot's slope-support rule fired 4 and 6 times on the downhill,
where the drone skims the straw, so light contact with the hillside is not
excluded. fast-brain-07 took 5:45.792 and 5:46.846 with the previous pilot; the
fast PD's 5:02.933 was flown with an earlier pilot (before arc turns and the
downhill fix) and has not been flown with this one. Minus Two and Pine Valley were
not finished: obstacles on or beside the line to the next checkpoint, as for
fast-brain-06, fast-brain-07 and the fast PD at 6 m/s (the slower scene09 stack did
finish Minus Two). These are seen development courses, not an unseen-race or
freestyle result, and far from the 1:34.807 target.

The same downhill, 2x speed: fast-brain-07 with the previous pilot (left) yaws
back and forth; fast-brain-08 with the fixed pilot (right) holds its line.

What changed from fast-brain-07

• Pilot: no yaw sweep on the downhill. A slow yaw sweep meant for rings
  clipped at the top edge also ran on rings clipped at the bottom, so every
  downhill wove from side to side. It now runs only for top clips, and a ring that
  stays clipped below makes the descent steeper over time
  (downhill card). This pilot change is
  what removes the weave.
• The brain can follow that pilot. fast-brain-07 crashed 2/2 with the fix: it
  barely braked or sank when asked. Replays against the teacher showed a
  distillation gap, not a simulator mismatch: the readout's ridge of 10 held
  braking back (a pitch gap of +0.78 to the teacher on steep, slow ticks), and the
  goal neurons stop responding beyond a vertical goal of about 1 m. brain-08
  encodes the vertical goal as the request x 0.4 s (was 1.0 s) and was distilled
  under the fixed pilot with steep-sink samples up-weighted, a lower ridge (0.3)
  and a penalty on the command change per 10 ms tick
  (brain-08 card).
• Pilot: support on a slope. Sliding down a hillside still sinks at the
  hill's slope, so the old support rule (sink stopped) never fired. A sink
  shortfall with the issued throttle held below hover now also triggers a short
  climb.
• Pilot: gentle search. A lost ring used to trigger a stop at the 10 m/s²
  command limit; near the ground the pitch-up and the sharp turn that followed hit
  the hilltop twice. The search now slows at up to 3 m/s² and rises at 0.5 m/s
  for 1.5 s. In the surrogate, with the ring blanked 1.4 s every 9 s, brain-08
  crashed on 2 of 16 courses instead of 8 with the old search (emulated as a
  10 m/s² stop without the climb).

Weight audit against the parent (motor10 candidate05): only readout.weight and
readout.bias rows 0-2 changed (max 0.026 / 0.0009); the yaw readout, connectome
wiring, transmitter signs and every other tensor are identical. The teacher is
absent at inference.

Evidence

Live on Straw Bale, brain-08 (pilot at e71f4c0) against brain-07 (pilot at
58a0cfe), same course and drone. Both the weights and the pilot differ between
the columns, and no same-pilot brain-07 or PD race exists (brain-07 crashed 2/2 on
the fixed pilot), so these are whole-stack results. Every metric is computed with
the same code for both; pairs are in run order (brain-07 runs 02, 03; brain-08 runs
04, 06):

Primary switches are those more than 1 s after the previous one. Stick commands
are smoother over the whole race and equal after switches, and yaw is much calmer
on the downhill; roll/pitch body rates are about 10% higher than brain-07's, and
sharp (45-75 degree) switches are slower and slightly rougher.

Every attempt with this checkpoint (original [Copy] New Drone, hidden Anode
desktop, CPU brain, CUDA vision, 6 m/s, processed-control ground check per pad):

Times are from the start of the flight log.

Surrogate (not flight evidence): on the 16-course development gate (8 flat and 8
steep synthetic courses; the steep ones are the distillation's evaluation courses
and the gate chose between refits), with the downhill-fixed pilot before slope
support and gentle search (c6e57c5; brain-08's results are identical with the
released pilot), brain-08 finished 15/16 with 1 crash (brain-07: 15/16, 0 crashes,
1 timeout), steep-request sink shortfall 0.16 m/s (brain-07 0.69, PD 0.15),
roll/pitch stick chatter 0.0030 (0.0027). The finish count accepts a pass anywhere
within 3 m of a checkpoint, which hid brain-07's high descending passes (5 of 10
more than 1.5 m above the checkpoint, brain-08 0 of 10). These gate outputs are
not in the archive; the bundled evaluation.json is the refit's own 8-course
check (8/8, chatter 0.005, steep shortfall 0.28, its own metric definitions). The
full suite passed 787 tests at 09898a1.

