Use Case MBG: Collecting and Analyzing Indonesian News on Makan Bergizi Gratis

This guide walks the public workflow, from collection through aggregate analysis, for the Makan Bergizi Gratis (MBG) policy research corpus covering 2025-01-05 through 2026-07-31, using the newswatch registry as currently configured. It is the companion to practical-guide.md and does not duplicate installation, configuration, or troubleshooting content.

Last updated 2026-08-08 against the 2026-08-06 collection run (10,915 cleaned documents, 52 sources). This supersedes an earlier edition built on a smaller run over a shorter window; every figure and count below has been regenerated, and the superseded totals are not carried forward. The topic layer now carries a stability disclosure: see Topic resolution and stability before quoting any topic-level number.

Background: What Is MBG?

Makan Bergizi Gratis (MBG) is Indonesia’s national free nutritious-meals program, administered by the Badan Gizi Nasional (BGN). It provides meals designed against daily nutritional adequacy standards for school students, pregnant women, breastfeeding mothers, and young children, delivered through Satuan Pelayanan Pemenuhan Gizi (SPPG), the local service units that prepare and distribute meals.

BGN describes MBG as both a nutrition intervention and a platform for nutrition education, and its operating model links SPPG procurement with local farmers, fishers, cooperatives, and small businesses. That scale makes MBG a useful news research case: reporting spans beneficiary access, kitchen expansion, food safety, procurement, public finance, regional implementation, oversight, and political accountability.

The corpus starts on 5 January 2025, when BGN formally introduced the 2025 program, and includes implementation from 6 January 2025 onward. Official program context and operating details:

• BGN’s program launch statement
• BGN’s MBG frequently asked questions
• BGN’s SPPG quality-oversight statement

Collection Command

Run the stable, search-capable registry subset against the MBG keywords for the declared window. During the documented collection, --scrapers all resolved every stable entry with supports_search=True, not all 75 sources then registered.

uv run newswatch \
  --method search \
  --keywords "mbg,makan bergizi gratis,program MBG,satuan pelayanan pemenuhan gizi,SPPG,badan gizi nasional" \
  --start_date "2025-01-05" \
  --daterange "2025-01-05/2026-07-31" \
  --scrapers all \
  --scraper-timeout 180 \
  --max-concurrent-scrapers 6 \
  --output_format jsonl \
  --output_path mbg-all.jsonl \
  --progress

Sources run in waves under the concurrency cap, so wall-clock time is roughly (sources ÷ cap) × --scraper-timeout, not a single flat timeout.

What --scrapers all resolves to in this repo:

• 75 registry entries total.
• 68 stable entries support search and are resolved by --scrapers all.

Corpus Validation

Validate each collection before analysis; retrieval totals are evidence about that run, not fixed properties of the news ecosystem.

1. Check the schema. Require title, publish_date, content, link, and source on every record. Reject malformed rows rather than filling missing evidence with inferred values.
2. Enforce the study window. Parse publication timestamps and retain only 2025-01-05 00:00:00 through 2026-07-31 23:59:59 (both full calendar days).
3. Confirm relevance. Keep a record only when its title contains a standalone program term (MBG, SPPG, or BGN) or an explicit program phrase: Makan Bergizi Gratis, Program MBG, satuan pelayanan pemenuhan gizi, or Badan Gizi Nasional. Matching is case-insensitive, acronym boundaries reject collisions such as PVMBG, and title anchoring excludes tangential articles that mention an MBG term only in the body.
4. Remove duplicates in order. Deduplicate exact article links first, then lowercase and collapse whitespace in titles before removing repeated titles. Preserve the number removed at each step so the final corpus is auditable.
5. Publish aggregates only. Reconcile retained records by source and calendar month, but keep article text, titles, URLs, and document-level outputs in the private research workspace.

