Out-of-Distribution Signaling to Operators

The single most important trust signal: telling the operator plainly when the machine has entered a regime the AI was never trained on.

When the model is out of its depth

An AI model is reliable inside the regime it was trained on and unreliable — often confidently so — outside it. For machines still in design and simulation and heading toward first-of-a-kind first tritium (~2030), operators will inevitably push into states no training set covered. The most important thing L6 can tell an operator is: the AI is now extrapolating beyond what it knows. Out-of-distribution (OOD) signaling makes that explicit rather than leaving the operator to discover it through a bad recommendation.

How OOD is detected and shown

Each model reports an OOD score — how far the current input is from its training distribution — using density estimates, ensemble disagreement, or reconstruction error. When the score crosses a threshold, the affected estimate's provenance badge flips its in-distribution flag, its confidence widens, and any recommendation built on it is explicitly de-weighted. On the 3D overlay, OOD regions are marked distinctly from merely low-confidence ones, because the operator's response differs: low confidence means noisy; OOD means untrusted.

def ood_flag(x, model):
    epi   = ensemble_disagreement(x, model)     # members diverge off-distribution
    dens  = model.density_estimate(x)           # low density -> unfamiliar input
    recon = model.reconstruction_error(x)       # autoencoder-style novelty
    score = combine(epi, 1.0 - dens, recon)
    return score > OOD_THRESHOLD                 # -> badge flips, confidence widens

The right operator response

OOD is a cue to shift weight from the AI toward first-principles physics and human judgment, and often to lower the automation level. The interface actively suggests this: when a safety-relevant loop's model goes OOD, the trust-and-override surface recommends more human involvement, and the copilot flags that its reasoning is now less grounded. This is the honest posture the whole stack takes — the AI advises within what it knows and says so clearly when it doesn't.

OOD signaling is the sharp edge of uncertainty visualization and the provenance badges, and it directly shapes trust and override decisions during novel operation.