47. Decision trace and confidence¶

Every advisory is a graph you can walk backwards, from a farmer-facing recommendation to the raw signals beneath it. This is what makes GAP Agentic auditable rather than a black box, and it is the technical answer to two questions a national meteorological service will ask: "how do you know?" and "how sure are you?"

47.1 The trace is a DAG¶

Each advisory is a directed acyclic graph (DAG) of typed nodes. Any recommendation can be walked back through its ancestors to the inputs that produced it.

Node type	Purpose
Signal	A raw input as received (name, value, source, timestamp).
Feature	A quantity derived from one or more signals.
Assessment	One calculator's verdict, carrying its confidence.
Policy decision	A rule firing during resolution.
Recommendation	The farmer-facing output, with its ancestry.

flowchart LR
    S1[Signal<br/>rainfall forecast] --> F1[Feature<br/>rainfall accumulation]
    S2[Signal<br/>soil texture] --> F1
    F1 --> A1[Assessment<br/>verdict + confidence]
    A1 --> P1[PolicyDecision<br/>uncertainty gate passed]
    P1 --> R1[Recommendation]
    R1 -.->|"walk ancestry back"| S1

47.2 Deterministic node identity¶

Trace nodes are immutable, and a node's identity is derived from its own content and its parents. An identical decision made twice therefore produces identical node identities, so any divergence between two runs is meaningful rather than incidental. That property makes trace comparison usable as a regression check.

47.3 Multi-axis confidence¶

Rather than collapsing uncertainty into a single number, every assessment carries a confidence vector across several axes, banded as HIGH / MEDIUM / LOW for farmers. The axes capture distinct kinds of uncertainty, for example:

• meteorology — how trustworthy the forecast is, from model agreement and lead time;
• data quality — input completeness, lowered by missing fields and recorded degradations;
• spatial representativeness — how well a grid cell represents the specific farm;
• temporal relevance — how current the data is;
• rule coverage — whether the decision rules fit the case;
• cross-signal agreement — whether independent signals corroborate.

Confidence is re-weighted per action. A hazard alert leans on meteorology and temporal relevance; a planting recommendation leans on rule coverage and cross-signal agreement. The same vector can therefore yield different bands for different decisions — this is intended behaviour.

Confidence is not cosmetic: the data-quality axis feeds the uncertainty gates in the policy engine, which can withhold advice the platform cannot stand behind. See Observability and degradation for how a source failure propagates into confidence.