The reference run (collected 2026-08-06) documented in this guide yielded 27,927 well-formed records before cleaning, 20,992 after relevance filtering, 10,957 after URL deduplication, and 10,915 cleaned documents from 52 contributing sources across all 19 calendar months in the window. These are one run’s retrieval counts, not estimates of article production; inspect the actual numbers in each run rather than treating one retrieval as a fixed benchmark.

Aggregate Analysis

All figures below describe the reference run’s cleaned corpus. Topic annotations are provisional English summaries derived from auto-generated Indonesian term statistics and a manual review of private topic assignments; named entities retain their source-language proper names. Treat both as working labels pending further validation before citing them as facts.

Topic landscape and prevalence

The cleaned corpus resolves to 14 substantive topics plus an outlier class. The substantive topics span program governance and institutional oversight, corruption investigations and prosecutions, student food-poisoning incidents, budget allocation and disbursement, SPPG kitchen construction and police-run operations, electric-motorcycle and operational-vehicle procurement, Jakarta policing and school incident response, dairy production and cattle supply, SPPG site locations and regional coordinators, health ministry programmes and TB treatment, music and viral video mentions, listed agrifood companies and share prices, fisheries, aquaculture, and marine food supply, and children, infants, and family nutrition; the outlier class captures documents that do not cluster cleanly with any dominant theme.

The outlier class is large and is disclosed rather than hidden: 4,031 documents (36.9%) carry no topic, so any topic-level statement below describes the remaining 63%. That 36.9% outlier share is one of the stable properties of the corpus: refitting under five different projection seeds moves it only between 34.6% and 37.8%, and the substantive topic count stays at 14 every time. Several other topic-level numbers below are far less stable, and each is qualified where it appears.

A two-dimensional UMAP projection of the cleaned documents, colored by topic; each point is one document. All 14 substantive topics are labeled directly over their cluster regions, and grey points are documents the clustering model left unassigned. Separation and overlap are diagnostic patterns, not proof that the generated labels are definitive categories. The scatter uses a dedicated readability projection (random_state=42, n_neighbors=30, min_dist=0.3, metric=cosine), distinct from the analysis UMAP inside the BERTopic pipeline that produces the topic assignments; only the two-dimensional layout differs, never the assignments.

Topic-size distribution ordered largest to smallest, including the outlier class. The three largest topics account for 5,248 of 10,915 cleaned documents (48.1%); T0 alone holds 3,724, so topic structure outside T0 rests on roughly a quarter of the corpus. The remainder spreads across the other 11 substantive topics and the outlier class.

Both figures are seed-dependent. Across five projection seeds the three largest topics span 4,745 to 5,796 documents (43.5% to 53.1% of the corpus), and T0 spans 2,815 to 4,209, which is 39.5% to 60.3% of assigned documents against the 54.1% shown here. The qualitative shape, one dominant topic and a long thin tail, holds in every seed; the specific counts do not.

Per-topic volume over the 19 calendar months for the six largest topics: T0 program governance and institutional oversight, T1 corruption investigations and prosecutions, T2 student food-poisoning incidents, T3 budget allocation and disbursement, T4 SPPG kitchen construction and police-run operations, and T5 electric-motorcycle and operational-vehicle procurement. Each topic uses a distinct colorblind-safe color and line style, with the topic label printed directly at the right-hand endpoint of every line, so each series stays legible without relying on color alone.

Five callouts mark each annotated peak. The month and the document count are derived from the corpus at render time; the event text is curated and names the headline terms it answers to, so a reader can check it against the published evidence table rather than take it on trust:

• September 2025, T2, 179 documents: mass poisoning wave, Bandung Barat the largest cluster, kitchens closed over SOP breaches.
• October 2025, T0, 250 documents: hygiene certification required for SPPG, governance Perpres drafted.
• February 2026, T0, 359 documents: Ramadan menu switched to dry food, MBG extended to elderly and disabled recipients.
• June 2026, T1, 508 documents: prosecutors name Dadan Hindayana and Sony Sonjaya as suspects.
• June 2026, T0, 889 documents: SPPG halted over school holidays, with students demanding a stop and SPPG staff rallying to keep it.

