Monsoon Ledger
AI governance from the archipelago, promises checked against the record.
Catatan janji dan kenyataan tata kelola AI di Nusantara.
Live site: https://monsoonledger.vercel.app
What this is
Monsoon Ledger is a public, bilingual (English and Bahasa Indonesia) tracker of how Indonesia governs artificial intelligence. Every entry is tied to primary or official sources, labelled with how strong the evidence is, and compared with the EU AI Act and frontier AI safety practice. ASEAN neighbours come later, once there are twelve Indonesian entries. The scanner already watches them.
The monsoon is forecast and announced long before it arrives. Indonesia's AI rules have worked the same way. The ledger records what was promised and what actually showed up.
It is an independent project by Stephen Behren Tanjung. It is not legal advice.
Why it exists
Indonesia's AI policy is announced, debated and missed in Bahasa Indonesia. Ministers give deadlines in interviews, drafts are described in official statements, and consultations are reported by Indonesian outlets. Very little of that detail reaches readers who do not read Indonesian. I read Bahasa Indonesia natively and I live under EU rules, so I read the original and set it beside the European rules for comparison.
From outside, an announced rule, a drafted rule, a signed rule and a rule in force can look like the same thing. The ledger dates every promise and links every source, so the difference stays visible. The About page sets this out as a short research abstract: background, problem, objectives, scope, contribution.
What is on the site
| Page | What it does |
|---|
| Home | The oldest missed promise and how many days past its deadline it is, the headline counts, the Safety lens matrix, the latest entries and the Radar. |
| Entries | Every entry, newest first, with filters for status, evidence level and safety topic, and an overview of dates, evidence strength and source languages that follows the filters. |
| Promise vs Reality | Janji vs Kenyataan: every public promise officials have made, its computed outcome, what happened next, and a replay slider that rebuilds the ledger date by date. Each row links to its receipt. |
| Receipts | One permanent page per promise: the wording in Bahasa Indonesia and English, the deadline and how it was derived, the computed outcome and day count, the evidence, the source and its archived copy, citations and a share card. |
| Instruments | An interactive map of every law, draft and guide, where each sits in Indonesia's legal hierarchy, and a dossier for each with an EU comparison and numbered references. |
| Analysis | Longer written work. Analysis 01, Nobody to Call, with numbered references, archived sources, a figure computed from the ledger, a Word download, citation formats and a two-page brief. Open questions and the planned Analysis 02. |
| Safety lens | Seven questions asked of every entry, the definitions behind each answer, what would change each row, and how the questions relate to threat models. |
|
Every page exists in English (/en/) and Bahasa Indonesia (/id/). The build checks this. The Bahasa Indonesia text of Analysis 01 is a draft translation of its key sections, marked as such on the page.
Recurring features
- Janji vs Kenyataan (Promise vs Reality). Every public deadline officials have given, set against what happened next, with a missed, open or met stamp.
- A Safety lens on every entry. The same seven questions each time: does it require incident reporting, does it name unacceptable uses, does it address general-purpose models, how strongly is it enforced, who receives incident reports and does that body exist, does it address incidents that cross borders, and does it address frontier AI. Yes means a cited source states it; no needs the full text or a source that says so; unclear is the default.
- Indonesian first. Key quotes stay in the original Indonesian with an English line under them, and Indonesian-language sources come first.
How to read an entry
Each entry shows, at the top, its status (announced, draft, signed, in force, withdrawn), its evidence level and the last checked date, which is the last day I re-opened the sources. Then come a summary, an optional quote, the Safety lens panel, a comparison with the EU, my own take, and the full list of sources.
Evidence levels
| Level | Meaning |
|---|
| Legal text | I have read the signed or published text of the instrument itself. The strongest level. |
| Official statement | A minister, official or government body said it on the record, as reported by a named outlet or an official channel. It tells you what the government says it intends, not what the law says. |
| Press report | The claim rests on journalism or secondary analysis, such as a law firm summary, without the primary text. Provisional until the primary text is read. |
When the legal text becomes available, I read it and upgrade the evidence level.
Promises and receipts
data/deadlines.yml records every public promise about Indonesia's AI rules in the Open Promise Ledger format (below). Each row has a claim type:
- dated: it has a deadline that can pass;
- conditional: it depends on an event, such as a signature, and has no date;
- status_claim: it reports progress, such as "only the signature is left", and has no deadline.
For dated rows, deadline_rule records how the deadline follows from the wording (for example end 2025-06), and deadline_basis_en and deadline_basis_id quote the wording and explain the reading. Where the wording allows more than one date, the latest plausible one is used, so every "missed" can be defended.
