When the Forecast Becomes a Meal Decision: The Provenance Record MBG Needs

MBG Watch · 2026-09-09

The premise

A weather forecast becomes more than a forecast when it changes a meal decision.

For MBG, that change can be small and practical: a route leaves thirty minutes earlier; a school handoff is moved under cover; a hot menu is replaced with a safer item; a batch is held, chilled, or discarded; a kitchen pauses delivery while a district is under emergency school closure. It can also be larger: after the Anak Krakatau ash disruption, BGN temporarily adjusted MBG distribution for schools under distance learning, saying operations should remain safe, targeted, and aligned with emergency policy, and that future operating steps would follow local conditions, local-government policy, school readiness, and SPPG readiness.

That is the right class of problem. Weather is not only a background condition. It enters the food-safety clock, the worker-exposure record, the route record, and the school-day record.

MBG Watch has already argued for this from several angles: “Weather at the Dispatch Door” asked for a stop-go record; “When Climate Stops Being an Exception” asked for a baseline stress test; “When the Last Mile Goes Quiet” asked for communications readiness; “When Guidance Runs Locally” separated local assistance from local authority; “When the Score Becomes a Gate” warned against unaudited scores becoming gates; and “Not the Dashboard, the Guidance” asked whether an interface changes the action at the edge.

This piece narrows the point. If AI weather models, BMKG public forecasts, route-weather tools, or locally derived guidance begin to affect MBG kitchen decisions, BGN should not publish or rely on a black-box “weather risk score.” It should publish a small forecast-decision provenance record whenever forecast guidance materially changes cooking, routing, menu substitution, handoff, worker exposure, cold-chain protection, or pause/restart decisions.

What better forecasts can and cannot do

The public forecast landscape is moving toward shorter update cycles and more usable local products.

Google DeepMind describes WeatherNext 3 as a global AI weather model that generates forecasts every hour, draws directly from raw satellite imagery, and can produce higher-resolution forecasts, including weather-station-targeted temperature and humidity at 5 km resolution and other surface variables such as wind at 10 km. It also says the model can deliver local data into Google products such as Search, Maps, and Gemini.

That matters because MBG decisions happen in short windows. A kitchen does not need a climate essay at 6 a.m.; it needs to know whether the route to a school will be unsafe, delayed, flooded, ash-affected, smoky, or too hot for a planned holding time. More frequent forecasts and route-aware interfaces could help.

But the record does not support treating any AI weather model as direct MBG authority. Google’s public description is not an MBG operating protocol. It does not say that its forecasts are validated for Indonesian school-meal logistics, SPPG cold-chain decisions, food-holding safety, last-mile routes, or kitchen stop-go thresholds. It also does not replace Indonesia’s public meteorological authority.

BMKG already has public products closer to the operating edge. Its open forecast data page describes village and ward-level forecasts for all Indonesian kelurahan and desa, covering three days, with eight forecast entries per day, updated twice daily. The exposed fields include UTC and local forecast time, temperature, humidity, weather description, wind speed and direction, cloud cover, visibility, and the analysis date — the production time of the forecast.

BMKG’s open early-warning page is even closer to operational accountability. It describes nowcast weather warnings across Indonesia down to the kecamatan level, based on the Common Alerting Protocol, updated at any time, with RSS and CAP XML feeds. The CAP fields BMKG documents include event, effective time, expiry time, sender name, headline, description, web link, and affected area polygon. BMKG also requires applications using the data to cite BMKG as the source.

For routes, BMKG’s Digital Weather for Traffic was described by ANTARA in 2025 as a real-time digital channel for travel-weather information, with features for rain warnings, land-route weather, travel-route weather, district-point weather, railway-route weather, aviation weather, and maritime weather. BMKG officials said the route feature could show where rain was occurring along a route, using radar data.

The practical conclusion is sober. Better forecasts can improve MBG decisions when they arrive in time, at the right location, with the right uncertainty and operating threshold. They cannot by themselves answer whether food should be cooked, moved, held, substituted, or stopped. That is still a public-health operations decision.

The provenance gap when forecast becomes instruction

The gap appears at the exact moment a forecast turns into an instruction.

A kitchen sees “heavy rain likely,” “route unsafe,” “ash affected,” “high heat,” “haze,” “flood warning,” or a color-coded weather risk. Someone changes the plan. The delivery is delayed. A batch stays in a vehicle longer. A menu is replaced. A school handoff is moved. A driver is told to wait. A SPPG is paused and later restarted.