The peak months are stable; the peak counts are not. Under five projection seeds every one of these topics peaks in the same month it does here, but the counts move: the T0 June peak spans 606 to 889, the T1 June peak 435 to 508, and the T2 September peak 179 to 218. The values printed above are the seed the published model uses, and for T0 and T1 they are the highest of the five.

Across all documents, June 2026 is 3.20× its predecessor (2,761 against 863 in October 2025). That comparison uses no topic assignment and is the more reliable statement of the spike. Restricted to the three largest topics, the same month reaches 1,412 documents, about 3.5× their previous combined monthly peak of 406; that multiple ranges from 1.95× to 3.5× across projection seeds, and the value quoted here is the largest of the five. Prefer the all-document figure when citing the spike.

Treat the peak as real and the multiple as an upper bound. Distinct sources contributing per month rise from 17 to 45 across the window, so later months are searchable by more sources than earlier ones and raw month-over-month growth is partly a coverage artifact. Any published trendline should ship the sources-per-month series beside it. Annotations describe coverage families contemporaneous with each peak, not claims that those events caused the volume; the pattern describes this retrieved corpus and should not be extrapolated beyond 2026-07-31.

Independent public reporting documents the same coverage families without quoting private records:

• Bandung Barat mass-poisoning response, ANTARA (September 2025)
• Ramadan dry-food adaptation statement, ANTARA
• BGN school-holiday audit of MBG kitchens, ANTARA
• SPPG safety certification push after poisoning cases, Kompas

Topic resolution and stability

The pipeline reduces the clustering to a fixed number of display topics. That number is an editorial choice, and this section states plainly how it was made and what it does not support.

How it was chosen, and why that no longer holds. The reduction target was selected from the candidates 10, 12, 15, and 18, on a corpus roughly a third the size of the current one, using the largest topic’s share of assigned documents as the criterion: 10 and 12 produced a mega-topic and were rejected, 15 and 18 did not. Two things follow. Twenty and above were never candidates, so the published value was not chosen over them. And on the current corpus the criterion no longer endorses the choice: the largest topic now holds 54.1% of assigned documents, nearer the level that was rejected than the level that was accepted.

What was run to test it. A resolution was swept from 8 to 40 in steps of 2, under five projection seeds, with twenty bootstrap resamples at 90% of the corpus, against three criteria fixed before the sweep: no dominant residual bucket and no shattering into slivers; agreement across seeds at adjusted Rand index 0.60 or better; and no more than a third of topics failing to survive resampling.

No resolution passes more than one. The intersection is empty, and it stays empty under every way of aggregating the shape criterion across seeds. Mean cross-seed agreement never exceeds 0.606 anywhere on the grid, so the instability is not a property of the reduction target: it originates in the projection and clustering stage, which is also where the 36.9% outlier share is set. Refitting at a different target would not repair either.

What that means for reading the topics. The fourteen topics are a reading aid for the volume and sentiment figures, not a set of categories the corpus contains. The three largest are reasonably identifiable across seeds, matching at 0.50 to 0.79 by document overlap. Below them the correspondence degrades sharply: the budget topic matches at 0.12 to 0.55, and the SPPG-construction topic at 0.10 to 0.77, meaning a topic present in one run may have no counterpart in another. Topic counts, labels, and per-topic shares should be read as one plausible partition rather than the partition.

No claim is made that this resolution is optimal. Any earlier implication to that effect is withdrawn.

What the largest topic contains

Reduction merges clusters; it never re-clusters. Every display topic therefore has a fan-in, the set of underlying clusters merged into it. For twelve of the fourteen topics that number is small, a median of 2.

