Not the Dashboard, the Guidance: What Human-Amplifying Interfaces Can Teach MBG Kitchens

MBG Watch · 2026-08-23

The premise

MBG Watch’s earlier analysis, “From SOP Posters to Guided Practice”, argued that MBG’s food-safety problem is not only a shortage of written rules. It is a capability problem at the moment of action: the cook starts too early, the route runs late, the school day shifts, the weather changes, the water is unsafe, a child wants to take food home, or a shelter meal has to be served outside the normal school chain.

This piece extends that argument. It asks what MBG can learn from a wider class of human-amplifying interfaces: tools that help a person perform complex work closer to expert level by reducing unnecessary mental translation, surfacing the next safe action, and leaving an inspectable record of exceptions.

The recent signal is not from food service. On 20 August 2026, MIT News described a new excavator training interface in which trainees grasp a miniature excavator-like arm instead of learning joystick mappings. In a seven-day simulator study, MIT reported that novices using the “World-Space Interface” were “just as good as experts from the start,” while novices on joystick controls remained below experts even as they improved. The useful comparison is not “MBG needs excavator technology.” It is more modest: when a task is physically or procedurally complex, better interface design can reduce the gap between a rule and a safe action.

For MBG, the interface does not need to be robotic or AI-led. It may be a paper card, a route phone prompt, a dispatch timer, a colored exception tag, a school receiving checklist, or a supervisor screen that asks one practical question at the right time. The standard is not sophistication. The standard is whether it improves judgement without taking judgement away.

What human-amplifying interfaces can teach

The excavator example is useful because it names the burden that ordinary training often hides: mental mapping. A new operator does not only need to know the goal. They must translate a physical world problem into a control language that is not natural to them. MIT’s interface works by making the control resemble the thing being controlled.

MBG has its own mental mappings. A kitchen worker may be told that food must be safe within a time window, but still has to translate that rule into a decision when cooking began early, dispatch was delayed, the classroom schedule slipped, or the route temperature rose. A teacher may be told not to serve unsafe food, but still has to decide what “unsafe” means when the package looks acceptable but the time record is unclear. A posyandu worker or shelter coordinator may receive meals outside the school pattern and need to know what counts as an acceptable handoff.

The WHO’s Five Keys to Safer Food are simple at the level of principle: keep clean; separate raw and cooked; cook thoroughly; keep food at safe temperatures; and use safe water and raw materials. The difficulty is carrying those principles through a large operating chain. The FDA Food Code page likewise frames the code as a model for safe handling in retail and institutional food settings, giving regulators a technical basis for rules. But neither a principle nor a code enforces itself at the dispatch door.

Clinical checklists offer a closer operating lesson. WHO describes clinical checklists as tools that guide decision-making, support consistent care, and reduce medical errors while remaining low-cost and scalable. Its surgical checklist was designed not to replace surgeons, but to improve communication, teamwork, and the capture of critical steps before harm occurs. That is the right analogy for MBG: not automation replacing accountable people, but structured guidance that makes the next safe action harder to miss.

The MBG moments where guidance matters

The public record shows why this cannot remain abstract. ANTARA reported in September 2025 that President Prabowo called for food-safety testing kits in MBG kitchens after dozens of poisoning cases, with BGN reporting 70 food-safety incidents affecting 5,914 recipients between January and September 2025. ANTARA also reported causes including E. coli, Staphylococcus aureus, Salmonella, Bacillus cereus, coliform and other contamination in food and water. The Indonesian National Police site, citing BGN, reported that 45 kitchens had failed to follow health protocols and 40 were shut indefinitely pending investigation and improvements.

Those numbers do not prove that guidance would have prevented each case. They do show that MBG’s control problem is not theoretical. It lives in routine work: water condition, kitchen hygiene, raw-cooked separation, holding time, route timing, meal inspection, and the authority to stop service.

This is where MBG Watch’s recent pieces connect. “When the Kitchen Gets Hot,” “Hourly Heat, Not Daily Heat,” “Weather at the Dispatch Door,” and “When the Day Has to Move” all pointed to conditions that change inside a day. “The Last Meter of the Meal” and “The Care Has to Travel” extended the same point beyond the classroom: the meal remains a safety responsibility through the final handoff, including care routes, posyandu settings, and shelters. “Seen Without Being Watched” set the privacy boundary: beneficiaries and workers should be visible enough for the program to correct errors, but not turned into surveillance subjects.

A guided-practice layer would sit exactly at that boundary. It would not ask, “Can we collect more data about everyone?” It would ask, “At this handoff, what condition matters, what decision was made, what exception occurred, and what correction followed?”

What MBG should borrow

First, MBG can borrow timely prompts. A kitchen should not only record that cooking began at 04:30. The system, paper or digital, should translate that fact into the next decision: last safe dispatch time, required holding condition, school receiving deadline, and what to do if the route is late.

Second, MBG can borrow condition-tied checklists. A generic “follow SOP” reminder is weak. A route-day checklist that changes when heat, rain, damaged roads, school closures, water disruption, or shelter feeding is present is stronger. The worker does not need a lecture at the critical moment. They need the three checks that matter now.

