Datasetpublic

Sejibeji/causal-internalization-physical-law

IOO — Conservation-Slip Signatures on Hyper-Graph Fiber Bundles Verified results bundle for the working manuscript "Machine Intelligence for Physical Systems: Conservation-Slip Signatures on Hyper-Graph Fiber Bundles — From Simulated Aircraft to Satellite and Robot Telemetry" (Sehaj Randhir Singh, NYU ECE). This dataset hosts the machine-readable verification artifacts of the IOO (Index of Operators / Index of Operations) framework: one JSON per experiment, each produced and… See the full description on the dataset page: https://huggingface.co/datasets/Sejibeji/causal-internalization-physical-law.

sourceHugging Faceupdated 17d agoView on Hugging Face

0likes143downloads

Dataset Card

IOO — Conservation-Slip Signatures on Hyper-Graph Fiber Bundles

Verified results bundle for the working manuscript "Machine Intelligence for Physical Systems: Conservation-Slip Signatures on Hyper-Graph Fiber Bundles — From Simulated Aircraft to Satellite and Robot Telemetry" (Sehaj Randhir Singh, NYU ECE).

This dataset hosts the machine-readable verification artifacts of the IOO (Index of Operators / Index of Operations) framework: one JSON per experiment, each produced and verified end-to-end on Kaggle compute at the full protocol (except where noted). The framework introduces a conservation-slip signature: a physics-derived, unit-invariant statistic that measures the lagged asynchrony a physical boundary accumulates relative to its healthy coupling profile. Faults decouple boundaries; slip rises; the machine's graph manifold deforms. One representation — monitor, transfer, control, design.

Headline verified results (all reproduced end-to-end on Kaggle)

Experiment	Kernel	Result
Simulated aircraft, run-disjoint detection	`ioo-hypermechanism`	AUC 0.992 (MLP 0.482); zero-shot 0.933
Real ESA satellite (OPS-SAT-AD), official split	`ioo-reald-conservation-v2`	slip adds +0.040 AUROC; unsupervised 0.659
Satellite strong recalibration	same	slip −0.013 AUROC drop (univariate −0.432)
Real NASA fleet (N-CMAPSS DS02), zero-shot units	`ioo-hybrid-ncmapss`	RUL RMSE 20.11 vs retrained-MLP 33.90
Real hexapod robot, detection / localization	same reald v2 kernel	0.990 / macro-AUC 1.000
Fleet compiler representation	`ioo-shape-compiler-fleet`	shape curvature 5.6× more unit-invariant than raw channels
Shape classification boundary (honest)	same	supervised stats 0.985 vs shape 0.602 on level drift
Decisive baselines, third fleet (S2d)	same	under per-unit sensor recalibration: stats 0.985→0.675 (drop 0.310), shape measured drop ≤0.005, DANN 0.983→0.573 (drop 0.409)
Zero-shot MPC re-control (hardware swap)	`ioo-zero-shot-mpc-re-control-c1`	settle err 0.0001 (flat MLP 0.0237)
Slip-conditioned control (frozen model, 5× gain drop)	`ioo-slip-conditioned-mpc`	ISE 0.067 vs plain 0.442 (6.6×)
Same under noise + 5% dropouts	`ioo-slip-conditioned-mpc-noise`	ISE 8.2×, ss err 11.4× better
Generative physical design over healthy shape manifold	`ioo-shape-generative-design`	interpolations at 1.44× real spacing; 90.2% inside envelope
Fixed public benchmark: CruiseBench, all nine subdatasets (eta5-W256-S10)	`ioo-on-cruisebench-ds02-cpu-run` + `ioo-cruisebench-sweep-{a,b,c}`	IOO-CONV wins 6/9 subdatasets (DS02 1.83 vs GRU 2.02 prior best); benchmark mean 3.32 vs best-published 3.37 / TSMixer 3.46; losses reported (DS03/DS04/DS08c)
Wind-farm operating envelope (honest 4th-fleet boundary)	`ioo-windfarm-conservation` v3	no representation detects pre-fault at 10-min SCADA (stats 0.585, slip 0.497, DANN 0.539) — decoupling dynamics averaged out

Layout

—manuscript.pdf — current compiled working manuscript (22 pp).
—results/ — per-experiment verified JSONs (full protocol on Kaggle; each file records its own quick flag).
—Regeneration: kernels are rebuilt from ioo_kaggle/build_kernels.py; figures from ioo_paper/make_figs.py; manuscript via pdflatex ioo_paper/manuscript.tex. Kernels are private during review and will be made public on acceptance (access on request).

Why this is a candidate representation paradigm

1.Invariance by construction, not by learning. Per-channel affine recalibration cancels in the slip asynchrony ratio (Proposition 1 in the manuscript) — verified: univariate statistics collapse (−0.432 AUROC) under test-telemetry recalibration while slip does not degrade (−0.013); nonlinear sweeps degrade ≤ 0.007 AUROC.
2.Capability, not just scores. A frozen dynamics model re-normalizes commands through its own boundary readout and holds a plant through a 5× actuator-gain drop at healthy setpoint error — no retraining, no fault flag — and the same controller survives measurement noise and dropouts.
3.Honest boundaries. Calibrated statistics carry absolute level/magnitude (supervised level-drift classification 0.985 vs 0.602; RUL regression); the shape representation is blind to level drift by design and its home is the decoupling regime (hexapod localization, satellite drift-resistance, cross-fault transfer).
4.Generative. A VAE over healthy dev-unit shape signatures synthesizes designs that land near the healthy manifold and back-compute physically valid operating points.

Provenance

—Datasets: OPS-SAT-AD (Zenodo 12588359, CC-BY-4.0; ESA OPS-SAT operated by the European Space Agency), hexapod joints (Kaggle), N-CMAPSS DS02 (U.S. Government Works; Chao et al., Data 6, 5 (2021)).
—All experiments deterministic or seed-fixed; learner variance reported where it exists.