Independent research · Physical intelligence under runtime assurance

Runtime assurance for frozen driving planners.

A guardian layer studies a planner's internal failure boundary, intervenes before impact in the tested setting, and keeps the claim boundary visible.

What the live surface shows

  • The useful moment is before impact.

    The surface renders a schematic trace from the committed campaign, not fake road video. One path shows the unguarded failure pressure, the other shows the guardian's measured intervention window.

    Evidence: NeuroNCAP released union improved the official closed-loop score while preserving a tight deployment-metric null.
  • The story is not every iteration.

    The page compresses the campaign into the real scientific spine: signal discovery, monitor release, localization, failed steering attempts, transfer boundary, replay support, and HUGSIM mechanism audit.

    Evidence: the milestones shown are the committed mechanism frontier; the head node is the latest gate in the published snapshot.
  • A stable benchmark result, not a general safety result.

    Sentinel can show the closed-loop NeuroNCAP gain because it is regenerated from committed evidence. The same screen shows the transfer null because that is the credibility signal.

    Evidence: OFF 2.12, released union 2.91, delta +0.783 with 95% CI +0.605 to +0.928.
  • The failure boundary is part of the product language.

    The honest page does not hide where the claim stops. It says the HUGSIM transfer gate was null, and it blocks deployment language until evidence exists.

    Evidence: HUGSIM transfer null, no deployment or safety claim, no retuning claim.