When Evidence Becomes Reusable: The Record MBG Needs Before Public Claims Become Tools

MBG Watch · 2026-10-09

The premise

Scientific evidence is moving from static reading material toward interactive reuse. In September 2026, Nature published Paper2Agent, a framework that converts a research paper, code, data, and workflows into an AI agent that can answer questions and reproduce analyses through tools exposed by the paper itself. The authors describe the shift plainly: a paper becomes less like a finished PDF and more like an active system connected to its own methods, datasets, and tests.

That does not mean Indonesia’s Makan Bergizi Gratis program should turn every dashboard, nutrition claim, complaint, kitchen record, or audit finding into an AI agent. That would be the wrong lesson.

The useful lesson is narrower: public evidence is now likely to be queried, scraped, summarized, remixed, scored, and reused whether the evidence owner designs for that reuse or not. If the evidence package is thin, reuse can make uncertainty look settled. If the evidence package is well-formed, reuse can make accountability easier to check.

This is the next layer after MBG Watch’s earlier work on claim cards, durable logs, failure transparency, outcome attribution, route-specific evaluation, and hostile instructions entering records. Those pieces asked whether an individual number, log, status, or digital route could be trusted. This one asks what must travel with the record when other people — and future tools — carry it somewhere else.

What the current MBG record already contains

BGN’s public record is no longer only speeches and press releases. It is becoming an operating layer.

The clearest public-facing example is Radar MBG, which presents itself as “Menu MBG Hari Ini” and asks a user to select province, city or regency, district, village, and institution to see the day’s menu and the supplying SPPG. The page also invites users to send experience, suggestions, and obstacles through a “Hubungi Kami” channel.

BGN’s July 2026 transparency release describes a wider digital ecosystem: parents would be able to see which schools receive service, what menu is served, which kitchen cooks it, who the responsible SPPG head is, and what service coverage applies. The same release says BGN is developing a public dashboard with operating SPPG counts, geographic distribution, served schools, and national implementation progress.

BGN has also described a more direct quality-record layer. In May 2026 it launched Reviu MBG, an application through which designated PICs — including school teachers, posyandu heads, and pesantren administrators — can assess food when it arrives. BGN named four initial parameters: distribution punctuality, aroma, taste, and menu variation compared with the previous day. It also said the results would become evaluation indicators for each SPPG, while noting that the early stage was not yet a sanction basis.

The beneficiary record is also being digitized. In April 2026, BGN said it opened validasidata.bgn.go.id for local governments, schools, and posyandu to check whether beneficiary data had been recorded. The same release says BGN planned API-based integration across ministries and agencies, with data sources including education, religious affairs, and health ministries.

The complaint record is public enough to matter. BGN says SAGI 127 operates 24 hours for complaints, input, and clarification, and that incoming reports will be verified and followed up under the applicable mechanism.

The safety record is public enough to require a stronger evidence package. In August 2026, BGN said 504 MBG kitchens had been permanently closed after being found sanitation-unfit based on health-office reports transmitted through the Ministry of Health, while about 3,000 more kitchens were under inspection.

Taken together, these are not just communications. They are reusable public facts: menus, responsible kitchens, beneficiary validation, quality review, complaint handling, kitchen eligibility, and food-safety status.

The reusable evidence package

A reusable MBG record should be small enough to publish consistently, but complete enough that a reader can tell what the record is and what it cannot prove.

For each public claim or operating record, the evidence package should carry at least:

• Record identity: a stable record ID, title, issuing unit, responsible owner, and contact or appeal route.
• Time boundary: creation time, observation time, update time, and whether the record is current, superseded, corrected, or withdrawn.
• Scope: province, district, SPPG, school, route, beneficiary category, or national aggregate — without exposing child-level identity.
• Denominator: the base number behind any percentage, count, delay rate, quality score, incident rate, or coverage claim.
• Method: how the number was produced — direct observation, kitchen self-report, PIC review, ministry data feed, health-office inspection, complaint intake, laboratory confirmation, or administrative reconciliation.
• Source chain: which office, system, form, API, or inspection produced the underlying record.
• Uncertainty and limits: missing districts, late reports, unverified complaints, sampling limits, duplicate risks, self-reporting risks, and exclusions.
• Version history: what changed, when it changed, who approved the change, and whether the old value remains visible.
• Correction path: how a school, family, local government, journalist, or auditor can contest a record and see the resolution.
• Machine-readable attachment where safe: CSV, JSON, or a package manifest for aggregate data, with field definitions and redaction rules.