Without a provenance record, no one can later tell what happened. Was the input from BMKG, a district official, a Google-powered interface, a local route app, a manual phone call, a kitchen supervisor, or a derived score inside an MBG dashboard? What time was it issued? What location did it cover? Which threshold was crossed? Who made the decision? Was the decision communicated to the driver, school, family, health post, or local government in time? What was later observed on the route? Was the forecast a false alarm, a miss, or a useful early signal?

This is where weather accountability and AI accountability meet. NIST’s AI Risk Management Framework is voluntary, but its framing is useful here: AI risks need to be governed, mapped, measured, and managed across design, use, and evaluation. OASIS’s Common Alerting Protocol is useful for a different reason: it shows that serious alerting already carries structured fields for identity, source, sent time, effective time, expiry, area, severity, certainty, and related messages.

MBG does not need to become a meteorological agency. It does need to keep the decision trail when forecast-derived guidance affects meals.

The MBG forecast-decision record

The record should be small enough for a kitchen to use and precise enough for an auditor to test. For every material forecast-driven operating change, BGN should retain and publish a summarized record with these fields:

• Forecast source: BMKG product, local government instruction, route-weather tool, AI model/interface, manual field report, or combined source.
• Source provenance: URL, feed, bulletin ID, CAP reference, model name/version where available, issuing agency, and required source attribution.
• Issue and validity time: when the forecast or warning was produced, when it was received by the kitchen, and the window it covered.
• Location and route: SPPG, school, district, route segment, kecamatan, village/ward, or affected area polygon where available.
• Operating threshold: rain intensity, flood/ash/haze warning, temperature/humidity/visibility threshold, road closure, school closure, or food-holding-time risk.
• Decision owner: the named role, not necessarily the person’s public identity — kitchen head, district coordinator, BGN command post, school authority, local disaster agency, or health official.
• Decision taken: route delay, menu substitution, cold-chain protection, worker exposure limit, handoff change, batch hold/discard, suspension, restart, or no change.
• Message trail: what was sent, to whom, by which channel, and when it was acknowledged.
• Observed condition: what drivers, schools, local officials, sensors, or public reports later confirmed.
• Performance review: whether the action was a hit, false alarm, miss, late warning, unclear instruction, or communication failure.
• Correction trail: what threshold, source, route rule, contact list, or restart condition changed after review.
• Privacy boundary: no child-level location trails, no family exposure, no unnecessary worker surveillance, and no public disclosure that turns a delivery route into a security risk.

The form can be shorter than the list. The discipline matters more than the dashboard. A forecast that changes a meal should leave a record showing why the change was made and what happened afterward.

The least-harm path

The least-harm path is not blanket shutdown. It is selective, inspectable adjustment.

A forecast should have authority only through a pre-declared operating rule. “Heavy rain in the district” is not enough. “BMKG CAP warning covering this kecamatan during the delivery window, combined with a route segment that has flooded twice this season, triggers supervisor review and school confirmation before departure” is closer to an accountable rule.

BGN’s own 2027 statement says the program is strengthening SPPG standardization, food-handler capacity, supply-chain resilience, quality and food-safety monitoring, information systems, supervision, governance, and wider ecosystem capacity for a program targeted at 72,464,886 beneficiaries. That is the right institutional home for this record. Forecast provenance should not sit as an isolated climate feature. It belongs in the operating record that already governs food safety, logistics, supervision, and restart decisions.

The backward test from last-mile warnings applies here too. A forecast has not succeeded because it exists centrally. It succeeds only when the kitchen, driver, school, family, health post, or local government receives a usable decision in time. If a forecast changes MBG operations but the message does not reach the people who must act, the failure is not meteorological. It is operational.

The minimum rule should be this: no opaque weather risk score may stop, restart, reroute, substitute, hold, or discard meals unless the underlying source, threshold, human decision, message, observed condition, and correction trail can be inspected.

What remains uncertain

I did not find public evidence that BGN is currently using Google WeatherNext, an AI weather model, or an internal weather risk score for MBG operating decisions. The argument here is therefore preventive: build the provenance record before forecast-derived tools become operational authority.

I also did not find a public MBG-specific protocol linking BMKG village forecasts, BMKG nowcast CAP warnings, Digital Weather for Traffic, local disaster decisions, and SPPG stop-go rules into one operating standard. Such a protocol may exist internally. If it does, the public question is narrower: which parts can be disclosed without exposing children, routes, or staff?

The hard boundary is between useful weather awareness and automated overreach. MBG needs better foresight. It also needs meals to keep moving when movement is safe. A narrow provenance record protects both aims: it makes weather guidance usable, and it keeps final authority where it belongs — in accountable public-health operations.