Most labelled turns already bend in the last 2 seconds

Lab research, not a vehicle function. advisory_not_promoted. Measurement only — not a fit, not a map grant. Frozen cards stay 0.620 / 0.922.

Heading change in the 2 seconds we already have

Every window here is a labelled moving turn (≥ 3 m/s, future heading ≥ 10°). Δθ is the path heading change over the last 2 seconds, from the dots. Flat means under 5°. The old “already turning” bit was last 0.2 s at 2 rad/s — a much higher bar.

Windows with Δθ under 5°

Median Δθ over 2 s

Already turning at the gated-CTRA cutoff (2 rad/s)

Neighbors around the turner

Other tracks within 40 m of the labelled vehicle, in that vehicle’s forward frame. Heading is scored only if the neighbor is moving (≥ 1 m/s). Red cells are those already bending ≥ 5°. Parked jitter is not a turn cue.

Turns with a neighbor within 40 m

Of those with a 2 s neighbor trace, all of them still straight

At least one neighbor already bending

Neighbor at t₀ (count) Share already bending Origin is the labelled vehicle. +x is its last heading. Not TTC.

One scene at a time

Grey is the last 2 seconds. Black is where the labelled vehicle actually went. Blue is last-step constant velocity. Dots are neighbors; a red path means that neighbor’s own 2 seconds already bent.

Map-less ceiling

With only cuboid history, the number a learned model has to beat is last-step constant velocity. That is already the Gate A bar. It is not a new ADE table.

Frozen test ADE (Gate A bar)

Expanded moving-turning CV ADE

Extra-input net (spent grant)

Gated CTRA overall ADE 0.590 is not a second bar; it lost on turning. Catalog still has no lane prefix. This page does not license maps, camera, or another fit.

Expanded moving turns · Neighbors and extra seconds · History-only tree vs the simple rule. Not a brake, not a steer, not a command.