Bharat Tracker

A source-cited record of Indian defence, infrastructure, trade and manufacturing since 2001 — with an honest assessment of how good it actually is, measured against global benchmarks rather than against India's own past.

The core rule: every number rendered on this site must be traceable to a named source with a URL and a verification date. This is enforced mechanically — npm run validate fails the build if any series lacks a definition, a unit, a resolvable source, or a confidence grade.

---

Quick start

npm ci
npm run dev            # http://localhost:3000

No environment variables are needed. The site ships with committed data and works offline.

npm test               # validate data + audit registry + schema tests + typecheck
npm run etl            # refresh World Bank series and the news tracker
npm run etl:dry        # show what would be fetched, touch no network
npm run db:check       # is the bharat_tracker schema reachable?
npm run db:push        # seed Postgres from the committed JSON
npm run db:pull        # hydrate the JSON from Postgres (runs before every build)

Getting it live on Vercel (free tier)

1. Push this branch and open https://vercel.com/new.
2. Import chad3456/india-defence-infra-corridor.
3. Framework preset auto-detects as Next.js. Leave the defaults — vercel.json already sets the build command so the data-validation gate runs before every deploy.
4. Click Deploy. Nothing else is required; there are no mandatory env vars.

Region is pinned to bom1 (Mumbai) in vercel.json, which is closest to the audience.

Optional: the Postgres backend

The site works with no database at all — chart data is committed as JSON and the tracker falls back to data/live/news.json. Connecting Postgres makes it the store of record and lets the news tracker refresh continuously instead of on rebuild.

Everything lives in a dedicated bharat_tracker schema. Nothing is created in public, no public object is altered, and grants are scoped to that schema — so this is safe in a Supabase project shared with other projects. drop schema bharat_tracker cascade removes this project completely and touches nothing else. The schema tests assert the isolation (npm run test:schema).

Getting data into the schema

Two ways. Both are idempotent and produce the same result: 41 sources, 114 series, 1,742 data points, 434 peer values, 20 events.

A. Paste the SQL (no setup at all). Open supabase/seed.sql, copy it, paste into the Supabase SQL Editor, run. Apply supabase/migrations/0001_init.sql first if the schema does not exist yet.

B. One secret, then a button. Add SUPABASE_DB_URL under Settings → Secrets and variables → Actions, then Actions → Provision database → Run workflow. It applies the schema, seeds it, and prints the row counts in the run summary.

SUPABASE_DB_URL is Supabase → Settings → Database → Connection string (URI), with your real password substituted for [YOUR-PASSWORD]. If Supabase was connected through the Vercel integration, POSTGRES_URL_NON_POOLING is the same value.

Seeding runs over plain Postgres rather than the REST API, so it needs nothing else and does not depend on the schema being exposed to PostgREST first.

supabase/seed.sql is committed so it can be read and pasted, and CI fails if it drifts from the JSON it is generated from (npm run db:seed-check).

Optional: let the app read from Postgres

Without these the site renders from the committed JSON, which is complete. With them, the news tracker and the events map refresh without a redeploy.

Secret / env var	Where to find it
NEXT_PUBLIC_SUPABASE_URL	Settings → API → Project URL
NEXT_PUBLIC_SUPABASE_ANON_KEY	Settings → API → anon / publishable key
SUPABASE_SERVICE_ROLE_KEY	Settings → API → service role key — repo secret only, never the browser

Set them in Vercel → Settings → Environment Variables for the app, and as repository secrets for the scheduled pipeline. The Vercel Supabase integration injects SUPABASE_URL / SUPABASE_ANON_KEY automatically and the app reads either naming.

Reading over the API also needs bharat_tracker in Settings → API → Exposed schemas. The migration tries to add it itself and warns if it lacks the rights.

Locally: copy .env.example to .env.local, then npm run db:check / db:push / db:pull. npm run db:seed-sql prints the same SQL the workflow runs, if you would rather paste it into the SQL editor.

---

How it is put together

data/
  sources.json            # the source register — every publisher, URL, tier, access date
  series/*.json           # chart data, one file per domain; ETL owns wdi.json
  geo/                    # India basemap + corridor geometry
  live/                   # pipeline output: news.json, last-run.json
lib/
  types.ts                # the data model, and the rules encoded in it
  data.ts                 # loaders
  transforms.ts           # yoy / index / cagr / share — all labelled as derived
  registry.ts             # generates the chart specs
  wdi-catalogue.ts        # 66 World Bank indicators the ETL tracks
  assessment.ts           # the graded verdicts
scripts/
  validate.ts             # the data gate — runs before every build
  audit-registry.ts       # fails if the gallery drops below its target size
  etl/                    # connectors + orchestrator

Charts are specs, not components

570 charts are generated from 117 series × a set of transforms, rendered by one SVG chart component. Two rules keep the gallery honest:

1. A spec is only emitted when the transform is meaningful for that series — no year-on-year chart drawn across non-adjacent periods, no index chart off a zero base.
2. Specs for series the ETL has not yet populated are still emitted, flagged pending, and render an explicit awaiting data state rather than an empty axis pretending to be a finding.

At the time of writing all 570 render on committed data; see docs/data-sources.mdx for where each series comes from.

The data model encodes the editorial rules

scripts/lib/validate-series.ts refuses to publish a series that breaks any of these:

• no sources → rejected
• confidence: "low" without a note explaining the uncertainty → rejected
• provenance: "think-tank" graded high confidence → rejected (estimates are not records)
• backed only by tier-3 press sources but graded above low → warned

Gaps stay gaps. Where a figure could not be located the period is stored as null and the chart shows a break in the line. Nothing is interpolated or back-filled.

The pipeline

Three connectors, all fail-soft:

• World Bank — full 2001-present history for 76 indicators across India and five comparator countries. Rewritten in full each run, never appended to, because the World Bank revises history.
• Ingest — 27 RSS feeds from PIB, PMO, The Hindu, NDTV, India Today, Swarajya, OpIndia, Economic Times, Business Standard, BusinessLine and Rest of World. Headlines feed the tracker; those that pass the sector, action and geo gates become map pins.
• X / official handles — dormant unless X_BEARER_TOKEN is set; X has no free read tier.

Nothing from the news connector ever becomes chart data. Press reports of government figures are tier-3 evidence: they flag that a number may have moved, and a human verifies against the primary release before any series changes.

A connector that fails leaves the previous committed data in place and marks the run partial, so the site keeps serving last-known-good values. Output is written only after validation passes.

.github/workflows/pipeline.yml runs this every 6 hours — and on any push touching the ingest code — committing only if data changed. docs/data-sources.mdx documents every provider, tier and classification gate.

---

Known limitations

Read /methodology on the site for the full list. The big ones:

• Nominal rupees. Indian financial series are stored as published. A large share of the apparent growth in defence exports, production and budget is inflation and depreciation.
• Definitional drift. Airport counts vary between releases; highway network length grows partly by re-designating existing state roads.
• Estimates presented as data. Nuclear warhead counts are external inferences — India publishes nothing. Graded low confidence; order-of-magnitude only.
• Schematic map geometry. Corridor lines join real geocoded endpoint cities in straight segments. They are not surveyed alignments. The opening years are the load-bearing fact.
• The assessment page is editorial. Grades are judgements, not computed scores, and say so. Each ships with the strongest argument against itself. The data underneath is separable — you can reject every grade and still use every number.

Licence and attribution

Underlying data is public-domain government and multilateral output; each figure links to its publisher. Not affiliated with any government body, party or contractor.