When Training Becomes Evidence: The Competence Record MBG Kitchens Need

MBG Watch · 2026-09-24

The premise

The useful lesson from recent AI tutoring research is not that MBG should rush to put chatbots in every kitchen. It is narrower and more important: learning systems are beginning to separate an answer that looks competent from a learner who can perform the task later, without help, under pressure.

StudentBench, released on 23 September 2026, reports that AI tutoring produced GRE learning gains statistically equivalent to expert human tutoring in a study of 2,383 participants and more than 175,000 student-AI messages. That is evidence about tutoring, not about food safety. But it sharpens the right question for MBG: if a situated helper can improve practice, what public record proves that a kitchen worker, route coordinator, school handoff staff member, cadre, or supervisor can actually perform the safety-critical work when the meal is moving?

MBG Watch has already argued for related layers: guided practice at the moment of action in “From SOP Posters to Guided Practice,” rehearsal records before disruption, workforce proof for 3B and 3T care, bounded edge assistance, and hard gates for controls that cannot be left to discretion. This piece adds one missing layer: a public competence and retraining record.

A certificate says a condition was met. A training attendance sheet says a person was exposed to material. A competence record says the system has observed the work, found the weak points, corrected them, and retested them.

What current evidence shows

BGN’s public record now contains several food-safety controls that matter.

First, BGN has made the Sertifikat Laik Higiene Sanitasi, or SLHS, a hard operating condition. In its 7 August 2026 statement, BGN said SLHS is not an administrative formality but a “syarat mutlak” for SPPG operation, and that a kitchen found unfit on hygiene and sanitation grounds cannot continue operating. The same statement said BGN had received reports of roughly 950 kitchens suspected of not meeting hygiene and sanitation standards and would verify them before deciding which SPPG should be permanently stopped.

Second, BGN’s September 2025 SLHS statement describes the content of that hygiene floor. It says every SPPG must pass sanitation inspection, and describes five food-safety keys: keep the environment and equipment clean, separate raw and cooked food, cook at safe temperatures, store food at proper temperatures, and use safe and verified water and raw materials. It also says local health offices and environmental agencies conduct inspections under BGN technical supervision, and that SLHS is valid for one year and re-evaluated annually.

Third, BGN has begun to use suspension and inspection as an operating response. On 7 September 2026, BGN said it temporarily closed 1,999 MBG kitchens that had not registered for SLHS, after checking 27,022 kitchens against health-ministry coordination data. On 10 September, BGN said it and the Pati district government would evaluate all 172 SPPG in Pati after a food-poisoning incident, including facility condition and whether kitchen capacity matched beneficiary load.

Fourth, BGN has started to treat the handoff to schools as a food-safety point. CNN Indonesia reported in January 2026 that BGN ordered SPPG to put safe-consumption time labels on MBG meals, asked SPPG heads to make written agreements with schools, and discouraged taking meals home. The same report quotes BGN’s public-communication and investigation deputy connecting many poisoning cases to meals eaten after the safe time had passed.

These are real controls. They move MBG away from a purely declarative safety posture.

But they still leave a gap. None of them tells the public whether the people doing the work can perform the sequence reliably: receiving ingredients, separating raw and cooked food, cooking to a safe point, holding hot or cold food properly, labeling the tray, rejecting a late handoff, discarding unsafe food, escalating a complaint, preserving samples, cooperating with investigation, and retraining after failure.

Why SOPs, certificates, and training claims are not enough

Food safety is not mainly a memory test. It is a timed, physical, interrupted sequence of work.

A 2026 Lancet Regional Health – Southeast Asia article on Indonesia’s free school meals program describes why this distinction matters. The authors report that by 31 December 2025 the program had caused 177 food-poisoning outbreaks affecting more than 20,000 schoolchildren across 127 districts in 33 provinces, while noting the true magnitude was likely higher. They also describe a demanding SPPG operating model: typically 3,000–4,000 meal portions daily for 15–25 schools, plus monthly contributions to 5–10 community health posts, often with overnight cooking.

The same article reports field findings that point to competence, not merely paperwork: food handlers preparing meals without proper protective equipment and handwashing procedures; high-risk ingredients stored outside safe temperature ranges; cooked meals left at room temperature for 7–8 hours, beyond the WHO 4-hour threshold cited by the authors; trays dried with unsterile cloth; inadequately rinsed fruit; surfaces wiped rather than sterilized; and sanitation failures severe enough to allow flies or maggots to contaminate meals.

This is the difference between a standard existing and a standard being performed.

HACCP guidance makes the same point in system language. The FDA’s HACCP principles include hazard analysis, critical control points, critical limits, monitoring, corrective actions, verification, and record-keeping. Its explanation is plain: if a deviation shows control has been lost, the deviation should be detected and timely steps taken to re-establish control so hazardous products do not reach the consumer. USDA’s school food-safety guidance points in the same direction: a school food-safety program based on HACCP principles depends on a written plan and records of monitoring, corrective action, and calibration.

