Bounded Embodiment: hostile run on an SO-ARM101 arm, measured by its own encoders
9 October 2026, 17:39:40 to 17:49:40 US Pacific (600 s)
https://craigmerry.com/bounded-embodiment/

WHAT THIS IS
The raw log behind the numbers on the "First results" slide of my ITU workshop talk
(16 October 2026): 57,350 commands, 0 samples outside the declared box, top tool speed
0.23 m/s against 0.25 declared, closest approach 10.2 mm.

- Arm: SO-ARM101, 12 V STS3215 servos, driven through LeRobot 0.6.1's driver.
- Commands: a seeded hostile generator (seed 7, amplitude 1.0) switching between fast,
  partial, boundary, malformed, uniform, sign-flip, max and drive-out strategies. This is an
  EARLIER generator, not the Bounded Embodiment method's be-hostile/1 stream.
- Gate: the reference gate, with LeRobot's own step limit off (max_relative_target=None).
- The run would have aborted if the tool went 5 cm outside the box. It finished.

WHAT THIS IS NOT
- Not an EV-03 result. EV-03 requires an external instrument. Every position here is
  forward kinematics of the arm's own encoder readings.
- Those encoder positions are approximate. The joint offsets in envelope.json (joint_map) were
  fitted from two photos to about +/-10 degrees, so the tool positions, and the 10.2 mm
  closest approach, carry that calibration uncertainty. The camera-measured run is next.
- Not scored by the bounded-embodiment core package, and not signed. It predates the kit.
- Not verified by anyone else. One maintainer, no third-party verification.
- The 10.2 mm is against THIS envelope (envelope.json). Scored against a different box,
  the same motion can be outside it.

FILES
run.csv.gz      one row per command (57,350 rows), columns below
envelope.json   the declared envelope the run was scored against (base frame, metres)
summary.json    the runner's own summary, as written at the end of the run
console.log     the runner's console output
recheck.py      recomputes every number above from run.csv.gz + envelope.json (stdlib only)
SHA256SUMS      checksums of the files above

COLUMNS (run.csv.gz)
wall_time                  Unix time, seconds
t                          seconds since the first command
strategy                   which hostile strategy produced the command
decision                   the gate's decision: allow, clamp, reject_hold, invalid_hold
present_<joint>            joint positions read from the encoders (LeRobot degrees)
sent_<joint>               joint positions the gate sent to the servos (LeRobot degrees)
fk_x, fk_y, fk_z           tool point from forward kinematics of present_*, metres
tracking_mm                distance between commanded and measured tool point, mm
error                      runner error, if any (empty in every row)
gate_ms                    time the gate took to decide, ms
elbow_*, wrist_*           elbow and wrist joint centres from forward kinematics, metres

RECHECK
    python3 recheck.py
prints: 57,350 commands over 600.0 s; 0 position, 0 speed and 0 link excursions;
top tool speed 0.23 m/s; closest approach 10.2 mm (to the box's +y face).

LICENCE
CC BY 4.0. Attribute to Craig Merry and link to this README.