Third, MBG can borrow exception capture. The safest systems do not pretend abnormal days are rare. They ask workers to record the exception plainly: “meal held longer than plan,” “water test unavailable,” “route vehicle delayed,” “school refused tray,” “beneficiary absent,” “meal moved to shelter site.” An exception record protects the child first, and it protects honest workers from being forced to hide the reality of the day.

Fourth, MBG can borrow stop-go support. A guided interface should be allowed to say, “do not dispatch,” “do not serve,” “serve only after supervisor check,” or “replace menu for this route.” That authority must be public enough to inspect. If a dashboard only reports after the fact, it is not guidance.

Fifth, MBG can borrow practice loops. The excavator study was conducted in simulation before real-world operation. WHO’s emergency checklist materials include simulated use. MBG should treat high-risk meal days the same way: rehearse a route delay, a water problem, a late school serving time, a shelter diversion, and a take-home refusal before the first crisis makes the rehearsal real.

What MBG should not borrow

MBG should not borrow opaque model scores. MBG Watch’s “Inspectable by Design” argued that the useful lesson from public AI evaluation is not to adopt AI, but to make failure modes, test conditions, audit trails, and correction loops visible. A kitchen score that cannot explain its evidence is not enough. A beneficiary score that cannot show which failure mode it is trying to prevent is not enough.

MBG should not borrow biometric shortcuts. If a child, pregnant mother, toddler, worker, or caregiver must become a faceprint before receiving food or being trusted, the program has crossed a line it does not need to cross. Beneficiary validation can be specific without being intimate. Attendance, school rosters, posyandu records, exception logs, and local correction mechanisms can do much of the work without collecting sensitive identifiers by default.

MBG should not borrow worker surveillance. A guided-practice layer should make unsafe conditions visible, not rank workers by constant watchfulness. If the interface becomes a tool for punishing every deviation, workers will learn to hide exceptions. Food safety needs the opposite: fast, low-shame reporting when the day has gone off plan.

MBG should not borrow vendor lock-in. A kitchen safety prompt, dispatch timer, and receiving checklist should be understandable in public language and portable across systems. If only one vendor can explain the control logic, the public cannot audit it.

Most of all, MBG should not borrow automation that displaces accountability. A related AGA governance question from work on agentic operational risk travels well here: who gave the system authority, what can it touch, what is it forbidden to touch, what evidence remains after it acts, and who can revoke or correct it? For MBG, the answer must remain human and public. Tools may guide. They should not quietly govern.

The minimum public record

Any guided-practice tool used in MBG should leave a small public control record. It does not need to expose children’s identities, worker biometrics, family details, or exact household locations. It does need to show the action chain.

The minimum record should include: the action prompted; the condition observed; the human decision made; the exception, if any; the correction or escalation; the data minimized or deliberately not collected; and the revocation path if the tool gives bad guidance.

For example, a route record might say: cooked 05:10; planned dispatch 07:00; actual dispatch 07:35; ambient heat flag active; school receiving deadline 09:10; received 08:42; no visible spoilage; served 09:00; leftovers not allowed; exception: route delay; correction: adjust next-day cooking start for route B. That is not surveillance. It is an operating memory.

A shelter feeding record might say: meals diverted from school route to temporary shelter; beneficiary count estimated; no individual biometric validation used; receiving coordinator named; safe serving window posted; unused meals discarded after deadline; next-day demand corrected. Again, the purpose is not to watch people. It is to keep care coherent when the normal chain breaks.

The least-harm path

The least-harm path is not a national procurement for a “smart kitchen” platform. It is a small, public, reversible pilot in a few high-risk contexts: one heat-prone route, one water-constrained kitchen, one long-distance 3T route, one school cluster with repeated schedule delays, and one emergency shelter feeding scenario.

The pilot should begin as training and evidence, not as enforcement. Workers should help design the prompts. Teachers, posyandu workers, route coordinators, kitchen staff, and parents should be asked where the current SOP becomes unclear. The first success metric should not be a dashboard score. It should be whether workers catch more unsafe conditions early, record exceptions more honestly, and know who can authorize a stop.

The failure modes should be published before the pilot scales: missed prompt, false alarm, worker override, unclear authority, data collected unnecessarily, route condition not represented, school unable to respond, beneficiary privacy risk, vendor explanation gap. Corrections should be published with the same discipline.

This is the worker-facing companion to “Inspectable by Design.” MBG should borrow interface discipline, not automation ideology. It should build tools that make the right action easier at the moment of risk, while keeping children, mothers, toddlers, workers, and caregivers protected from needless surveillance.

What I am uncertain about

The evidence for human-amplifying interfaces is strongest as analogy, not as direct proof for MBG. The MIT excavator study concerns simulator training for heavy equipment, not food-service governance. WHO checklist evidence comes from health care, where teams, authority structures, and legal duties differ from school meal delivery. The MBG incident record shows repeated food-safety failures, but public reporting is still too incomplete to assign each failure to training, equipment, time-temperature control, water, staffing, procurement, route design, or supervision.

That uncertainty argues for a narrower conclusion, not a stronger one. MBG does not yet need a grand interface platform. It needs a guided-practice pilot with a public control record, a privacy boundary, and the humility to stop if the tool produces compliance theatre instead of safer meals.