Before the Number Becomes a Fact: The Provenance Standard MBG Needs for Public Claims

MBG Watch · 2026-08-18

The number that moves faster than the evidence

A large public program cannot run without numbers. MBG needs beneficiary totals, kitchen counts, suspension lists, food-safety reports, emergency-feeding figures, budget realization, and nutrition indicators. The danger is not that these numbers are provisional. In a program operating across thousands of kitchens and tens of millions of claimed recipients, many numbers will be provisional at first.

The danger is that a provisional number becomes a public fact before its source, denominator, method, date, uncertainty, and correction path are inspectable.

That is what MBG Watch saw in “The Jayapura Stunting Claim.” A public narrative linked MBG to sharp changes in stunting without the minimum provenance needed to test the biological timeline, the unit of geography, the age cohort, the baseline, the measurement method, and the alternative explanations. The issue was not whether better meals can help child nutrition over time. They can. The issue was whether a specific public claim had earned the weight placed on it.

The same problem now appears across MBG’s operating record. BGN’s own January 2026 release said MBG restarted the year with 19,188 SPPGs and projected 82.9 million beneficiaries and about 21 billion meal portions in 2026. Another BGN article described the 2026 target as 82.9 million beneficiaries. The Ministry of Finance later reported, through Tempo, Rp36.6 trillion in MBG budget realization by February 21, 2026 — 10.9 percent of a Rp335 trillion allocation — while citing 60.24 million recipients and more than 23,678 SPPGs. In July, several outlets reported BGN was checking the validity of about 63 million beneficiary records and the SPPG incentive scheme.

Those figures may each be reasonable in their own context. They are not automatically comparable. “Target,” “reached,” “registered,” “served,” “validated,” “unique beneficiary,” “daily recipient,” and “eligible person” are different claims. If they are published without a claim card, the public cannot tell whether the number changed because the program changed, the denominator changed, the definition changed, or an earlier estimate was corrected.

What the evidence supports

The public record supports four narrow conclusions.

First, MBG has reached a scale where small definitional errors become large fiscal and safety errors. BGN’s January release projected 82.9 million beneficiaries and about 21 billion portions for 2026. Tempo’s February report put budget realization at Rp36.6 trillion out of Rp335 trillion by February 21 and cited 60.24 million recipients. At this scale, a one-percent counting error is not administrative noise. It can mean hundreds of thousands of people and large procurement, staffing, and reimbursement consequences.

Second, beneficiary validation is no longer a side question. It is central evidence. The July reports that BGN was rechecking about 63 million beneficiary records matter because duplication, overcounting, school overlap, and mismatched eligibility do not only distort a dashboard. They change the denominator for coverage, the numerator for cost per meal, and the fraud surface for payments. This is why MBG Watch’s earlier piece, “BGN Asks the Question It Should Have Asked First: Is 63 Million Beneficiaries Real?”, treated the count as program infrastructure, not a footnote.

Third, kitchen-status claims need the same discipline. BGN reported that from January 6, 2025 to May 29, 2026, 8,182 SPPGs had at some point been suspended and 2,213 remained suspended. Its own detail separated suspensions linked to “kejadian menonjol” from those tied to infrastructure, organization management, and nutrition quality. The same record named concrete compliance gaps, including 1,043 SPPGs that had not registered hygiene-sanitation certificates and 443 without a standard wastewater treatment installation. That is useful public information. It becomes much more useful if every suspension and reopening number carries a status date, category, province, responsible verifier, reopening condition, and correction log.

Fourth, food-safety counts need a lab and care-status trail, not only a victim total. Tempo reported January incidents in Muaro Jambi, Kudus, and Cikalongkulon involving 110, 112, and 205 affected students or beneficiaries. ANTARA later reported BGN’s assurance that medical treatment and investigation were proceeding in Papua after hundreds of MBG recipients were suspected of food poisoning. Searchable public reports also cite Jayapura counts in the hundreds and lab findings such as E. coli in some cases. The least-harm public record should distinguish suspected cases, treated cases, hospitalized cases, lab-confirmed causes, kitchen suspensions, reopened kitchens, and follow-up support for children and families.

What the evidence does not support

The evidence does not support treating every official number as equally decision-grade.

A target is not a count. A press statement is not an audit. A self-reported kitchen status is not an inspection-verified kitchen status. A media-reported poisoning count is not the same as a lab-confirmed outbreak investigation. A nutrition trend is not an MBG-attributed outcome unless the timing, cohort, comparator, and confounders are made visible.

The evidence also does not support the opposite mistake: dismissing every provisional figure until a perfect audit exists. MBG is too large, and children’s food safety is too time-sensitive, for silence to be the standard. Officials need to speak early during emergencies and incidents. The correction is not silence. The correction is honest labeling.

The public should be able to see whether a number is preliminary, estimated, self-reported, inspection-verified, lab-confirmed, audited, corrected, or superseded. Uncertainty is not a scandal. Hidden uncertainty is.

The MBG claim card

MBG Watch proposes a compact public “claim card” for every major MBG number. It should be short enough to attach to a press release, dashboard cell, spreadsheet row, or social-media graphic.