T0 is 83 of the 150 underlying clusters merged into one. Its constituents include chicken prices and poultry farmers (240 documents), SPPG water treatment installations (214), viral menu videos (187), Ramadan menus (111), hygiene certification (105), breastfeeding and infant nutrition (87), and a Persebaya Surabaya football cluster (56). These are not facets of a single theme. T0 functions as a residual bucket, and the label “program governance and institutional oversight” describes its largest tendency rather than its contents. This is structural rather than an artifact of the chosen resolution: the largest topic still absorbs 73 clusters at a target of 20 and 51 at 28.

The practical consequence is that a real, well-covered theme can be invisible in the topic list. Street mobilisation is one. Four of T0’s clusters carry it, 260 documents in total, with these leading terms:

Titles were matched on demonstrasi, unjuk rasa, aksi massa, turun ke jalan, long march, massa aksi, berunjuk, and demo, giving 162 distinct titles across the whole corpus, of which 135 fall inside T0 and 128 inside the four clusters above. All 138 that matched only on the ambiguous token demo were read individually, since the word also denotes a product or cooking demonstration in Indonesian: 138 of 138 are genuine street mobilisation.

Two features of the coverage are worth stating, because a topic label would hide both. Roughly half is mobilisation in support of the programme, by SPPG staff, kitchen volunteers, and organised supporter groups demanding it continue, alongside student protests demanding it stop and a sub-thread disputing whether the supporter rallies were paid or state-directed. And the coverage is not confined to one month: of the 162 titles, 137 fall in 2026-06, 13 in 2026-07 and one in 2026-02, but an earlier and separate cluster of 11 sits in 2025-02, covering student refusals of the programme in Nabire and Jayapura, Papua.

Document-similarity network

The method builds an undirected graph over the cleaned-document embeddings. Each node is one document; an edge connects two documents when their cosine similarity is at least 0.90 among each document’s k=5 nearest neighbors.

The full graph contains 2,636 active nodes and 2,847 edges across 735 active connected components, with 8,279 isolates. The bounded figure displays 36 communities within the 4 largest component groups under a 500-node cap, rendering 499 linked documents; smaller components are shown in grey and carry no group annotations.

The four leading groups shown are:

• G1, budget allocation and disbursement: 222 documents, 64.4% dominant.
• G2, corruption investigations and prosecutions: 114 documents, 71.9% dominant.
• G3, student food-poisoning incidents: 107 documents, 78.5% dominant.
• G4, corruption investigations and prosecutions (separate component): 56 documents, 100% dominant.

The fitted color areas follow the displayed component extent only: they are not confidence regions, ground-truth boundaries, or calibrated estimates. “Dominant” is the share of documents inside the displayed component carrying the indicated topic label, not a probability that the component is exclusively about that topic.

Interpretation boundary. Proximity and edges mean embedding similarity between retrieved documents, not shared event identity, factual equivalence, causation, coordination, or editorial influence. The deterministic spring layout (seed=42) makes the figure reproducible, but placement is meaningful only insofar as it reveals the connected components. Article-level evidence remains private.

Sentiment by topic

News tone varies substantially by topic. Among the larger substantive topics, student food-poisoning incidents (T2, n=599) show the clearest negative pattern: 59.9% of documents have negative as their highest-probability label and the mean probability score is −0.47. Corruption investigations and prosecutions (T1, n=925) carry a nearly identical mean score (−0.46) with 50.9% negative labels and almost no positive probability. By contrast, SPPG kitchen construction and police-run operations (T4, n=399) have a positive mean score of +0.25.

The first two directions are robust; the third is not. Holding the document-level predictions fixed and regrouping them under five projection seeds, corruption stays between −0.46 and −0.41 and food poisoning between −0.47 and −0.43, so both are strongly negative in every seed. Food poisoning’s negative share ranges 54.2% to 59.9%, and the 59.9% quoted above is the highest of the five. SPPG kitchen construction does not hold: its mean score ranges from +0.25 down to −0.04, so the sign itself flips, and in one seed the topic has no recognisable counterpart at all. Read it as “less negative than the rest”, not as a positive counterweight.