That is the standard MBG should adapt to its own scale. The public should not be asked to accept “training has been done” as proof that the control chain works.

The public competence record MBG should publish

The least-harm version is not a worker surveillance layer. It should not expose individual cooks, drivers, cadres, or school staff to public naming and blame. MBG is operating under intense pressure; turning each mistake into a personal spectacle would push errors underground.

The record should be aggregate, role-based, and inspectable. For each SPPG, district, or region, BGN could publish a monthly table with these fields:

• role category: kitchen handler, cook, quality-control lead, route coordinator, school handoff lead, cadre, supervisor, incident-response lead;
• required module: hygiene, raw/cooked separation, cooking and holding temperature, label and consume-by handoff, route disruption, complaint intake, sample preservation, incident escalation, waste and IPAL duties;
• completion count and coverage: number trained, number active, percentage current;
• practice date: when the task was last observed or drilled, not just taught;
• scenario tested: ordinary service, late delivery, heat/haze/flood disruption, power outage, ingredient rejection, suspected contamination, vomiting cluster, 3T delivery delay, kitchen suspension restart;
• observed failure mode: late label, incomplete temperature log, unsafe holding time, missing handwashing step, overloaded route, unclear school responsibility, complaint not escalated, sample not preserved;
• corrective action ordered: change route timing, reduce production load, retrain handoff lead, repair equipment, add supervisor check, stop menu item, suspend kitchen, retest before restart;
• retraining and retest date: when the failed step was practiced again and whether it passed;
• accountable supervisor role: not the worker’s name, but the office or role responsible for closure;
• privacy boundary: no individual names, no medical details of children, no punitive public scoring of low-level workers.

This would do three things at once.

It would show whether SLHS and SOPs have reached the daily work. It would show whether a poisoning incident, suspension, or failed inspection caused retraining rather than just a press statement. And it would let communities see whether kitchens in harder conditions — 3T routes, disaster-mode feeding, heat, haze, flood, power interruption — have been tested against the conditions they actually face.

Where situated helpers may help, and where they cannot be proof

AI tutoring, guided practice, simulation, and local/offline assistants may be useful in MBG, especially where supervisors are thin and kitchens are under time pressure.

A safe use would be rehearsal: a kitchen lead practices the decision chain for a late school handoff; a route coordinator simulates a bridge closure; a cadre practices what to do when several children report nausea; a supervisor reviews whether the corrective action after a failed temperature log is specific enough. A local assistant might help workers remember the next step when connectivity is poor, translate a checklist into local language, or generate a practice scenario for a weekly drill.

But assistance is not competence.

An AI helper can prompt a worker through the right answer while leaving no proof that the worker can act unaided during an interruption. A checklist can make the task easier while hiding the fact that the route is overloaded. A simulation can produce good answers while failing to test whether thermometers, labels, containers, cold storage, and supervisor authority are actually present.

So MBG should keep three categories separate:

• assisted performance: what a worker can do with a prompt, checklist, tutor, supervisor, or device;
• certification: what a kitchen or person has been formally cleared to do;
• durable competence: what the system has observed the role performing under ordinary and disrupted conditions, with correction and retest after failure.

Only the third should count as evidence that training has become operational safety.

What changes after harm or disruption

After a food-poisoning incident, a generic “retraining conducted” line is too thin. The public record should change in visible, bounded ways:

• the affected kitchen or district should show which task failed or remains unknown;
• the relevant roles should show retraining and retest dates before full restart;
• the same failure mode should be checked across similar kitchens, not only the kitchen where children became ill;
• the school handoff record should show whether labels, consume-by time, and no-take-home instructions were understood and practiced;
• route and capacity records should show whether kitchen load was reduced when beneficiary numbers exceeded safe production capacity;
• complaint channels should show aggregate response times, escalation completion, and closure categories, without exposing children’s medical details.

In disaster-mode feeding, the record should change again. Flood, heat, haze, earthquake, power outage, bridge closure, and 3T travel are not rare edge cases in Indonesia. They are predictable operating conditions. A competence record that only covers ordinary service will overstate readiness.

What MBG Watch will monitor

MBG Watch will look for evidence that BGN’s control layer is moving from declaration to performed competence.

The strongest signs would be public, dated records showing role-based drills, failed scenarios, corrective actions, retraining, retesting, and supervisor closure. The weakest sign would be a growing stack of certificates, training claims, and AI-enabled guidance with no public proof that the work can be performed without assistance under pressure.

What I am uncertain about is the present state of BGN’s internal training records. The public record shows SLHS requirements, inspections, suspensions, labels, and district evaluations. It does not yet show a national, privacy-protecting competence loop for the people who make those controls real.

That is the next evidence MBG needs.

Not a public list of workers to blame. Not a claim that every kitchen is now trained. A competence record: what was practiced, what failed, what changed, and what passed when tested again.