Pillar: agent plane. Schema lands here; SDK types are generated from it.
WHY — The wire format for agent camera intent is the single most consequential design decision in the agent plane, and it is a land-grab: whoever publishes the neutral literal owns the interop point between any agent and any spatial engine. Published research finds that world-space 6-DoF trajectories transfer as "geometrically plausible but narratively meaningless" across scenes (Auteur, arXiv:2606.01900). An agent that sends {shot_scale: MS, orientation: FrontL, look_at: Chest, target: player:23} is portable across venues, sports, engines and vendors; an agent that sends a camera matrix is not — and it steps directly on the renderer's job, which is the fastest way to lose an integration. Enum-quantised axes are additionally the correct shape for constrained decoding: small enum-heavy schemas hold under grammar enforcement where large nested ones degrade.
GROUNDED — Auteur's six axes with published vocabularies and ranges: Camera Level (Ground/Low/Eye/High/Overhead, 0.1–4.0 m), Look-at Level (Feet/Knees/Waist/Chest/Eyes), Shot Scale (ECU…EWS, 0.10–1.00 of frame height), Framing Offset (FarLeft…FarRight, −0.35…+0.35), Orientation Angle (Front/FrontL/SideL/Back…, 0…±180°), Dutch Angle (None/Slight/Strong, −25°…+25°). Small open-weight vision-language-action policies commonly ship a 32-wide action and state vector (max_action_dim: 32, max_state_dim: 32 in a published reference config) — SE(3)=6 plus zoom/focus/iris plus an ROI rect fits with room to spare, so the literal is architecturally compatible with a 32-dimension action space without dictating one. WAVE side: nothing exists — zero viewport hits in source (grep, 2026-07-24), so this is honestly G0.
UNVERIFIED and to be re-checked before any of it is quoted externally: the sibling citations behind this design — arXiv 2606.01900 (Auteur), 2601.14945 (TIDAL, "400 ms blind window drops reactive success 0.31 to 0.09") and 2412.07889 ("many-track throughput peaks at N=5") — were not re-verified in the adversarial pass. Pin the figure or table reference, not just the arXiv ID.
BOUNDARY — The literal describes what shot is wanted, never how to fly the lens. No 6-DoF world-space poses, no keyframes, no interpolation curves, no camera matrices — those are trajectory, and trajectory belongs to whoever renders. No free-text prose field: it defeats constrained decoding and smuggles in ambiguity. Not a rendering format and not a splat format. Do not call it a standard until it is submitted somewhere — until then it is "WAVE's published neutral literal."
TAXONOMY — Production agents (emit), orchestration agents (compose), integration developers (bind SDKs), the technical director (reads the human-legible rendering), and any downstream policy model as the consumer on the far side.
GATE — G0 → G2 SCOPED on landing → G5 only when a downstream engine accepts it as the intent format.
DONE-WHEN — A dated receipt showing (a) at least 24 golden intent literals validate against the published JSON Schema and at least 8 adversarial ones fail closed with a typed error (world-space matrix, prose field, out-of-range dutch angle, unknown enum, missing target, oversized constraint set, negative TTL, unbounded hold); (b) a round-trip test packing every golden literal into a 32-dimension float vector and recovering it byte-identically; (c) the schema served at a stable public URL and resolvable by $ref.
DEPENDENCIES — None. Consumed by the viewport tool surface and the framing arbiter.
Pillar: agent plane. Schema lands here; SDK types are generated from it.
WHY — The wire format for agent camera intent is the single most consequential design decision in the agent plane, and it is a land-grab: whoever publishes the neutral literal owns the interop point between any agent and any spatial engine. Published research finds that world-space 6-DoF trajectories transfer as "geometrically plausible but narratively meaningless" across scenes (Auteur, arXiv:2606.01900). An agent that sends
{shot_scale: MS, orientation: FrontL, look_at: Chest, target: player:23}is portable across venues, sports, engines and vendors; an agent that sends a camera matrix is not — and it steps directly on the renderer's job, which is the fastest way to lose an integration. Enum-quantised axes are additionally the correct shape for constrained decoding: small enum-heavy schemas hold under grammar enforcement where large nested ones degrade.GROUNDED — Auteur's six axes with published vocabularies and ranges: Camera Level (Ground/Low/Eye/High/Overhead, 0.1–4.0 m), Look-at Level (Feet/Knees/Waist/Chest/Eyes), Shot Scale (ECU…EWS, 0.10–1.00 of frame height), Framing Offset (FarLeft…FarRight, −0.35…+0.35), Orientation Angle (Front/FrontL/SideL/Back…, 0…±180°), Dutch Angle (None/Slight/Strong, −25°…+25°). Small open-weight vision-language-action policies commonly ship a 32-wide action and state vector (
max_action_dim: 32,max_state_dim: 32in a published reference config) — SE(3)=6 plus zoom/focus/iris plus an ROI rect fits with room to spare, so the literal is architecturally compatible with a 32-dimension action space without dictating one. WAVE side: nothing exists — zeroviewporthits in source (grep, 2026-07-24), so this is honestly G0.UNVERIFIED and to be re-checked before any of it is quoted externally: the sibling citations behind this design — arXiv 2606.01900 (Auteur), 2601.14945 (TIDAL, "400 ms blind window drops reactive success 0.31 to 0.09") and 2412.07889 ("many-track throughput peaks at N=5") — were not re-verified in the adversarial pass. Pin the figure or table reference, not just the arXiv ID.
BOUNDARY — The literal describes what shot is wanted, never how to fly the lens. No 6-DoF world-space poses, no keyframes, no interpolation curves, no camera matrices — those are trajectory, and trajectory belongs to whoever renders. No free-text prose field: it defeats constrained decoding and smuggles in ambiguity. Not a rendering format and not a splat format. Do not call it a standard until it is submitted somewhere — until then it is "WAVE's published neutral literal."
TAXONOMY — Production agents (emit), orchestration agents (compose), integration developers (bind SDKs), the technical director (reads the human-legible rendering), and any downstream policy model as the consumer on the far side.
GATE — G0 → G2 SCOPED on landing → G5 only when a downstream engine accepts it as the intent format.
DONE-WHEN — A dated receipt showing (a) at least 24 golden intent literals validate against the published JSON Schema and at least 8 adversarial ones fail closed with a typed error (world-space matrix, prose field, out-of-range dutch angle, unknown enum, missing target, oversized constraint set, negative TTL, unbounded hold); (b) a round-trip test packing every golden literal into a 32-dimension float vector and recovering it byte-identically; (c) the schema served at a stable public URL and resolvable by
$ref.DEPENDENCIES — None. Consumed by the viewport tool surface and the framing arbiter.