Bind immutable Speech context and approved household language

ALYT supplies the original Speech context and consumes complete semantics. The developer workflow is to create a full-home environmental context, call the internal speech-turn-context route with the current context ID and source revision, and then manage bounded household aliases through the metadata route. The model receives opaque room, Thing, and routine references, not native IDs or credentials. Renames, moves, deleted objects, capability changes, approved hints, and permission changes invalidate the original context.

Context binding and compiler limits

Create an original full-home environmental context, then call the internal speech-turn-context route with the context ID or source revision, current hearing device or null, audio session ID, and SHADOW or ACTIVE execution mode. Supply the internal service key and current household claims. An enrolled microphone is a shared home device; it does not establish the speaker’s identity. Current authorization is checked independently.

The compiler is decoder-prefix-speech-small-v4-host-metadata, with an exact 256-token context and 128-token result. Context ID and source revision remain bound in host metadata. Descriptor selection is atomic per Thing, ordered by current-room, hearing, and pending relevance and counted with the installed tokenizer. Omission accounting is explicit. A referenced descriptor omitted from the prefix cannot authorize an action. The fake tokenizer identifies itself separately and proves interface mechanics. Installed Voice replies also carry the loaded manifest calibration revision, and a mismatch is refused even when the semantic result is otherwise valid. Measured log likelihood, margin, out-of-distribution score, and audio quality belong to the installed adapter under that frozen calibration; only explicitly identified fake adapters provide fixture assessments. Voice compile and lifecycle keep their separate 64KiB transport bound, and mono16kHz PCM keeps its separate 800,000-byte request allowance; neither changes the independent Reasoning frame budget. Preserve original context expiry and the existing voice/cloud route on refusal.

Metadata route and household aliases

Optional household-approved aliases use the existing Room, Device, Endpoint, and Routine rows. GET and PATCH on /properties/{propertyId}/speech-metadata/{kind}/{id}?tenantId={tenantId} require current home access; changes require property management. Send expectedRevision, aliases, purposes, and effects. Purposes and effects belong only to routines; no alias changes names, controls, saved actions, or permissions. Concurrent edits or conflicting aliases require a new read. Inventory revisions withhold stale hints after behavioral changes. Device acquisition remains independent of hints.

Schema, units, and stale-data rules

The pinned Speech context0.2, semantics0.1, and decoder-prefix-speech-small-v4-host-metadata schema and fixture sets are the authority for the complete JSON wire shapes. The internal custom JSON-schema validator uses these pins; generic OpenAPI custom-type output does not replace them. Pending component IDs are pcN; semantic units are PERCENT, CELSIUS, FAHRENHEIT, KELVIN, DECIBEL, LEVEL, and NONE. Optional direction and qualifier fields are absent when unused. Compiler compatibility runs against the original Python oracle and golden fixtures without weights.

Immutable turn maps expire within 60 seconds. Source, enrollment, or consent changes invalidate them rather than extending a stale context. Missing metadata preserves existing names and behavior. A stale hint is ignored by Voice, but it never changes object ownership or control authority. The read path returns the original revision and the compare-and-set value, and invalid or colliding normalized aliases are refused.

speech-context-contract

The frozen alyt.physical-ai-qualified-interface/1.0.0 release preserves Speech context0.2, complete semantics0.1, the pinned compiler, and the 256-context/128-result token limits under immutable canonical schema and fixture pins. The Voice adapter must negotiate the exact contractVersion/contextVersion, provide the exact multilingual tokenizer, grammar, and compiler, and maintain deterministic execution with original context expiry and permissions. Feasibility acknowledgment does not supply trained weights or calibrated scores, and the complete inference artifact, including outer FFN tensors, remains a separate adapter obligation. ALYT authority keeps current-source validation, while physical outcome evidence, not the context contract itself, is required for acceptance.