Weather at the Dispatch Door: The Stop-Go Record MBG Kitchens Should Publish

MBG Watch · 2026-08-13

The premise

MBG Watch has already argued that the food-safety clock for Makan Bergizi Gratis cannot be read by calendar day alone. In “Hourly Heat, Not Daily Heat,” the relevant record was batch-level: when cooking started, when food was declared matang, how long it was held, when it left the kitchen, when it arrived, when it was served, and what happened when there was delay. In “Heat at the Kitchen Door,” heat was treated as an amplifier of known weaknesses, not a stand-alone explanation. In “The Power Behind the Plate,” the same logic was extended to 3T energy readiness. In “The Visibility Standard,” the central test was whether BGN’s controls could be inspected from the outside before any operational pivot scaled.

The next layer is weather and route condition. Not weather as a reason given after children fall ill. Weather as a pre-dispatch control.

BGN has recently moved in the right direction on parts of the time record. On 8 August 2026, it said 27,489 operating SPPGs were required to use a production monitoring system that records stages from food preparation through cooking, maturity, packaging, delivery, and service at school. The same release says the school PIC must check meals before distribution, including an organoleptic check by sight, smell, and taste, and withhold food that does not meet the standard. BGN has also announced labels on each ompreng stating the latest safe consumption time, with MBG food generally to be eaten within four hours after cooking because it is fresh and not preserved.

Those are important controls. But they are still incomplete if the dispatch decision treats a wet, humid, power-unstable, slow route the same as a dry, short, reliable one.

The public question is not whether “AI weather” is exciting. It is simpler: before a kitchen dispatches perishable meals, what local weather and route facts did it use, what threshold did it apply, and what did it do when the threshold was crossed?

What local prediction could add

BMKG already publishes open weather-forecast data for Indonesian villages and urban wards, covering three days ahead, eight forecast entries per day at three-hour intervals, with fields including temperature, humidity, weather condition, wind, cloud cover, and visibility. BMKG requires attribution and sets an access limit of 60 requests per minute per IP. That is not a perfect route-level operational system. But it means MBG does not have to start from no public weather layer.

There is also a technical signal worth noting carefully. A 2026 arXiv paper from University of Cambridge researchers reports that Earth-observation embeddings can act as sub-grid surface descriptors for probabilistic weather downscaling. In their experiment, a downscaling model augmented with compressed 10 m TESSERA embeddings improved probabilistic skill for 2 m temperature and 10 m wind speed across five regions, including 11.5% CRPS skill improvement for temperature and 6.2% for wind speed. The useful point is not that this paper certifies any MBG system. It does not. The useful point is that local prediction is becoming more grounded in persistent surface features below the coarse weather grid.

For MBG operations, that should translate into a modest control layer:

• route forecast window: the forecast temperature, humidity, rain, wind, and visibility expected during cooking, holding, transport, arrival, and serving;
• route disruption flag: heavy rain, flooding, road closure, ferry disruption, landslide warning, or other delay condition affecting delivery time;
• energy-readiness flag: whether refrigeration, hot holding, water pumping, and packaging equipment had reliable power during the batch window;
• menu risk category: whether the meal contains higher-risk perishable components, wet sauces, coconut milk, egg, fish, chicken, or items intended to be held hot;
• time-temperature margin: how much of the safe window remains when the food leaves the kitchen and when it reaches the school.

The stop-go logic should be visible. For example: if the meal is high-risk, the route is expected to be delayed, the temperature is high, and the kitchen cannot prove hot holding or refrigeration, the default should not be “dispatch and hope.” The default should be a documented substitute, smaller batch, colder-safe menu, nearer school routing, delayed cooking time, or temporary non-dispatch.

This is not a blanket pause of MBG. It is a reversible, route-specific safety control.

What weather intelligence cannot add

Weather data cannot prove the cause of a poisoning incident after the fact. A hot day does not identify a pathogen. Rain does not replace laboratory testing. Humidity does not certify whether water, hands, surfaces, ingredients, or utensils were safe. A forecast model cannot make an uncertified kitchen safe.

This matters because model scores can easily become a new place for accountability to hide. If BGN were to publish only a “weather risk score,” the public would still be unable to inspect the kitchen decision. The score could be wrong, stale, over-smoothed, or quietly overridden. It could also be used to blame the weather for failures that actually came from sanitation, procurement, cooking, holding, or route management.

