VisionAssist is the GPT-S core perception capability: a visual-semantic cognition layer for evidence-grounded observations, competing intent hypotheses, counterevidence, uncertainty, invalidation, and explicit human-AI revision.
The R29 lineage preserves four stages:
Assistive vision
-> chart semantic pivot
-> Active Inference research architecture
-> current visual-semantic cognition canon
Accessibility is preserved as the optional android-pilot adapter. It is the
historical origin, not the current module identity. The machine-readable canon
is VisionAssist Lineage R29.
VISIONASSIST_CHART_INTENT_AND_HUMAN_AI_FUSION_PROOF
60 frozen market chart cases + 15 non-market visual cases
-> human-only prior
-> AI-only intent assessment
-> human+AI revision
-> outcome reveal
-> blinded scoring
The offline harness, phase gates, scorer, and proof checklist are implemented
under benchmarks/chart-intent-r26. The complete 75-case corpus is frozen at
CASE_FROZEN: 60 market cases plus 15 deterministic synthetic visual-control
cases under separate pre-registered commitments. Human, AI, fusion, baseline,
reveal, and adjudication records are not yet frozen. No calibration,
human-AI uplift, natural-scene generality, or production-value claim exists.
R33 handoff intake is bound to the exact R29 handoff. A local custody snapshot
now commits the ignored cases/ and outcome-vault/ trees without publishing
their inventory. Git identity, custody transfer, independent custody, and
human operator identity remain unbound.
image + human prior -> observations -> competing hypotheses
-> counterevidence -> invalidation
-> NO_ACTION / HOLD / DENY
Start the bounded runtime:
powershell -ExecutionPolicy Bypass -File .\scripts\start-intent-r26.ps1Then open:
http://127.0.0.1:8787/intent-r26
See the Intent R26 runtime guide.
camera capture -> /v1/describe -> TTS summary
-> hazard warning -> low-confidence warning
The adapter remains bounded and parked behind its own device-evidence gate.
Enables the prepared Describe, Intent R26, OCR, and Realtime development routes together. It is not evidence that every capability is pilot-ready.
.
|-- apps/mobile/ Android assistive client
|-- apps/intent-console/ R26 visual diagnostic console
|-- benchmarks/ P1 blinded proof harness
|-- contracts/lineage-r29/ canonical identity and portfolio contract
|-- contracts/intent-r26/ canonical schema, example, provenance
|-- services/edge/ profile-gated Node gateway
|-- docs/pilot/ Android proof pack
|-- docs/research/ benchmark and source-proof materials
`-- scripts/ startup/bootstrap helpers
Canonical boundaries:
powershell -ExecutionPolicy Bypass -File .\scripts\verify-p1.ps1The Intent module is implemented and offline-testable. A live model-backed image run still requires OPENAI_API_KEY.