Methodology

Hand-authored · written 5 August 2026 · describes the build as it stands on that date

A reference site asks for trust on every page, so it owes you one page that explains why the trust is warranted — where the numbers come from, what the rules are when data is missing or wrong, and what happened to the thousands of pages this estate used to be. This is that page. Nothing here is aspiration; each mechanism described below is running in the build today, and where the mechanism has a file name, the file name is given.

The data spine

Every figure on a directory page is read from one place: the data spine under _data/, a set of JSON shards built deterministically from the 347 source files pulled from the legacy server in July 2026. A build script (_data/build_spine.py) generates the shards; a separate verifier (_data/verify_spine.py) checks them; and the spine’s own manifest states the contract in one sentence we hold ourselves to: “Every count below is computed from those source files… at build time — none is typed by hand. If a future page shows a number, it should read it from here.” When you see “568 ports” or “273 agreements” on an intel directory, that number was counted by a program from named inputs, and the page footer carries the matching caveat: Compiled reference — verify current specifics at the source.

The four honesty rules

The spine ships with rules, written into its manifest, and the pages inherit them.

1 · No invented data. A page family with no backing record gets a stated zero and a reason — never a guess, never padding borrowed from a different source. You can see the rule working on the country atlas: the legacy registry’s header claims 198 countries, the computed truth is 161, and the page publishes 161 with the arithmetic explained. Same on the FTAs directory, where 37 of 273 records ship as bare slugs with empty names — they are listed bare, because an empty field reported is information and an empty field filled is fiction.

2 · No stale prices. Cached live quotes in the legacy data are excluded from the spine entirely. Anything market-priced on this estate — the market desk — draws on institutional series (European Central Bank, World Bank) and says so next to the figure.

3 · Declared is not actual. The old site’s own headline claims (“384,000 pages”, “526 FTAs”) are kept in a separate bucket, cross-checked against computed counts, and never substituted for them. Where a claim and a count disagree, the count wins and the disagreement is disclosed on the page that uses it.

4 · Every source file accounted for. All 347 source JSON files are classified exactly once — 279 used, 48 used for reconciliation, 20 excluded with a written reason. Nothing is quietly dropped; nothing is double-counted.

Dated provenance, everywhere

Dates are the cheapest honesty there is, so the estate spends them freely. Directory pages carry a “verified” date under the title — the day the table was last rebuilt against the spine. Hand-written pages carry a “written” date. External facts in prose cite the institution and an accessed date inline — the convention looks like “(UNCTAD; accessed 5 August 2026)” — and the sources registry lists the 140 institutions the intelligence layer treats as citable, typed by what kind of authority each one is. Two footer lines tell you which regime a page lives under: Compiled reference means the content is read from the data layer; Hand-authored guidance means a person wrote it and checked the external claims on the stated date. Either way the instruction is the same — verify current specifics at the source — because a library’s job is to get you to the primary document faster, not to replace it.

The consolidation: from thin pages to directories

The estate’s previous incarnation made the classic programmatic-SEO trade: hundreds of thousands of generated URLs, most of them thin. An external audit of the live site on 4 August 2026 found the predictable failure modes — headline sections that 404’d, country and FTA deep-links that all rendered the same generic fallback page, and page counts that disagreed with each other from one page to the next. The rebuild reversed the trade. What used to be thousands of single-entity pages — one URL per port, per zone, per agreement — became eleven consolidated directories: one page per family, every entity a sortable row, every table backed by a named shard. 568 port pages became one ports table worth reading; the pattern repeats for SEZs, HS chapters, institutions, trade bodies, cities, corridors, business studies, agreements, economies and sources. Legacy URLs were not abandoned: the edge worker carries permanent 301 redirects from the old paths into the consolidated pages, so a decade of external links keeps resolving. And the standing rule going forward is the simple one: no thin pages. A page ships when it has a data shard behind it or considered writing on it — ideally both — and not otherwise.

What “hand-authored” means here

It means a person chose the words for the page you are reading — no template filled per entity, no generated prose wearing a byline. Hand-authored pages on this estate include the narrative family (thirty-two long-form pages, mapped here), the orientation and role paths, the deep-dive profiles inside the directories, and this page. Hand-authored does not mean unsourced: it means the sourcing is visible in the sentence instead of implied by the template.

Corrections

The failure mode of any reference is silent wrongness, and the antidote is a reader with a sharper eye than the build. If a figure here disagrees with the primary source, the primary source is right and we want to know: write to Amit from the front page, name the page and the number, and the fix lands in the next build with the verification date moved forward. That loop — shard, page, date, reader, correction — is the whole methodology. Everything above is just its moving parts.