Anatomy of a Run
One question, followed through the instrument. The result appears first; the machinery is offered second. Every stage below corresponds to a real execution state the system reports, and every visual change on the plate is driven by that state, not by decoration.
This page renders a real captured execution, not a synthetic walkthrough. The run identifier, task, specialist models, evidence regions, confidence, timing and config hash all come from the server's own ResultEnvelope; the plate is the actual image that was analysed. Captured live from the deployed stack (Cloudflare Pages → Render → Codespace) on 2026-09-25.
What changed here?
The question
Everything downstream is a consequence of this string. It arrives with a scene reference — which plate, which epoch pair, which ground sampling distance — and nothing else.
Query processor
The question is parsed into a task and a set of slots. This is disambiguation, not reasoning: the output is a label the controller can act on, and it is shown to you verbatim.
Policy controller
A deterministic rule table selects the analysis path. There is no learned router and no hidden preference — the rules that fired are listed, and the ones that did not are listed too.
Specialist analysis
The selected specialists run, and only those. For this grounding request the image encoder proposes candidate regions and the grounding head scores them — no other specialist is invoked.
Grounding
Change proposals are resolved to located regions. A proposal without a location is not evidence; it is a guess with a coordinate missing.
Evidence
Each region is measured in metres, scored, and tied to the component that produced it. Evidence is what the answer is allowed to refer to.
Confidence
Reported confidence is temperature-scaled and checked against observed accuracy on a held-out calibration split. When calibration fails, the system is expected to abstain rather than sound certain.
Answer
The answer is assembled last, from evidence that already exists. It is bounded by what the imagery can support at this ground sampling distance — and the bound is stated, not hidden.
What is deliberately not shown
Never exposed
- Chain-of-thought or internal reasoning traces
- Intermediate prompts, unless the user asked for them
- Uncalibrated scores presented as confidence
- Evidence regions without a source component
Always exposed
- Current stage and active specialist
- Component and model identity, plus execution timing
- Registration residual before any change claim
- Region areas, scores, provenance and run identifiers