The figure reports each document’s highest-probability label as a share of its topic; the centered grey marker indicates neutral classification, and the ordering uses the topic mean of P(positive) − P(negative). The two measures can differ: program governance and institutional oversight (T0, n=3,724) has 1,145 positive, 1,310 neutral, and 1,269 negative highest-probability labels with a mean score of −0.02. That mean stays mildly negative in every seed, at −0.07 to −0.01. Smaller substantive topics contain at most 186 documents each (T6 Jakarta policing and school incident response), so their directions are especially provisional, and per Topic resolution and stability the smaller topics are also the least likely to reappear under a different projection seed. The heterogeneous outlier class is retained in aggregate reconciliation but omitted from this substantive-topic figure.

Overall, the cleaned corpus shows 2,620 positive (24.0%), 4,313 neutral (39.5%), and 3,982 negative (36.5%) highest-probability labels. The figure is the public aggregate output; document-level predictions and the underlying review tables remain private.

Interpretation boundary. The pinned Indonesian RoBERTa classifier was trained on IndoNLU SmSA comments and reviews, not MBG news. Its outputs describe the language tone of retrieved articles after right truncation at 512 model tokens, not public opinion, policy effectiveness, factuality, or stance, and the probabilities are ordinal comparisons, not calibrated population estimates.

Named entities and SPPG kitchens

Provisional entity extraction surfaces the most-mentioned people, event locations, and SPPG/Dapur kitchen references in the corpus. The figures below are aggregate counts; surface-form resolution is provisional.

Top multi-token people: 60,886 mentions resolved to 6,343 normalized surfaces. Ranks describe prominence in this retrieved corpus, not policy importance. Audited aliases combine Purbaya, Purba, Yudhi, and Yudhi Sadewa with Purbaya Yudhi Sadewa (1,773 mentions across 409 documents). Ambiguous single-token surfaces such as Yusuf are retained in aggregate tables but excluded from this precision-oriented chart rather than assigned to one person.

The three entity charts rank by documents, not mentions. A single article naming an entity a dozen times is one article’s worth of coverage, so ranking on raw mentions lets one repetitive story outrank an entity covered across many. Each bar prints its mention total alongside the document count, so repetition stays visible.

Top event locations: 1,962 mentions resolving to 562 unique places. Counts exclude publisher datelines and general geographic framing: only locations anchored to a described event (visit, launch, incident, audit) are counted, so the chart under-represents places that appear only as byline cities or background geography.

Places are settlements and administrative regions only. Venues such as Kompleks Parlemen, Istana Merdeka, and Gedung DPR are excluded: each sits inside a city that is itself counted, so ranking them together both double-counts the coverage and puts a building beside a province. Spelling and granularity variants are merged, and truncation fragments are rejected.

Top SPPG/Dapur (satuan pelayanan pemenuhan gizi) kitchen references: 1,760 mentions resolving to 804 unique kitchen surfaces. Unit numbers are preserved where available; strict normalization excludes regional collectives and malformed identifiers. These automated identifiers remain provisional until reconciled against operational records.

Person co-mention network

The method builds an undirected graph over the cleaned corpus. Each node is one normalized person surface; an edge connects two nodes when both people occur in the same document, with repeated mentions of the same normalized person deduplicated within a document before document and edge counts are incremented.

These gates produce 30 nodes and 42 edges across 7 connected components and 8 detected communities. The figure draws the largest component only: 17 of the 30 eligible nodes, with the remaining 13 spread across 6 smaller components that are not shown. Node area encodes document count, and community is deliberately not shown as colour, because Louvain community IDs are arbitrary integers and rendering them as colour on a corpus about a government programme invites a political reading the data does not support.