Limits

• Seen development courses only; unseen races and freestyle are untested.
• No obstacle sense beside or on the line to a checkpoint (Minus Two pillar, Pine
  Valley terrain and boulder); a learned clearance model is in development.
• brain-08 tracks the pilot's requested line more closely than brain-07, which
  climbed well above it; near the ground it relies on the pilot's slope support
  and gentle search, and it touches the straw on the downhill. Sharp checkpoint
  switches are slower and slightly rougher.
• Runs 04-06 and the Minus Two and Pine Valley flights ran with an Anode 0.9.0
  build whose HidHide seat isolation kept the virtual pad inside the seat
  (pads_seat_only true in preflight and postflight); runs 01-03 predate that
  check. Without it, the pad is machine-wide and the game guard unplugs it when
  another game runs.

Download and use

Extract fast-brain08-inference.zip into the repository root at 09898a1 or a
later compatible revision (the finishes ran at e71f4c0, which differs only in the
flight-log writer). It contains the checkpoint, the GateNet detector and stick
mapping it expects, the measured dynamics profile, the training source and
configuration of both the collection run and the refit, the audit, evaluation,
flight sidecars, ground checks, finish frames and per-file hashes
(fast-brain08-manifest.json).

.venv/Scripts/python.exe -m haltere.liftoff.visual_brain runs/fast-brain-08-vgs04-s10r03m30/candidate.pt `
  --mapping runs/pine-route-collection-01/liftoff-original-drone.yaml --device cpu --vision-device cuda `
  --pilot-assistance race-cue --pilot-profile fast --assist-speed 6 --motor-controller brain `
  --dynamics-profile runs/measured-dynamics-low-speed-20260923/profile.json `
  --udp-out 127.0.0.1:9003 --pause-on-stop --log runs/my-flight.csv

Reviewed on 2026-09-25. Download from
GitHub Releases
or Hugging Face.

Straw Bale lap 1, hilltop through the descending checkpoints, at 4x speed:
connectome activity (left) beside gameplay (right). Brain motors, fast race-cue
pilot with arc turns, assisted yaw.

The fastest and smoothest brain-motor races so far. Under the fast race-cue
pilot at a declared 6 m/s, fast-brain-07 finished the full three-lap Straw Bale /
Field Day race twice: 5:45.792 and 5:46.846, with no detected contact.
fast-brain-06 took 5:50.204 and 5:50.489; the earlier brain best was 14:05.703.
Minus Two and Pine Valley were not finished, as with brain-06 and the fast PD:
obstacles on or beside the line to the next checkpoint. These are seen development
courses, not an unseen-race or freestyle result, and far from the 1:34.807 target.

What changed from fast-brain-06

• The pilot flies checkpoint switches as coordinated turns. When the ring
  jumped to the next checkpoint, the old pilot braked hard and re-accelerated
  (a pitch swing the user saw as jitter). It now rotates the velocity request at
  up to 8 m/s² centripetal and follows a side-clamped ring at 65% speed instead of
  30% (fast racing, commit 58a0cfe).
• The brain was re-distilled under that pilot (same recipe as brain-06:
  FastMotorPD labels, DAgger in the measured-drone surrogate, readout rows 0-2
  only, scene currents blanked), so it learned to carry speed through a turn.

Weight audit against the parent (motor10 candidate05): only readout.weight and
readout.bias rows 0-2 changed (max 0.0060 / 0.0005); the yaw readout, connectome
wiring, transmitter signs and every other tensor are identical. The teacher is
absent at inference.

Evidence

Live on Straw Bale, brain-07 against brain-06 (same course and drone):

Every attempt with this checkpoint (original [Copy] New Drone, hidden Anode
desktop, CPU brain, CUDA vision, 6 m/s, processed-control ground check per pad):

The downhill fix removed a real yaw weave, but brain-07 does not follow the
steeper descent and slower horizontal request it asked for, so without the weave
it reached a lower ring from above. Those pilot changes were reverted; see the
downhill flight card. Surrogate evaluation
(not flight evidence): 7/8 steep held-out synthetic courses, 0 crashes. The full
suite passed 596 tests at a7dbbac.

Limits

• Seen development courses only; unseen races and freestyle are untested.
• No obstacle sense beside or on the line to a checkpoint (Minus Two pillar, Pine
  Valley terrain and boulder); a learned clearance model is in development.
• On long descents the pilot's yaw sweeps side to side (see above); the fix waits
  for a brain that tracks steeper descents.
• The virtual gamepad used for Liftoff is machine-wide: do not play another
  gamepad game while it flies.

