VA.2026.012 · Floor 01 · hall 09
“before a mind can speak of the world, something must first look at it.”
VA-Sense is the eyes and ears of the collection — nothing more. Frames and clips enter, sense events leave: a segmented silhouette, a detected face with five landmarks, an identity label from a local gallery, an audio queue waiting for its provider. Summaries fold events into snapshots, and bridges carry text to memory and nodes to the graph. The kernel is framework-free by law; adapters for HTTP, models, and storage live outside it.
Models are data, never source: pulled by script, verified by hash, overridden by environment, absent from the repository. Segmentation and face run live on CPU; detection is registered but unwired; speech-to-text waits for its provider; text-to-speech is skipped permanently and says so with an honest error instead of a fake success. A single-file test page proves vision, snapshot-to-memory roundtrips, and audio stubs against the living service.
Identity here is orchestration, not biometrics: embeddings never leave the provider, the gallery holds local templates, and only a label — or nothing — travels in the event. Joining that label to a who belongs to the identity room, not this one. Every face in this hall is withheld — the gallery appears as count and dimension only — because the exhibit studies how perception flows, not whom it perceives.
VA.2026.012
VA-Sense
Perception Engine · Vision + Audio · ONNX CPU · TypeScript Library
Skeleton, blind kernel, model registry, segmentation and face live — the room opens its eyes.
Events fold into snapshots and sync live to memory as text and to the graph as anchored nodes — perception reaches the brain.
Object detection service plus realtime transport wait as plan documents.
Speech-to-text via provider; speech synthesis stays skipped on purpose.
Tracking, identity calibration, mask cropping, and budgets — the eyes learn discipline.
concepts — kept & discarded
kept — framework-free kernel with adapters outside; models pulled, never committed
rejected — fake-success stubs and biometric storage in the sense layer
held — detection service, audio provider, and tracking wait for their phases
archive ref — VA.2026.012/CON
cross ref — timeline fig. 01
curator annotations
weights are weather, not architecture — pull them, verify them, never frame them.
a detected face is not a recognized person — the gallery decides, the threshold doubts.
a permanent no is better documentation than a perpetual soon.
| materials | ONNX CPU models · model registry + sha verify · fps + confidence sampler · snapshot summaries · memory + graph bridges · honest 501 stubs | dimensions | phases — 2/5 green · tests — 156 + 23 e2e · modalities — segmentation · face · identity · audio · models — pulled, never committed · TTS — skipped — permanently |
|---|---|---|---|
| condition | dev · 156 pass · 23 e2e · models pulled, not committed | acquired | 2026 |
| revision | Rev. A | archive location | DWG-012 · floor Floor 01 |
| internal reference | VA.2026.012/M | origin | built because the voice needed eyes it did not have, 2026 |
failed experiments
associating before remembering answered node-not-found. bridges now carry their own anchors.
a mismatched embedding silently poisoned the gallery. wrong-sized vectors are now discarded, re-enrollment required on model change.
masks cover the whole frame instead of the box. cropping waits in phase 05 — documented, not hidden.
draft concepts
research notes
— kernel blindness scales down too — banning model imports from the core keeps tests instant.
— remember before associate, or the graph answers with absence.
— calibrate thresholds against your own gallery size — borrowed numbers flatter or accuse.
The room perceives and passes on; storage belongs to memory and the graph. A sense layer that hoards becomes a second brain with worse recall.
Every unbuilt modality answers with an explicit error, and one voice is refused forever. Honesty about absence is a feature visitors can rely on.
Biometrics never leave the provider and the gallery never leaves the room. What crosses the bridge is a string or a silence — enough to join, never enough to expose.
within the collection
future vision
VA-Sense grows eyes first, ears second: detection and streaming, then speech, then the discipline of tracking and budgets. The voice it refused stays refused — some silences are the design. Until then the room keeps looking, honestly reporting everything it sees and everything it cannot yet hear.