This is not exotic. The web already has mature language for this kind of work. The W3C PROV family defines provenance as information about the entities, activities, and people involved in producing data, useful for assessing its quality, reliability, or trustworthiness. RO-Crate, a research-object packaging specification, shows a practical pattern for bundling data and metadata so files, workflows, authorship, context, license, and provenance travel together.

MBG does not need to copy either standard mechanically. A meal program is not a genomics paper. But the transferable principle is exact: the record must carry enough context for reuse to remain accountable.

What evidence reuse must not become

Reusable evidence is not permission to publish everything.

The public package should not include child-level identity, individual health status, household location, pregnancy details, complaint-maker identity, private school-level contact information beyond official channels, or security-sensitive implementation details that would make kitchens or delivery routes more vulnerable.

It should also not turn an aggregate score into automated authority. A Reviu MBG percentage can help show where distribution is late or where quality complaints cluster. It should not, by itself, become a sanction, funding cut, school label, or kitchen blacklist unless the record shows method, denominator, appeal, validation, and human responsibility.

The same boundary applies to future AI use. A system may summarize public evidence, compare districts, or flag missing denominators. It should not pretend to know what the underlying record never observed. If a complaint is unverified, an agent’s answer must say unverified. If a nutrition outcome claim rests on before-after association rather than causal attribution, reuse must not promote it into proof.

The danger is not artificial intelligence as a category. The danger is portable certainty detached from its evidence boundary.

The least-harm path

BGN does not need to build the perfect public data architecture before publishing better records. It can start with five high-risk record families.

First, nutrition-outcome claims. Any claim that MBG improved attendance, anemia, stunting, learning, or household burden should travel with measurement dates, comparison group, denominator, instrument, attrition, confounders considered, and whether the claim is administrative, observational, or causal.

Second, food-poisoning and food-safety incidents. Each public incident record should show report date, location granularity, affected count range, verification status, suspected source, lab or health-office confirmation status, kitchen action, service continuity decision, and remedy route. Privacy should be strict; accountability should not be vague.

Third, kitchen operating status. If a kitchen is active, suspended, remediated, or closed, the public record should state the inspection basis, sanitation status, date of last inspection, responsible authority, appeal status, and what happens to beneficiaries during service interruption.

Fourth, beneficiary counts and validation. Since BGN has described ministry data sources and API integration, public aggregates should show which source contributed which population, how duplicates are handled, how often records refresh, and how local governments or institutions correct errors.

Fifth, complaint and remedy receipts. SAGI 127 can become more than a hotline if each complaint category has aggregate counts, verification status, response-time bands, resolution outcomes, and a clear way to challenge closure — without revealing complainants.

This would make MBG evidence more useful to journalists, parents, local governments, researchers, auditors, and civil-society monitors. It would also make future software reuse safer, because a tool would have structured uncertainty to carry forward instead of inventing confidence from a headline.

What remains uncertain

I could verify the public existence of Radar MBG, public transparency plans, Reviu MBG, beneficiary validation, SAGI 127, and BGN’s 504-kitchen closure statement. I could not verify from public pages which of these systems already preserve full internal provenance, immutable version history, appeal records, API documentation, or machine-readable public extracts.

It is also unclear how much detail can safely be published at district, SPPG, school, or service-route level without creating privacy, stigma, retaliation, or security risks. That line should be drawn conservatively and revisited with health, child-protection, data-protection, local-government, and beneficiary representatives.

The right standard is not maximum openness. It is safe contestability: enough public evidence that claims can be checked, corrected, and reused without exposing the people the program exists to serve.