Download and use

Extract fast-brain07-inference.zip into the repository root at a7dbbac or a
later compatible revision. It contains the checkpoint, the GateNet detector and
stick mapping it expects, the measured dynamics profile, training source,
configuration, audit, evaluation, flight sidecars, ground checks, finish
screenshots and per-file hashes (fast-brain07-manifest.json).

.venv/Scripts/python.exe -m haltere.liftoff.visual_brain runs/fast-brain-07-6mps-arc/candidate.pt `
  --mapping runs/pine-route-collection-01/liftoff-original-drone.yaml --device cpu --vision-device cuda `
  --pilot-assistance race-cue --pilot-profile fast --assist-speed 6 --motor-controller brain `
  --dynamics-profile runs/measured-dynamics-low-speed-20260923/profile.json `
  --udp-out 127.0.0.1:9003 --pause-on-stop --log runs/my-flight.csv

Videos (release assets): the full 5:45 Straw Bale finish, a 4x version, and the
Minus Two and Pine Valley failures, each with connectome activity beside gameplay.

Reviewed on 2026-09-24. Download from
GitHub Releases
or Hugging Face.

Straw Bale lap 1, hilltop through the descending checkpoint, at 4x speed: connectome
activity (left) beside gameplay (right). Brain motors, fast race-cue pilot, assisted yaw.

The first brain-motor finishes at race speed. Under the fast race-cue pilot at a
declared 6 m/s, fast-brain-06 finished the full three-lap Straw Bale / Field Day
race twice: 5:50.489 and 5:50.204, with no detected contact. The previous
brain best on this race was 14:05.703 (scene09, 2.5 m/s pilot). The matched fast PD
with the same pilot and code finished in 5:02.933, so the brain is 16% slower than
its teacher. Minus Two and Pine Valley were not finished; the fast PD fails
them too, on obstacles lying on or beside the line to the next checkpoint.
This is a seen development course, not an unseen-race or freestyle result, and it
is far from the 1:34.807 Straw Bale target.

What changed

• Motor readout only. The fast velocity PD (FastMotorPD, measured thrust and
  rate curves of the original [Copy] New Drone) was distilled into the throttle,
  roll and pitch rows of readout.weight and readout.bias by DAgger in the
  measured-drone surrogate on seeded synthetic checkpoint courses (40% steep legs,
  5 rounds x 10 courses, speed-balanced ridge fit from the parent each round).
  The audit against the parent, motor10 candidate05, confirms that the yaw readout,
  connectome wiring, transmitter signs and all other saved tensors are identical
  (max change 0.0061 in the weight rows). The PD teacher is absent at inference.
• Scene currents blanked. Replays of live flights showed that the recorded
  scene currents shifted brain-05's mean throttle by up to 0.2 and pitch by 0.1
  depending only on which images it saw; live Straw Bale images meant less braking
  and more thrust, so it could not follow slow or descending requests. brain-06
  was trained with a zero retina and the checkpoint records
  recorded_scene_currents: false; the runtime then zeroes the brain's retina
  input. The camera still provides the checkpoint-ring cue to the pilot.
• Declared contract. The pilot's world velocity request becomes a body-frame
  goal (horizontal x0.4 s, vertical x1 s) and horizontal velocity senses are
  scaled by 0.4, so 6 m/s looks like 2.4 m/s to the brain. Speeds above 6 m/s are
  refused.

The pilot is the disclosed fast race-cue pilot: world velocity
along the causal bearing of the visible next-checkpoint HUD ring, continuous turn
slowdown, bottom-edge descents, a descent path governor and assisted yaw. No
course file, route, per-course parameter, teacher or navigation predictor is loaded.

Evidence

All attempts with this checkpoint, original drone, hidden Anode desktop, CPU brain
and CUDA vision, one configuration (6 m/s, no looming) unless noted:

Main-race batch with the declared configuration: 1/3 courses (2/2 Straw Bale
finishes after the guard fix). On the Straw finishes the brain's median speed was
4.87 m/s (PD 5.89), roll/pitch command change per 10 ms tick 0.0050 (PD 0.0055)
and body-rate RMS 0.66-0.69 rad/s (PD 1.04-1.09). Surrogate evaluation (not flight
evidence): 6/8 steep held-out synthetic courses finished, 0 crashes. Each flight
passed a processed-control ground check. The full suite passed 579 tests at
d5dd2d2.

