Finger millet is a drought-tolerant C4 crop with a high light-saturation point and modest N demand relative to rice/wheat. Its PI is most sensitive to N/chlorophyll deficits in early tillering and to waterlogging at any stage.

Live staging status from qagro-research-in — what's measured vs. twin-borrowed for this geography, as of the last sync.

Overall PI confidence: 0.83 — five of six sub-indices measured, K twin-borrowed from the reference kernel's optimum midpoint pending Soil Health Card or wet-chemistry confirmation.

Each crop/geography pair gets a season-normalized PAR and day-length target rather than a fixed absolute value — the Twin compares live sunshine hours against the seasonal norm for that latitude band before flagging a light deficit.

Soil + air moisture probes trigger drip cycles; NPK is dosed into the same cycle in proportion to the live PI sub-index gap, not on a fixed calendar.

Doses scale up or down each cycle from these baselines using the live sub-index gap (e.g. N dose rises when NDRE falls below the stage optimum) and scale down automatically above 80% field capacity to prevent leaching and waterlogging.

Real-time decisions run on-device, not in the cloud. A small always-on sensor set drives the control loop; the full assay stack (drone NDRE, XRF, wet chemistry) stays a periodic calibration input.

Closed loop: minimal sensors → edge inference (on-device PI approx. + nutrient gap) → irrigate/dose decision → valve + proportional injector (NPK metered in-line) → flow + EC feedback confirms delivered volume & ppm → feeds back to edge inference, with periodic sync to the cloud twin registry for recalibration.

Same PI engine and edge architecture across all four — only the weight schedule and sensor-to-actuator mapping change per crop-geography twin.

Rice is a deliberate stress-test: it replaces moisture% with a flood-depth threshold, confirming the control loop is a generic resource-gap → proportional-actuation pattern, not hardcoded to drip irrigation.