The outcome is never written by hand. scripts/promises.mjs computes it on the build day, in Western Indonesian Time: met if kept on or before the deadline, missed if the deadline passed without it (or it was kept late), open otherwise. Conditional promises and status claims are never counted as missed. Headline counts are taken among dated promises only, for example "4 of 5 dated promises missed", and every count says which day it is as of. The validator, the ontology, every page, the RSS feeds and the share cards all call the same module.
Every promise has a receipt at /en/promises/<key>/ and /id/promises/<key>/, where the key is the date said and the speaker, the same key as its ontology object. A receipt shows the promise, the speaker and role, the wording in Bahasa Indonesia then English, the deadline and its basis, the computed outcome with the days past or left, the evidence level, the source and its archived copy, and citations in plain text, APA and BibTeX with the data version. It also lists the entries and instruments the promise connects to. Each receipt has its own 1200 by 630 preview card in each language, showing the outcome, the day count and the day it was made, so a shared card is accurate on its own.
The Open Promise Ledger format
The promise format is published so that other ledgers can use it and anyone can check the counts. Version 0.1 is a draft.
- docs/standard/v0.1/README.md (Bahasa Indonesia: README.id.md) sets out the purpose, the scope, the object types, every field, the claim types, the deadline rules with worked examples, the outcome computation, the evidence levels, source preservation, what the format does not do, honest mappings to schema.org and ELI, related work and versioning.
- promise.schema.json is the JSON Schema for one promise row. The validator checks
data/deadlines.yml against this same file, so the spec and the data cannot drift apart. The site serves it at /standard/v0.1/promise.schema.json.
- examples/ holds a dated and a conditional row copied from the Indonesian ledger, also checked by the validator.
- adopters.yml lists ledgers using the format. It starts with this one. Open an issue to add another.
Start a ledger for another country
kit/ is a starting point: a data skeleton (deadlines.yml with one annotated example row, instruments.yml, and a sources.yml watch list) and a guide, Start a promise ledger for your country, in English and Bahasa Indonesia. Check a ledger folder with the same validator:
npm run validate -- --data kit/data
The format, the schema, the outcome module and the validator are reusable today. The website is built for Indonesia and is not yet reusable without real work.
Preserving evidence
News pages move, change and disappear. scripts/archive.mjs (npm run archive) gives every source url in the promises, the instruments and the entries a preserved copy in the Internet Archive. It uses an existing snapshot when there is one and otherwise asks Save Page Now, at most eight times a run with a pause between requests. It writes only archive_url and archived_at beside each source, editing lines so comments and folded text stay as they were. It runs after every scheduled scan and cannot block it. The validator warns, without failing, about sources still without a copy, and every page that shows a source links its archived copy.
The ontology layer
The ontology turns the ledger into one connected model. It has nine object types: instrument, entry, promise, official, institution, document, outlet, reference and radar lead. They are joined by sixteen named link types, such as concerns, cites, made, passed by, drafted by and implements.
- Most links are derived, not written. They are read straight from the data files at build time. An entry and a promise that cite the same report meet at one document object. That shared document is where the ledger fuses.
- Curated links live in
data/ontology.yml. Each one names its source and an evidence level. The validator rejects a curated link that points at something missing, joins the wrong object types, or cites a source the ledger does not already cite.
- Nothing is inferred. Officials are listed exactly as the ledger lists them. Promises are never tied to an instrument by guesswork. Radar leads appear only as machine hints. Promise objects link to their receipts and carry their computed outcome.
- Standards. Where a public standard defines the same relation, it is named, for example
schema:legislationPassedBy and eli:passed_by. The Instruments page publishes schema.org Legislation records built from the same data.
- Open data. The full graph is published at /ontology.json.
The fusion logic is one pure module, scripts/ontology.mjs. The validator and the site both use it, so they always see the same graph.
Automatic monitoring (the Radar)
Every Monday and Thursday at 06:17 WIB a GitHub Actions workflow runs npm run scan. It reads Indonesian and ASEAN news searches and official legal databases, and keeps headlines that score above a keyword threshold. It drops anything already seen or already in the ledger, and writes new leads to data/radar.yml. The site shows them on the Radar as unreviewed leads.
The scanner only ever writes data/radar.yml, data/scan-log.yml and data/watch/seen.json. It never creates or edits entries, instruments, promises or ontology links. The archive step that runs after it only adds archive_url and archived_at beside sources. A lead becomes an entry only after the author reads the source.
If the repository has an ANTHROPIC_API_KEY secret, the scanner also asks Claude for a second opinion on each new headline. The answer comes back as structured output and is checked before use. Without the key it runs on keyword rules alone. Claude's summaries restate headlines and are labelled as machine summaries on the site.