The WHO’s food-safety advice is deliberately practical: hot and humid weather helps germs multiply quickly; cooked food should not be left at room temperature for more than two hours; cooked food should be kept piping hot above 60°C before serving; cooked and perishable food should be refrigerated promptly, preferably below 5°C. Those rules are not replaced by a better forecast. They become more important when the forecast shows conditions that narrow the safety margin.

BGN’s own recent SLHS statement points in the same direction. On 7 August 2026, BGN said SLHS is not an administrative formality but a mandatory operating requirement for SPPGs. It said operating kitchens without SLHS were given until 10 August to complete certification, and that kitchens found not hygienically and sanitarily fit should be permanently suspended. Weather controls should sit underneath that standard, not beside it. A route forecast can tell a certified kitchen when a perishable delivery is riskier today. It cannot grant permission to an uncertified kitchen.

The minimum public control record

If weather intelligence is used, the minimum public record should be plain enough for a parent, school, journalist, local official, or auditor to read.

For each dispatch batch, BGN should publish or retain in an inspectable public-control log:

1. Kitchen and school route identifiers.
2. Menu risk category and any perishable high-risk components.
3. Cooking start time and matang time.
4. Holding method and temperature readings before dispatch.
5. Packaging time and dispatch time.
6. Forecast window used for dispatch: temperature, humidity, rain/weather condition, wind, visibility, and severe-weather warning status.
7. Power and water readiness status during the batch window.
8. Planned route time and actual arrival time.
9. School receipt time, PIC check result, and serving time.
10. Latest safe consumption time shown on the ompreng label.
11. Exception reason if the route was delayed, power failed, temperature logs were missing, or weather threshold was crossed.
12. Stop, substitute, or release decision, with the named threshold used.
13. Post-incident correction trail if any illness, complaint, rejected batch, or delayed service occurs.

This record should be batch-level, not just kitchen-level. It should preserve raw times and temperatures, not only compliance summaries. It should show the forecast used before dispatch and the actual conditions encountered afterward. And it should record overrides. If a supervisor releases a batch after a threshold is crossed, the public-control file should say who approved it, on what basis, and what compensating control was used.

The same record would strengthen BGN’s announced production monitoring system. If the system already records preparation, cooking, packaging, delivery, and service, then adding weather and route fields is not a new philosophy. It is a missing accountability column.

A least-harm path

The least-harm path is not to suspend a national meal program whenever the weather is bad. Children still need meals. Public money still needs disciplined use. Kitchens still need rules that can operate at scale.

A proportionate approach would start with the routes where the margin is smallest:

• long transport routes;
• schools without reliable arrival-to-serving control;
• high-risk menus;
• kitchens with recent incidents, missing logs, delayed SLHS, or weak temperature records;
• 3T routes with known power, road, water, or connectivity fragility;
• days with high heat, high humidity, heavy rain, flooding, or visibility disruption during the dispatch window.

For those routes, BGN should define pre-dispatch thresholds that trigger reversible actions: cook later, split the batch, change the menu, shorten the route, require insulated transport, move to shelf-stable components, add school-side temperature verification, or hold dispatch until a route delay clears. The control is not punishment. It is a way to avoid converting a forecastable condition into a child’s illness.

The governance test is whether the rule is auditable. A local forecast can be wrong. A route can clear faster than expected. A kitchen can perform well under difficult conditions. The system should allow correction. But correction should leave a record.

What I am uncertain about

First, I cannot verify from the public record that BGN has adopted weather-sensitive dispatch thresholds. The recent BGN materials I retrieved discuss production monitoring, school PIC checks, safe-consumption labels, and SLHS enforcement. They do not show a public stop-go rule tied to heat, humidity, rain, power reliability, or route disruption.

Second, BMKG’s open forecast data is useful but not the same as verified route-level exposure. It provides village/ward forecasts in three-hour intervals, not a guaranteed measurement of the temperature inside a transport container, school holding area, or congested road segment. MBG would still need temperature logs and field verification.

Third, the downscaling research is promising but early for this use. The paper supports the idea that surface-aware downscaling can improve local weather prediction skill; it does not validate an Indonesian MBG dispatch model, a kitchen risk score, or a foodborne-illness predictor.

Fourth, compliance incentives matter. If kitchens are punished only for reporting exceptions, records will become cleaner while food may not become safer. The safer design is to reward timely substitution and non-dispatch when thresholds are crossed, and to treat an honest stop as evidence of control rather than failure.

The core standard is therefore modest: do not let weather become either an excuse after harm or a black-box score before dispatch. Make it a public control record. Show the forecast, show the threshold, show the batch clock, show the temperatures, show the route exception, and show what was changed before children ate.