How does a question
become evidence?
Follow one question through the instrument. Each stage below names the conceptual step in plain language first, then the technical component that does the work. You can read the architecture without ever meeting a service diagram.
One question, six stations
-
01
See
The instrument looks at the imagery. It reads the question and the plate together, and decides what is being asked. Is this a temporal comparison? A grounding request? A SAR-only claim?
Component Vision-language specialist · produces a task label and a list of slots (temporal, spatial, modality).
-
02
Compare
When the question is about change, the temporal pair is co-registered before any pixel comparison runs. A residual above tolerance invalidates the comparison, so it is measured first, not last.
Component Change detector · co-registration residual is part of the evidence record; the comparison only proceeds if it is below the per-pixel tolerance.
-
03
Locate
The change proposal, or the grounding target, is resolved to a region of pixels. Coordinates, area and score are recorded. A proposal without a location is not evidence; it is a guess with a coordinate missing.
Component Grounding head · emits region rectangles, areas in m² and confidence scores.
-
04
Understand
The answer is written in the language of the question. A short sentence that names what changed, where, and over what period. The sentence is bounded by the ground sampling distance and the registration residual — those bounds are stated, not hidden.
Component Change-VQA · returns a label string and the bounds derived from the evidence record.
-
05
Multimodal
The same scene can be observed in optical or in SAR. The instrument treats them as the same question asked through different weather: the same answer, validated by a different sensor.
Component Optical / SAR stack · cross-validation of an answer between passive and active modalities.
-
06
Evidence
The answer is assembled last, from evidence that already exists. Every region has a source component, a registration residual, a confidence score, and a calibrated counterpart. That is the audit trail.
Component Evidence layer · writes region areas, scores, provenance and run identifiers; calibrates reported confidence against held-out accuracy.
The instrument, drawn
Specialist system
Deterministic policy controller
Chain-of-thought is never exposed. The trace above contains execution state only: stage, selected task, active component, component identity, timing, evidence, confidence, provenance and run identifiers. The routing decision is produced by a fixed rule table, so it can be audited without inspecting model internals.