Datasets:
physicalai-bmi
/
punch-energetics

Dataset card Data Studio Files
xet
 Community
2
text
stringlengths
0
163
# Release manifest — what is actually on the Hub
Every file below was uploaded to https://huggingface.co/datasets/physicalai-bmi/punch-energetics
and verified present by downloading the bundle back.
WHY THIS FILE EXISTS. TR-2026-45 shipped stating that the benches behind its Sections 3 through 7 were
released at that URL. They were not: the bundle had been uploaded before those benches existed, so a
published paper carried a false availability claim for five of its own sources. Nothing caught it. The
named-artefacts gate deliberately skips any block containing a URL, because a filename next to a link is
normally a citation rather than a promise, and that exemption is exactly the hole this fell through.
AND WHY IT NOW RECORDS DIGESTS. The first version of this manifest listed FILENAMES. That closed the
failure it was built for and left the neighbouring one open, and the neighbouring one then happened:
`thermal_shadow.rs` was corrected here after its release — a headline retracted, prior art cited, a gate
added — while its name sat in this file and the check stayed green. A released file that has since been
corrected is worse than an unreleased one: a reader who downloads it gets a confident wrong answer instead
of a missing file. Each line below is now `<sha256 prefix> <path>`, and a name whose bytes no longer
match the release is a failure.
The digests are taken from a copy DOWNLOADED BACK from the Hub, never from the working tree, and the
regeneration command refuses the working tree for that reason. A manifest built from local files would
faithfully record the bytes someone MEANT to upload, which is exactly the claim that has already been
wrong here once. This file is the receipt of a download, not of an intention.
Regenerate with: python3 scripts/check_release_manifest.py --update <downloaded-bundle> <date>
uploaded: 2026-08-28
files: 28
--- files ---
f58e72053f481bd6 DATASHEET.md
0e7e5618488dc3af LICENSE
b75dc217184754db README.md
49cdec54632659d8 bench/freefall_reorient.rs
b862b6c5857f4637 bench/preflex_latency.rs
88015bb6f13f7383 bench/punch_demand_frontier.rs
9214bd1697ecf79c bench/punch_energetics.rs
082d5093eaa1fea7 bench/punch_reward_learning.rs
3ace4197edf30d34 bench/punch_verify_c.rs
7a384aae82f8291f bench/standing_load.rs
4b03b0a5400bcb0d bench/thermal_shadow.rs
f56c73ff4c1ebe60 bench/total_task_energy.rs
138c9e112d84daf4 bench/windlass_gearing.rs
c6ba15ddd9b1e3fe bench/yielding_leg.rs
19a498622f8237b8 freefall_reorient.out.txt
c69f84c6c6b288e5 preflex_latency.out.txt
32f2125fb0d7a221 punch_demand_frontier.out.txt
1a64c0aa79ad9b99 punch_energetics.out.txt
e8fc9d4ec5b4467c punch_reward_learning.out.txt
231ec459e3995897 punch_verify_18.out.txt
e0303456398dc209 punch_verify_22.out.txt
a923bbc7cf8e080c punch_verify_9.out.txt
7221b4fefb835784 punch_verify_c.out.txt
7019200e6dc000e3 standing_load.out.txt
27d28ee9d4c47ebe thermal_shadow.out.txt
9127b5631c6ed0b1 total_task_energy.out.txt
607e389debeb7203 windlass_gearing.out.txt
94fa9dc3d86a83e6 yielding_leg.out.txt
TURNING IN MID-AIR FOR NOTHING — reorientation at zero angular momentum.
A trunk with two limbs, each on its own joint axis. Angular momentum is held at exactly zero
throughout. What changes is orientation, which is not conserved. Predictions P1-P4 in the header.
GATES
[pass] G1/P1 PARALLEL axes, one full shape cycle: net turn 0.000e0 deg — must be exactly zero
A planar-momentum system is holonomic: orientation is a function of shape alone.
[pass] G2 angular momentum held at zero throughout: worst |L| = 1.11e-16
[pass] G3 the shape loop closes: joint error at cycle end = 1.47e-16 rad
[pass] G4 orientation stays a rotation: worst orthonormality drift = 2.60e-14
[pass] G5 SKEW axes do reorient: net turn 0.877 deg, which is 6e11x the integrator drift and
unboundedly larger than the parallel control's exact zero.
P2 — is it area, or is it distance travelled?
amp A amp B net turn (deg) vs amplitude^2 path length (rad)
0.30 0.30 0.251 1.00x 1.20
0.42 0.42 0.472 1.88x 1.68
0.60 0.60 0.877 3.50x 2.40
0.85 0.85 1.436 5.73x 3.40
Compare the measured ratio against amplitude squared. A phase that scales with ENCLOSED AREA
goes as amplitude^2; one that scaled with distance travelled would go as amplitude.
a loop with NO enclosed area net turn (deg)
both joints driven in phase 0.0000
A line segment retraced encloses no area, so a flux through it is nothing. Same joints, same
travel, no turn. That is the sharpest evidence that this is geometry and not effort.
P3/P4 — the bill, and the alternative.
Shape change, one cycle, 0.9 deg of turn:
net mechanical work over the closed loop : 0.0000 J (P3: the shape returns, so this is ~0)
copper, the actual bill : 3.5067 J
A REAL TURN: 90 deg within 0.50 s of flight.
shape change : 2.36 deg per cycle, so 38.1 cycles needed, 83.0213 J of copper each
total 3161.40 J, and it needs 38.1 cycles in 0.50 s — DOES NOT FIT: the flight is too short