Limits

• Seen development courses only; unseen races and freestyle are untested.
• No obstacle sense beside or on the line to a checkpoint. Three offline side
  cues (split-field looming, optic-flow balance, monocular depth) did not detect
  the Minus Two pillar or Pine Valley boulder in time (flight card).
• The brain's sideways response lags the request by about 0.3 s and it flies
  below the requested speed on straights (about 5.5 of 6 m/s).
• Straw Bale in 5:50 is 3.7x the 1:34.807 target.

Download and use

Extract fast-brain06-inference.zip into the repository root at d5dd2d2 or a
later compatible revision. It contains the checkpoint, the GateNet detector and
stick mapping it expects, the measured dynamics profile, training source,
configuration, audit, evaluation, flight sidecars, ground checks, finish
screenshots and per-file hashes (fast-brain06-manifest.json). The connectome
graph is tracked in the repository. Training examples (~1 GB) and raw flight logs
are retained locally and not included.

.venv/Scripts/python.exe -m haltere.liftoff.visual_brain runs/fast-brain-06-6mps-noscene/candidate.pt `
  --mapping runs/pine-route-collection-01/liftoff-original-drone.yaml --device cpu --vision-device cuda `
  --pilot-assistance race-cue --pilot-profile fast --assist-speed 6 --motor-controller brain `
  --dynamics-profile runs/measured-dynamics-low-speed-20260923/profile.json `
  --udp-out 127.0.0.1:9003 --pause-on-stop --log runs/my-flight.csv

Videos (release assets): the full 5:50 Straw Bale finish, a 4x version, and the
Minus Two and Pine Valley failures, each with connectome activity beside gameplay.

The frozen geometry recovery experiment completed 2 of 3 full one-lap races on the known generated box course, using the original [Copy] New Drone. The two finishes were 2:28.763 and 2:39.492, confirmed by the game. The remaining attempt stalled at the final wall and was stopped under the declared 60-second/1-metre rule. All three videos are included and fully decode.

The videos show the standard brain panel beside gameplay, labelled PD motors; brain in shadow. The newest candidate05 brain is loaded, but it does not control the motors in this diagnostic. No neural weights changed. There were no detected impacts, resets, camera failures, controller deadline failures or geometry-freshness stops.

Runtime source: 946056e032a0c1bd4db8aa86c99bbb7b26a3e987. Identical 2.5 m/s assistance, model, settings and original drone across the batch. The archive includes all three logs, exact geometry input images, replay data, preflight/result cards, saved game evidence and frozen runtime source. One refused preflight and its operational PID-list amendment are preserved; both preceded any flight.

This is known-course development evidence. It does not establish 3/3 reliability, unseen-course transfer, brain-controlled obstacle avoidance, or the five main-track speed target. Two earlier memory revisions finished 0/3 each; the linked project record preserves those outcomes. A later tiny-detour selection fix is outside this flown source.

Model provenance and existing weights: motor10 experimental release. This release contains evaluation evidence, not new weights.

The first matched visual-geometry comparison on the known generated box course finished 0/1 with geometry off and 1/1 on (2:26.657). Two additional unchanged assisted attempts, declared together afterwards, finished once (2:24.148) and hit the wall once. Combined assisted results: 2/3, not reliable 3/3.

These are autonomous camera/telemetry/visible-race-cue flights, using the original [Copy] New Drone, PD motors and motor10 candidate05 in shadow. No stored route or course geometry entered control. No neural weights changed. This does not establish brain-controlled obstacle avoidance, unseen-course transfer, freestyle, or the five main-track speed target.

All four complete standard brain/gameplay recordings decode. No camera or controller deadline failure occurred. Successful runs stopped on live_pose=false when the game entered its results screen. The failed repeat lost nearby obstacle memory while nearly stationary, then resumed descent into the wall; it is retained alongside the successes.

• 01-geometry-off-wall-impact.mp4: matched off condition; incomplete at 1:56.918.
• 02-geometry-on-full-finish.mp4: matched on condition; full lap in 2:26.657.
• 03-geometry-on-repeat-finish.mp4: unchanged repeat; full lap in 2:24.148.
• 04-geometry-on-repeat-wall-impact.mp4: unchanged repeat; incomplete at 2:17.030.
• geometry-control-evidence.zip: frozen manifests and runtime source, all flight logs and neural replay inputs, exact geometry-worker camera images, accepted points/proposals, ground check, terminal evidence and scorer/review cards. Existing neural weights are pinned by hashes, not duplicated.
• asset-manifest.json: sizes and SHA-256 hashes.

