Compounding autoregressive drift
Latent errors accumulate over horizon T until the world model leaves the real state space: interpenetration, boundary breach, phantom contact.
METHOD
How verification
works.
Layered gates on measured behavior. Bounded claims, published uncertainty, missing evidence first.
SYNTHETIC REPLAY / SAME SCENE FAMILY AS THE FAILURE LAB
Same-frame fixture traces; draft boundaries.
01 / FAILURE MODESFive ways physics
Five ways physics
breaks the demo.
Every gate traces to a named failure mode with a measurement, never a vibe.
E[T] ~ O(T·ε)
The executability gap
Perceptual similarity (LPIPS, SSIM) says the motion looks right while commanded forces exceed torque, contact, or acceleration limits.
Perceptual hallucination
Out-of-distribution inputs collapse to the nearest latent codebook vector. The reconstruction residual trips the gate before control executes.
Action marginalization
The model plays back plausible motion while ignoring the action stream. Shuffled actions yielding identical predictions proves it.
Sim-to-real degradation
Positional divergence between simulated and physical trajectories in identical reference frames, measured as discrete Fréchet distance.
Match the model to the measurement
An architecture name does not establish causal understanding or physical validity. Record the model, task, observation and action contract.
Three schools
ARCHITECTURE / CAUSAL RUNGSCHOOLRUNG
RENDERER SCHOOLPixel-space video modelsVisual plausibility is not contact-dynamics evidence. Verify timing, action conditioning and physical constraints for the specific model and task.PIXELS
LATENT / JEPA SCHOOLLatent prediction and planningLatent rollouts can support planning. Inference speed, uncertainty and safety properties need model-specific measurements.LATENT STATE
SPATIAL / ENGINEERED SCHOOLVoxels + PINN constraintsMetric geometry with conservation-law penalties. The verifiable structure behind industrial digital twins and audits.RUNG II
03 / SYNTHETIC GOVERNANCEAccumulate,
Accumulate,
then filter.
Recursive training on uncurated synthetic data collapses variance and erases the tail events where safety lives. The fix is a law, not a hope.
D(n+1) = α·D_real + (1−α)·{x ∈ G : Verifier(x)}REAL RATIO α PRESERVED / EVERY SYNTHETIC SAMPLE VERIFIEDReal 50%Synthetic 50%GOVERNED
This ratio illustration is not a prediction of model collapse. A retained real anchor, verifier quality and labeled validation remain necessary.
Run the provenance checks04 / ENGINE MAPSeven modules.
Seven modules.
One chain of custody.
Local-first and air-gap-compatible: one service, one data directory, complete backup by archiving the folder.
Module responsibilities
M1–M7 / TYPED COREMODULEOUTPUT
M1Degradation ScorerSim-to-real divergence, horizon drift, action sensitivity → scorecard JSON.CORE
M2Executability GuardAction-chunk physics validation → supervisory HALT within one chunk of latency.CORE
M3Bounded VerifierExact queries over sub-policies → PROVEN / VIOLATED / UNKNOWN certificates.PROOF
M4Adversarial EngineSeeded friction / mass / noise search → edge-case catalogue with replay refs.HUNT
M5Embodiment BenchmarkCross-robot transfer with frame checks → centered-Fréchet transfer score.BENCH
M6Provenance LedgerRun seals, dataset cards, reviews → append-only SHA-256 hash chain.CHAIN
M7Compliance MappingScorecards + ledger heads → regulation-mapped dossier packs + OTA vault.FILE
Reproducibility, with a recorded scope
Inputs, profile, code version and seed are pinned. Same-platform determinism is tested; verify a second platform with the supplied golden-check tool.
Canonical inputs
Schema validation, key sorting, unit conversion, and reference-frame pinning before anything scores.
Double precision
Measurements use unrounded floating-point values. Display rounding does not change gate decisions; cross-platform reproduction requires the golden check.
Seeded stochasticity
Every random operation takes an explicit, recorded seed. Same seed, same scorecard.
Timestamp-free hashes
Canonical bodies exclude wall-clock time, so independent re-runs on the same platform hash byte-identically.
Verify the method
on your hardest case.
ONE ROBOT / ONE FAILURE MODE