Minimum fields:

1. Claim text: the exact sentence being asserted.
2. Claim class: beneficiary count, meal portions, kitchen status, food-safety incident, nutrition outcome, budget execution, reimbursement, emergency feeding, canteen pilot, or other.
3. Source institution: BGN, Ministry of Finance, Ministry of Health, BPOM, BNPB, local government, school, SPPG, hospital, police, media, or other.
4. Geographic unit: national, province, district, school, kitchen, evacuation center, or route.
5. Population and denominator: who is counted, who is excluded, and what “one” means — person, daily recipient, meal portion, kitchen, school, case, rupiah, or household.
6. Method: administrative record, field inspection, lab test, health-facility report, financial disbursement record, survey, media report, self-report, or model estimate.
7. Period covered: the exact dates the claim covers.
8. Timestamp published: when the number entered the public record.
9. Verification status: preliminary, self-reported, inspection-verified, lab-confirmed, audited, corrected, or superseded.
10. Uncertainty or confidence: confidence interval where statistical, error band where estimated, or a plain uncertainty note where formal intervals are not available.
11. Privacy protection: whether the public version is aggregated, de-identified, k-anonymized, or otherwise protected from identifying children, families, schools, or patients where exposure would cause harm.
12. Correction trail: a stable ID for the claim, prior versions, replacement figures, reason for correction, and date corrected.
13. Machine-readable archive: CSV, JSON, or other structured record that can be downloaded and cited later.
14. Contact or responsibility line: the office accountable for correcting the claim.

A claim card does not need to be elaborate. It needs to prevent category mistakes. If “60.24 million recipients” means unique registered beneficiaries, say that. If it means active served recipients on a certain date, say that. If it is estimated from SPPG capacity, say that. If it excludes pregnant women, breastfeeding mothers, toddlers, remote areas, or school absences, say that.

How the card changes the main MBG claim classes

Beneficiary totals: A decision-grade beneficiary claim should distinguish eligible population, registered beneficiaries, validated unique beneficiaries, daily served beneficiaries, and actual meals consumed. It should disclose the deduplication method and whether one school, one child, or one service unit is the counting anchor. This extends “Seen Without Being Watched”: validation can be inspectable without exposing children’s identities.

Nutrition outcomes: A decision-grade nutrition claim should disclose the indicator, age group, baseline period, measurement method, sample size, comparator, geography, and plausible timing. Stunting is a chronic growth indicator. It should not be publicly attributed to MBG without a credible time window and method. A short-term claim may be about meal consumption, attendance, dietary diversity, anemia screening, or acute illness; it should not borrow the authority of a long-term stunting outcome unless the evidence supports it.

Kitchen evaluated, suspended, and reopened: A decision-grade kitchen-status claim should separate “evaluated,” “failed,” “suspended,” “major suspended,” “reopened,” and “operating under corrective action.” It should include the inspection body, date, standard failed, remediation required, and reopening verifier. This extends “Inspectable by Design”: a score is only useful if the public can see what was scored.

Food-poisoning counts: A decision-grade incident claim should separate suspected exposure, symptomatic cases, medically assessed cases, hospitalized cases, lab-confirmed pathogen, deaths if any, and resolved cases. It should state whether the kitchen was suspended, whether food samples were tested, and whether families received follow-up. This is the standard MBG Watch named in “After the Incident: What MBG Owes Children and Families When Meals Cause Harm.”

Emergency meals served: A decision-grade emergency-feeding claim should separate normal MBG meals, shelter meals, social-assistance meals, and ad hoc local support. ANTARA reported seven emergency public kitchens for about 31,000 displaced residents after the NTT earthquake. If MBG kitchens are used in such a response, the public claim should show which kitchens served shelters, how many meals were delivered, which budget line paid, and whether normal school feeding was paused or continued. That extends the Flores emergency-feeding, reimbursement, and aftershock-status ledgers MBG Watch published this week.

Budget execution and per-meal cost: A decision-grade fiscal claim should state whether the figure is allocation, contract value, cash disbursement, accrued obligation, kitchen reimbursement, or verified meal cost. Per-meal cost should name the denominator: planned portions, delivered portions, consumed portions, or reimbursed portions. Without that, “cost per meal” can become a rhetorical number rather than a financial control.

Canteen-pivot results: A decision-grade canteen claim should separate schools enrolled, canteens certified, meals served, children reached, incidents, menu compliance, procurement route, and payment route. This extends “The Visibility Standard”: before a pivot scales, the public needs the record that tells whether the model is safer, cheaper, fairer, or merely less visible.

The privacy boundary

Claim provenance is not surveillance.

The public does not need children’s names, health records, household identities, faces, biometrics, or school-level details where small numbers would make a child identifiable. The public does need aggregated, inspectable evidence that the count is real and corrected when wrong.

The right boundary is: private identity, public method. BGN can publish the deduplication rule without publishing the child. It can publish an inspection result without exposing a kitchen worker’s personal data. It can publish a lab-confirmed pathogen without naming a sick student. It can publish a correction trail without exposing raw household records.

This is the continuity between “Seen Without Being Watched” and this piece. A privacy-preserving system is not one where the public is asked to trust a black box. It is one where the public can inspect the method, the aggregate result, the error handling, and the accountable office while sensitive personal data stays protected.

What MBG Watch will treat as decision-grade

Going forward, MBG Watch will treat a public MBG number as decision-grade only when it carries enough provenance to answer six questions:

1. What exactly is being counted?
2. Who counted it?
3. Over what place and time?
4. By what method?
5. How uncertain, preliminary, corrected, or verified is it?
6. Where is the correction trail?

If a number lacks these fields, MBG Watch may still report it, but will label it according to its evidentiary state: official target, official estimate, media-reported count, self-reported kitchen figure, inspection-verified status, lab-confirmed incident, audited expenditure, corrected figure, or unresolved claim.

This is not a demand for perfection before action. It is a demand for legibility while action happens.

The borrowed lesson from current evidence-integrity work, including AGA’s wider signal about provenance and calibrated uncertainty, is simple enough to state without jargon: a system becomes safer when answers carry their receipts. For MBG, the receipt is not decoration. It is how children’s wellbeing, public money, and public trust remain connected to reality as the program grows.