> ## Documentation Index
> Fetch the complete documentation index at: https://docs.wiseyield.co/llms.txt
> Use this file to discover all available pages before exploring further.

# Trust badges

> Three confidence levels stamped on every recommendation — data-driven, data-assisted, ai-generated.

Every recommendation that WiseYield surfaces carries a **trust badge** describing how the recommendation was produced. The badge tells operators (and integrators) how much confidence to place in the underlying numbers and whether they should be acted on directly or treated as a starting point.

## The three badges

| Badge               | What backs the recommendation                                                                                                                                                                                 | When you should trust it for action                                                        |
| ------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
| **`data-driven`**   | Authoritative reference data (ISRIC SoilGrids, FAO-56 crop coefficients, Köppen-Geiger climate, lab-measured leaf/water/soil samples) plus deterministic engines (ET₀, leaching fraction, fertigation dosing) | Act directly. Numbers are computed from measured inputs and published reference data.      |
| **`data-assisted`** | Same authoritative inputs as `data-driven`, plus an LLM layer that interprets, explains, or contextualises the deterministic output                                                                           | Act with light review. The numbers are still computed; the LLM only narrates them.         |
| **`ai-generated`**  | LLM-only inference (no measured inputs or reference data backing the specific numbers)                                                                                                                        | Treat as a starting point. Cross-reference with measured data before committing resources. |

## Where badges appear

In API responses, the badge is stamped on each recommendation object:

```json theme={null}
{
  "id": "rec_...",
  "trustBadge": "data-driven",
  "type": "fertigation_dose",
  "recommendation": {
    "nutrients": { "N": 4.2, "P": 1.1, "K": 6.8 },
    "unit": "kg/ha"
  },
  "sources": [
    "Leaf analysis 2026-04-22 (lab_id: ...)",
    "FAO-56 Kc for date palm at kimri stage",
    "ISRIC SoilGrids — 0–30 cm horizon for farm coordinates"
  ]
}
```

The `sources` array is required for `data-driven` and `data-assisted` badges. `ai-generated` recommendations carry an empty or descriptive-only `sources` list and an explicit caveat.

## How the badge is decided

The platform decides the badge based on which pipeline produced the recommendation:

1. **Deterministic engine ran successfully** with all required inputs present → `data-driven`.
2. **Deterministic engine ran**, then an LLM was called to interpret or explain → `data-assisted`.
3. **No deterministic engine ran** (LLM-only orchestration, often because required inputs were missing) → `ai-generated`.

A recommendation is **never upgraded** from `ai-generated` to `data-driven` by the LLM itself — only the presence of measured inputs and a deterministic computation moves it up the ladder.

## What this means for integrators

* **Building a dashboard?** Surface the badge prominently. A `data-driven` fertigation plan can be queued for execution; an `ai-generated` one should require operator confirmation.
* **Triggering automated actions?** Gate automation on `data-driven` badges only. Do not auto-execute on `data-assisted` or `ai-generated` recommendations.
* **Showing recommendations to end-users?** Match the badge to the operator's level of authority — a Grove-tier farm manager can act on `data-assisted`, a Seed-tier hobbyist should see all three with clear visual differentiation.

## Why the distinction matters

WiseYield's deterministic engines (ET₀ via FAO-56, leaching fraction, fertigation dosing, yield prediction with Holt-Winters) are reproducible: same inputs produce the same outputs every time, and the math is independently verifiable. LLM outputs are not reproducible and can drift between model versions. Stamping the difference on every recommendation is what keeps the integrity of the `data-driven` label honest — the moment the platform claims `data-driven` for an LLM-only output, the badge stops meaning anything.

## See also

* [Land hierarchy](/concepts/land-hierarchy) — badges always scope to the level whose data backed the recommendation
* [Errors](/concepts/errors) — when required inputs are missing, the API returns `VALIDATION_ERROR` rather than downgrading silently