Edward Y. Chang, Stanford University Position statement, v1.0. August 9, 2026.
Cognitive AI does not require a physical world to interact with directly. Physical AI does. The dividing question, which I take to be the central engineering question of embodiment, is not prediction quality but synchronization: how, and when, should a physical world and its digital representation be brought into agreement, and who is accountable when they were not?
My answer, TRW (TRACE-WorldModel), treats the world model as a materialized view over a base that can only be sampled at a price. The world state is observable through priced channels; between observations the view advances under a motion model without reading the base; an observation conditions the view as a partial refresh. On this framing, three disciplines follow. First, a commitment read: an action that will bind for a horizon is authorized only against a declared precision class, with bounds on calibrated uncertainty, support, out-of-distribution score, and maximum evidence age. Freshness stops being a virtue and becomes a typed, checkable, refusable precondition. Second, adaptive refresh: because channels are priced, refreshing is a decision, routed by value of information, with the design goal of converting undetected staleness into detected, repairable invalidations. Third, effects as four objects rather than one boolean: intent, dispatch, outcome, and completion, so that a provider failure after dispatch is visible divergence rather than imagined success.
The bridge is universal in a precise sense: the same contract governs fully digital, simulated, and physical regimes, with consequence severity as an orthogonal axis. In the measured flood search-and-rescue campaign over real delta geography, the world-audit console holds no mission truth of its own, calibration reached a seed-grouped out-of-fold expected calibration error of 0.00477, and the gate behaved as designed: a passability claim reported at confidence 0.92 was deferred for thin support and out-of-distribution evidence, with the missing evidence item named. Cognitive AI and physical AI stop being different fields; they become the same audited loop with the boundary crossings priced and typed.
Full development: The Path to AGI, Vol. 3, Part II (world under audit, physical anchoring, effects, deliberation cost, closing the loop).