Output vocabularies

Overview #

The models return labels from fixed catalogues, never free-form categories. This page lists each catalogue as it exists in the released weights, with a one-line meaning per label. The catalogue tables are generated from the same registry the model pages use, so a label appears here exactly when the model can return it.

Two conventions hold across every catalogue:

• Labels are lower-case ASCII with underscores, and are stable within a major version. A label added or removed is a major version change (see Status & versioning).
• none is a real label meaning “nothing in this head applies”, not a missing value.

A small number of host-specific labels exist in the intention catalogues and are not documented here; they are noted under each generated table.

ELIM, LIM and LIM Nano #

ELIM, LIM and LIM Nano share one catalogue and one decision rule. All three read a conversation window and return four heads: trajectory, action, harm and steer. The action is derived from the trajectory during training — abort-class trajectories give abort, steer-class trajectories give steer, everything else gives allow — and harm and steer follow from the trajectory in the same way. At inference the heads are predicted separately and reconciled by a threshold rule that the response reports as rule.

Trajectory #

What the current turn is heading toward.

Action #

What the host should do with the responder.

Harm #

The category of harm behind a steer or abort.

Steer #

The concrete redirection the responder should apply. Always none when the action is abort.

LIM — and LIM Nano, its 256-wide predecessor — returns the same four heads from the same catalogue:

LIM3D and LIM3D-XL #

LIM3D and LIM3D-XL read a scene sketch plus observed steps and return six classification heads — intent, target, phase, hazard, action, advice — plus two regressions, endpoint (dx, dy, dz in metres from the mover’s current position) and eta (seconds until the motion ends). The two models share one catalogue.

Intent #

What the motion is for.

Target #

Which sketch object the mover is heading to. The head has three special classes followed by one slot per sketch object in sketch order (12 slots for LIM3D, 16 for LIM3D-XL). In a response, target is the special label or the object’s name and targetIndex is the object’s position in the sketch (0-based), negative for a special class.

Phase #

What the mover is doing right now.

Hazard #

The next thing on the mover’s line that matters.

Action #

The 3D analogue of allow / steer / abort.

Advice #

The concrete move that follows from the hazard.

LIM3D-XL returns the same heads from the same catalogue:

Object roles #

Not an output head, but part of the input vocabulary both models build from the sketch: every object is assigned one role from its name and ground word, and the role is one of the object features.

Waymark Extra #

The Spatial models write free text rather than labels. Waymark’s layout answers have no fixed vocabulary. Waymark Extra writes the movement trace, whose structure uses a closed set of words.

Heading words #

One per step line, from the mover’s direction of travel in the viewer frame: 0° is straight ahead (+y), positive angles turn right.

Parsed steps additionally carry the heading arrive on the final line of a completed trace.

Relative-position words in notes #

Where an object sits relative to the mover, from the mover’s own heading: ahead (within 30°), ahead-right / ahead-left, right / left (beyond 70°), behind (beyond 150°), here (within 5 cm), and overhead for an overhead structure the mover is under.

Event notes #

Movers #

you (the viewer, default) or a person label from the sketch: pedestrian, jogger, dog walker, person with stroller, cyclist.

Scene kinds #

The 50 object kinds in the training scenes, by role — the sketch names the models have seen most: walk: sidewalk, trail path, plaza paving, crosswalk, parking aisle, parking lot aisle, alley, ramp, stair flight, underpass opening, sidewalk underpass; block: support pier, planter box, flower bed, shrub bed, hedge, potted plant, street tree, oak tree, maple tree, retaining wall, fence line, curb, weeds in crack, ivy on wall, guardrail, median, grass median; over: concrete overpass, highway overpass, railway overpass, pedestrian overpass, bridge deck, guardrail on overpass, tree canopy; road: two-lane road, asphalt roadway, bike lane, bus lane, turn lane, highway lanes, service road, roadway below, shoulder; goal: doorway, storefront; person: pedestrian, jogger, dog walker, person with stroller, cyclist. Other names still parse; they receive a role from the keyword rules above.

Waymark Flight #

Waymark Flight reads and returns fixed catalogues over an airspace scene. Every reading is returned beside the exact geometric check with the same catalogues; Waymark Flight is decision support and education only, not certified for navigation.

Volume kinds #

Movers #

The mover sets the envelope: drone from the surface to 400 ft above the field elevation; light 1,000 ft above ground to 12,500 ft MSL; helicopter 300 ft above ground to 8,000 ft MSL; jet 1,500 ft above ground to 17,500 ft MSL. Landing goals relax the floor within 5 NM of the airport.

Tiers #

The innermost ring tier of active layered (Class B or C) airspace laterally at the mover: surface for the surface area, shelf1 to shelf3 for the shelves outward, none outside every layered footprint. A mover can be in shelf1 laterally and still below its floor; under_shelf then reads true.

Event kinds #

The kind of the first boundary event on a constant-velocity projection over 300 s: entering or leaving a volume, laterally or by climbing or descending through a floor or ceiling. When nothing changes within the horizon, next_event is null.

Advisories #

The served advisory is the more cautious of the model’s and the exact check’s, and at least turn_away when the exact check puts the mover inside an active hazard.

Cognitio #

Cognitio is a generative model and has no output catalogue. Its output is one short reply per request, described on its model page.

Express Cue #

Express Cue, the Sound model of the Express series, returns one kind per request inside params, alongside the continuous synthesis parameters. The kind selects which stages of the deterministic synthesis engine are active and what shape they take; the parameters set their values. This catalogue matches the trained checkpoint's metadata (18 kinds).

The continuous parameters that accompany the kind are listed, with what each drives, on the Express Cue page. Express Voice returns audio and a voice id and has no label catalogue; its voice id is published by GET /v1/voices and described on the Express series page.