Runtime source: 732e140e06b0ee1132e24cb33a9cef87bd718756. The release target adds the reviewed evidence record; it does not retroactively change the flown runtime. Liftoff/capture/control ran inside hidden Anode, with the authorized idle LitHarness family disclosed in pre/postflight checks.

Sparse geometry, primitive motion prediction, unmeasured display delay and the observed memory-expiry failure remain limitations. Retained points do not certify unknown space. This is an experimental diagnostic release, not a promoted model.

Three complete standard brain/gameplay recordings retain the entire frozen dual-keyframe/consistent-detour batch, runtime c76722db6b7058b0487e7e4b10c8ecf1fbee9d71. Original [Copy] New Drone, known generated box-calibration course, one lap, PD motors, motor10 candidate05 in shadow, causal visible race cues and image/telemetry geometry at 2.5 m/s. No course geometry or learned depth model at runtime. No new neural weights.

Both finishes are confirmed by the saved game results. All three untrimmed videos fully decode. No resets or human flight interventions occurred. No camera or controller deadline stops occurred. Worker p95 was 104.2, 307.8 and 137.5 ms. The second attempt frequently rejected aging geometry proposals and spent several minutes at the wall. The third descended through a gap in observed coverage and then advanced into the visible wall; an automatic impact stop paused the game.

This is 2/3 on one known development course with conventional motor control. It does not establish reliable obstacle avoidance, brain-controlled navigation, held-out transfer, freestyle capability or the requested five-track speed target. Earlier failures remain retained separately. The evidence archive includes frozen manifests, preflight/result cards, telemetry, exact sensory replays and geometry-worker images. Videos explicitly label the brain as running in shadow.

Measured response of the original [Copy] New Drone, with standard gameplay and brain-panel recordings. These are empty-arena system-identification flights: PD plus input pulses, newest motor10 candidate05 brain in shadow. They are not autonomous race or freestyle completions, and do not change brain weights.

• Initial hover and all 12 throttle pulses completed. The first full-roll pulse exceeded the tilt limit; that complete failed recording is included.
• Revised bounds completed 24 roll, 24 pitch and 24 yaw pulses, with stable recovery after each.
• An independently declared batch at previously unused amplitudes 0.15, 0.35, 0.65 and 0.85 also completed all 72 pulses. Frozen prediction errors were 6.20, 7.68 and 5.69 degrees/s for roll, pitch and yaw, within the declared overall and per-amplitude limits, without refitting.
• No detected impact, camera failure or control-deadline failure occurred in these nine flight attempts. Ground-check binding failures and workload launch refusals remain separate in the evidence.

Runtime revisions: initial 9220911, revised 22c3014, independent 4fb5fda. Brain SHA-256: 64444a628c58b2c1615086bc84f8c5d3c0a0a12667df31d888c4587a802288b2.

Native telemetry timing matters: processed input applies forward, while quaternion differences measure an interval average. The empirical angular curve is an observed response model, not a claim about Liftoff internals. Sub-frame response time, high-speed drag, full trajectories, acrobatics, obstacle clearance and main-track speed remain unqualified. All raw attempts are retained; the evidence includes the initial incorrect response fits as well as their corrected analysis.

One failed generated-course diagnostic, with the untrimmed standard brain/gameplay video and reproducible image/motion geometry data.

• 0/1 complete courses. Five of six sections crossed geometrically, then a wall collision. HUD: lap 1/1, 1:55.611. No resets, operator flight interventions, camera failures or control-deadline stops.
• PD motor control; motor10 candidate05 brain in shadow. Visible race cues, 2.5 m/s, original [Copy] New Drone. Recorded runtime fd7cbce079c874cebcfba3d1df11c9c1e1f91137, Anode 0.8.0. No runtime course geometry, navigation predictor or live depth steering.
• The archive preserves the preflight manifest, raw logs, neural replay, generated course and offline geometry, terminal screenshot, all 599 passive images, and successive depth experiments. Extract it at the repository root without overwriting existing evidence.
• The current causal image/motion replay warns about the wall 1.98 s before impact, but still warns during successful passages. It has no live authority and does not certify free space or demonstrate obstacle avoidance. Pretrained metric-depth scale was inconsistent and that model remains outside steering.
• No new fly-brain weights were trained or promoted. This release does not meet the five-track speed/completion target and is not a generalization benchmark.

See the full diagnostic and limits and replay instructions.

Validation: 375 automated tests pass; full flight video decodes. Causal replay is checked with and without offline geometry labels to ensure labels do not affect predictions.