Visual layer and motion
The site is a static Astro build with small vanilla scripts. There is no client framework, no tracker and no external font. The interactive pieces are:
- the receipt headline on the home and ledger pages, the balance of dated promises and the replay slider;
- the receipts, with day counts brought up to the reader's day;
- the entries overview;
- the ontology console (graph, timeline and signals, with full screen), the scan history chart and the next-scan countdown;
- a right-hand rail on long pages, with reading progress, the current section and a page visual;
- counting figures and slowly drifting isobars behind every page header.
Every figure comes from the data files. The seven deadlines appear in two visual forms, the balance bar and the replay chart, plus the receipts. Animations pause when off screen or in a background tab, and stop entirely for readers who ask their system for reduced motion.
Links, icons and previews
The site ships an ICO and SVG favicon, an Apple touch icon, a web manifest and 1200 by 630 preview cards in English and Bahasa Indonesia: one for the site and one for every receipt. Large-image Open Graph and Twitter tags are on every page, including the root address people usually paste into chat apps, so a shared link shows the card and not a blank box. Tab titles are the page name followed by "Monsoon Ledger".
scripts/make-share-images.mjs draws every card as SVG and renders it to PNG with @resvg/resvg-js during the site build, using the Inter font files in scripts/fonts/ (SIL Open Font License, used for the images only, never loaded by the site). Running it directly regenerates the icons.
Analysis papers
Papers are Markdown files in writing/analysis/ (English) and writing/analisis/ (Bahasa Indonesia), with their numbered sources in data/analysis/<id>.sources.yml. The site renders them at build time with no client framework. Every [n] links to its reference and back; every direct quote is marked so the validator can enforce the quote rules; figures and tables built from data are computed as of the paper's facts day; numbers such as day counts are read from the data with tokens, so the text cannot drift from it. Changes after publication are listed on the paper from CHANGELOG.md lines tagged [analysis-01].
npm run check-quotes reads every quoted source, or its archived copy, and looks for the quote's exact words. Results are in data/quote-check.yml, and the Method page shows how many were found and when.
To cite the project, use CITATION.cff (GitHub shows a "Cite this repository" button) or the "Cite this" box on each paper. There is no DOI yet; docs/releasing.md explains how to get one through Zenodo.
Unpublished drafts, such as cross-posts, summaries and outlines, stay in writing/ and are never linked from the site.
Frontier incident reference and reading list
data/frontier-incidents.yml lists public reports of frontier AI models or agents acting on, or used against, systems outside the developer's control. Dates appear only as sources give them, and the day counts (access to notice, awareness to notice, notice to public) are computed by scripts/frontier.mjs. It is context for the ledger and for Analysis 01, never part of the ledger's counts, and it is not complete.
data/reading.yml is the reading list behind the analysis papers. An item marked read needs a finish date and the author's notes in both languages.
Quality gates
Every push and pull request runs, in this order:
npm test: the scanner tests on synthetic fixtures, and the tests for deadline derivation and outcome computation (leap years, month ends, the deadline day itself, conditional and status claims, the time zone), the archive script with the network mocked, the ontology, the analysis papers and their renderer, the quote check against fixture pages, the safety lens rules, the frontier incident counts, the reading list and the language parity check.npm run validate: the entry schema and the published promise schema, plus the rules a schema cannot express. These cover ids and file names, instruments that exist, quote rules, every deadline recomputed from its rule, the format's worked examples, the radar and scan log, ontology links and their sources, the build log, and a ban on em and en dashes anywhere in the project. Sources without an archived copy are reported as warnings, never failures.npm run build: the site build, which runs the validator again first, then builds, then runs the parity check (npm run parity): every page must exist in both languages unless marked partial, and every _en field in the data needs its _id twin.
The validator also covers the analysis papers (citations, sources, quotes, front matter), the quote check results, the safety lens rules, the frontier incident reference and the reading list. npm run links checks every internal link and anchor in the built site, then reports external sources that stopped responding.
npm run validate -- --data <folder> runs the same validator on a promise ledger's data folder only, such as the kit.
Build log
Every change the author asks for gets a row in data/build-log.yml, including fixes, reverts and rollbacks. The rows appear under "Build so far" on the Technical system page, each with its pull request and commits. The validator checks that rows are dated, in order, of a known kind and linked to real commit hashes.
Working on the ledger
One-time setup, with Node 22.23.4 or newer:
npm run setup
Adding an entry (weekly):
npm run new-entry -- --slug id-example-topic # creates data/entries/<today>-id-example-topic.yml
# fill in every field in the new file
npm run validate # must pass
git add data && git commit -m "Add entry: example topic" && git push
Other routine work:
- A new promise. Add a row to
data/deadlines.yml, newest first, with its claim type and, if dated, its deadline_rule, deadline, deadline_precision and the wording behind it. Never write the outcome.
- Something happens to a promise. Set
outcome_date and outcome_source_url, update the one-line outcome text, and add a dated line to CHANGELOG.md. The outcome itself is computed.