The method uses the full corpus and all eligible components; the figure displays only the largest connected component (17 nodes) for readability, with a deterministic spring layout (seed=42). Larger nodes represent people mentioned in more documents, and wider edges represent higher co-document counts. Every node in the displayed component is labelled, and a size legend keys node area to three real document counts.

By weighted degree, the most connected eligible surfaces are Prabowo Subianto (1,280, 3,675 documents), Sony Sonjaya (1,275, 810 documents), Dadan Hindayana (1,054, 1,497 documents), Nanik Sudaryati Deyang (890, 1,058 documents), and Asep Yusuf Somantri (417, 244 documents). The strongest single edge is Dadan Hindayana and Prabowo Subianto with 576 co-mentioned documents (Jaccard 0.13). Other top edges: Nanik Sudaryati Deyang and Prabowo Subianto (481), Dadan Hindayana and Sony Sonjaya (235), Prabowo Subianto and Sony Sonjaya (223), and Asep Yusuf Somantri and Sony Sonjaya (192). Names remain provisional NER surfaces: there is no general co-reference resolution, and ambiguous identities may still split or merge. Audited Purbaya/Yudhi Sadewa fragments are canonicalized to Purbaya Yudhi Sadewa; ambiguous bare Yusuf remains unresolved and is excluded by the two-token precision gate. No article-level evidence is published.

Interpretation boundary. Co-mention indicates only shared coverage within the same retrieved document, not a personal relationship, influence, endorsement, coordination, political alignment, or causality.

Collection Limitations

• Retrieval coverage. Only registry sources are searched. The stable search-capable subset is 68 of 75 entries (see the resolver breakdown in Collection Command), and 52 sources contributed documents retained after cleaning.
• Keyword recall. Retrieval uses six related queries: mbg, makan bergizi gratis, program MBG, satuan pelayanan pemenuhan gizi, SPPG, and badan gizi nasional. Articles that discuss implementation without any of those terms can still be missed.
• Completeness. A finished corpus is bounded by what each scraper’s search endpoint exposes. Some sources cap depth, return only top-N, or paginate inconsistently; re-running with a tighter window or source-by-source will not necessarily close those gaps.
• Copyright. Each row is a news article record: title, link, publish date, author, and a content excerpt as returned by the source’s own feed/page. The corpus supports aggregate analysis, citation (see How to Cite), and downstream modeling within fair-use research bounds; it is not a redistribution of full article text. Honor each publisher’s terms.

How to Cite

This guide. Cite the guide itself as part of the news-watch documentation:

@misc{mabruri_usecase_mbg_2026,
  author       = {Okky Mabruri},
  title        = {Use Case MBG: Collecting and Analyzing Indonesian News
                  on Makan Bergizi Gratis},
  year         = {2026},
  howpublished = {news-watch documentation, v1.2.0},
  doi          = {10.5281/zenodo.14908389},
  url          = {https://github.com/okkymabruri/news-watch/blob/main/docs/use-case-mbg.md}
}

Software. Cite news-watch using the repository’s CITATION.cff (v1.2.0, DOI 10.5281/zenodo.14908389):

@software{mabruri_newswatch,
  author = {Okky Mabruri},
  title = {news-watch},
  year = {2025},
  doi = {10.5281/zenodo.14908389}
}

Corpus and figures. Cite aggregate figures as products of this guide’s reference run, for example:

MBG news corpus (2025-01-05 to 2026-07-31), collected with news-watch v1.2.0; aggregate figures only. https://github.com/okkymabruri/news-watch

Article-level records are not redistributed, so cite counts and figures, not private document tables.

News articles. When quoting an individual article (including the linked ANTARA and Kompas reports above), cite the publisher, article title, publication date, and URL; the corpus record is a pointer, not the source of record.

Synced from okkymabruri/news-watch at main@f1c8fbf9 on 2026-08-17. These docs are ahead of the v1.2.5 release.