← Writing
Motion/Trajectory Geometry·6 min read

The Body Is Evidence

Why motion, speech, and agent work all end up needing replayable traces instead of confident stories.

The through-line in the motion work is not choreography for its own sake. It is evidence. A body moving through a room, a voice moving through an acoustic field, and an agent moving through a codebase all produce traces. Those traces are more reliable than a summary written after the fact because they preserve the structure of what actually happened.

This is why the anticipation geometry papers matter. The seven scalar signals are not decorations around a model. They are a way to make a trajectory inspectable. Commitment says how strongly a path has narrowed. Uncertainty says how many futures remain. Transition pressure says whether a system is collapsing toward a decision or opening back into possibility. Recovery margin asks whether the path can still return to coherence if the current move fails.

Claim

A system that cannot replay its own trajectory cannot honestly explain its own output.

That principle shows up in the speech work. A transcript is just a claim. The audio packet is the evidence. Once the recognizer collapses the waveform into characters, timing, tone, speaker energy, and many coarticulation cues have already been discarded. The acoustic verifier mattered because it reintroduced evidence where text-side governance was guessing.

It also shows up in the agent work. A successful final answer is not enough. The full session trace reveals whether the agent read the right files, checked its assumptions, ran tests, recovered from errors, or simply stumbled into something that looked correct. KARL exists because the trajectory of work carries more training signal than the final sentence alone.

The body makes this easiest to see. If someone says they are relaxed while their movement is stiff, asymmetrical, and arrhythmic, the body has introduced contradictory evidence. The right system should not erase that contradiction. It should preserve it, score it, and let the next decision account for it.

That is the larger architecture: not one model that speaks confidently, but a set of instruments that keep evidence alive long enough for governance to act on it. Speech, motion, graph traversal, and agent work become different instances of the same demand. Keep the trace. Score the trace. Learn only from what survives review.