Expand description
§pointlock-vision
The VisionVerifier plugin interface (verify role only, principle 7:
vision never locates or acts) and the default StubVisionVerifier.
The verifier is the runtime consumer of the vision verify channel
(spine §6.3): it answers an author-written prompt against localized
screenshot bytes with a three-valued verdict. A pass/fail answer is
a completed evaluation (fail is final, spine R5); an unknown answer
means the channel could not complete and the chain advances — for the
vision channel, which is only legal at the chain tail, that exhausts the
chain into assertion unknown (principle 4).
The v0.1 default is the stub: it always answers unknown with the
reason "vision verifier not configured". The runner treats an absent
verifier (RunOptions::vision == None) as exactly equivalent.
Structs§
- Anthropic
Vision Verifier - The first usable verifier (08 §6.4): asks an Anthropic vision model to
answer the author’s prompt against the screenshot, over raw HTTP (no
official Rust SDK exists). Discipline unchanged from the trait docs:
verify-only, chain tail only, and every failure mode — missing key,
transport, non-200, unparseable answer, model uncertainty — folds to
unknownwith a reason, never an error and never a guessed pass (principles 4/7). - Open
AiCompat Vision Verifier - A verifier over the OpenAI-compatible chat-completions wire format —
how graphics-focused open models (Qwen-VL, GLM-4V, …) are typically
served, both hosted and self-hosted (vLLM). Same discipline as the
Anthropic verifier: verify-only, chain tail only, every failure mode
folds to
unknownwith a reason (principles 4/7), and the same look-then-judge answer protocol so verdicts stay comparable across providers. - Stub
Vision Verifier - The v0.1 default verifier: always
unknownwithSTUB_REASON. - Vision
Judge - Identity of a vision judge implementation (evidence honesty: once the model is swappable, a verdict must say which model answered).
- Vision
Request - One vision verification request: the author-written prompt (never
synthesized by the compiler, principle 6), an optional region of
interest, and the localized screenshot bytes (evidence is localized
during
observing, spine §6.6 — the verifier never reaches back into a provider session). - Vision
Verdict - The verifier’s three-valued answer with a human-readable reason.
Constants§
- DEFAULT_
VISION_ MODEL - The default model of
AnthropicVisionVerifier. - MAX_
OBSERVATIONS - Cap on self-reported observations kept per answer, and per-observation character bound — both bound ledger growth: observations enter the durable assertion record verbatim.
- MAX_
OBSERVATION_ CHARS - See
MAX_OBSERVATIONS. - STUB_
REASON - The reason the stub (and an unconfigured runner) yields
unknown.
Traits§
- Vision
Verifier - The vision verification plugin interface (verify role only).