- Reviewing a radar lead. Set its
status to reviewed, dismissed or promoted. For promoted, also set promoted_to: <entry id>, and only after writing a full entry from the source.
- Adding an ontology link. Add a row to
data/ontology.yml with from, rel, to, evidence and a source_url the ledger already cites.
- Any change to the site. Add a row to
data/build-log.yml.
Useful commands:
npm run dev # local preview
npm test # every unit test, no network
npm run validate -- --data kit/data # check a promise ledger folder only
npm run scan -- --dry-run --no-claude # live scan, writes nothing
npm run archive -- --dry-run # check archive coverage, writes nothing
npm run check-quotes # checks each quote against its source, writes data/quote-check.yml
npm run parity # English and Bahasa Indonesia parity, after a build
npm run links # internal links in site/dist, then external sources (informational)
node scripts/make-share-images.mjs # regenerate the site icons
Pushing to main rebuilds the site on Vercel.
Project layout
data/
entries/ one YAML file per entry (the ledger itself)
schema.json entry schema (JSON Schema 2020-12)
instruments.yml laws, drafts and guidance, with dossier fields
deadlines.yml the Janji vs Kenyataan ledger, newest first, in the promise format
ontology.yml institutions and curated, sourced links
analysis/ numbered sources for each analysis paper
frontier-incidents.yml (+ .schema.json) the frontier incident reference
reading.yml (+ .schema.json) the reading list
quote-check.yml results of npm run check-quotes
build-log.yml every change asked for, newest first
radar.yml unreviewed leads written by the scanner
scan-log.yml one row per scan
watch/ what the scanner reads (sources.yml), how it scores (keywords.yml), what it has seen
docs/standard/ the Open Promise Ledger format: spec, JSON Schema, examples, adopters
kit/ a starting point for a promise ledger in another country
scripts/
validate.mjs schemas and project rules, run before every build
promises.mjs deadline rules and computed outcomes, shared by everything
archive.mjs preserved copies of sources in the Internet Archive
ontology.mjs the fusion layer shared by the validator and the site
analysis.mjs, analysis-render.mjs analysis papers: rules, and the HTML steps the site uses
quotes.mjs, check-quotes.mjs the quote check
lens.mjs the safety lens rules
frontier.mjs computed incident day counts and paper tokens
reading.mjs reading list rules
check-parity.mjs English and Bahasa Indonesia parity
scan.mjs, scan/ the monitoring layer and its tests
new-entry.mjs creates a dated entry from a template
check-links.mjs internal links in the built site, and dead source links
make-share-images.mjs link preview cards and icons, with fonts/ (OFL)
*.test.mjs unit tests, run with npm test
site/ the Astro static site (pages, components, styles, public assets)
writing/ analysis papers (analysis/ in English, analisis/ in Bahasa Indonesia) and unpublished drafts
docs/releasing.md how to archive a release on Zenodo and get a DOI
CITATION.cff, .zenodo.json citation metadata
.github/workflows/ ci.yml (tests, validation, build) and scan.yml (the scheduled scan)
.github/ISSUE_TEMPLATE/ forms for corrections, sources and promises
Corrections
Anyone can report a correction with a source. Errors are fixed in the open, logged with a date in CHANGELOG.md, and listed on the Corrections page. See CONTRIBUTING.md for sources and promises, CODE_OF_CONDUCT.md, and SECURITY.md for security problems.
How AI is used here
I research the sources, write and edit every entry and every My take myself, and set every evidence level. Claude Code, Anthropic's coding agent, writes the code and configuration and drafts text and translations for me to review. It also checks URLs and legal citations before they are used, and runs the tests, the validator and the build before every push. The rules it works under are in CLAUDE.md: never invent facts, quotes, dates or URLs, leave a marked TODO instead, never rewrite the author's words, and pass the validator. Every source is linked so you can check it yourself.
Licence
Code: MIT, see LICENSE. Data and writing: CC BY 4.0, see LICENSE-CONTENT.md.
Dalam bahasa Indonesia. Monsoon Ledger adalah pelacak publik dalam dua bahasa tentang cara Indonesia mengatur kecerdasan artifisial. Setiap catatan bersandar pada sumber primer atau resmi, diberi label seberapa kuat buktinya, lalu dibandingkan dengan EU AI Act dan praktik keselamatan AI terdepan. Bagian Janji vs Kenyataan mencatat setiap tenggat yang pernah diumumkan pejabat dan apa yang benar-benar terjadi. Lapisan ontologi menghubungkan catatan, instrumen, janji, pejabat, lembaga, dan dokumen dalam satu grafik, dan Radar memantau berita dua kali seminggu tanpa pernah menulis catatan sendiri. Setiap perubahan pada situs dicatat di log pembangunan. Kalau Anda menemukan kesalahan, silakan buka issue dengan menyertakan sumbernya.
