75 citations were being counted as zero, and all 15 affected studies are South Asia

[Varnasr]

The bug

countReferences() in scripts/build.mjs matched:

/## (?:References|Sources|Bibliography)\s*\n.../i

189 of the 204 studies head their citations ## References. The 15 written to the later, longer template head them ## Further Reading. The function returned 0 for all fifteen.

They are not missing citations. Each carries five, properly formatted:

1. Imbert, Clement, and John Papp. "Labor Market Effects of Social Programs: Evidence from India's Employment Guarantee." American Economic Journal: Applied Economics 7, no. 2 (2015): 233-263.
2. Dreze, Jean, and Reetika Khera. "Recent Social Security Initiatives in India." World Development 98 (2017): 555-572.
3. Sukhtankar, Sandip. "India's National Rural Employment Guarantee Scheme: What Do We Really Know?" India Policy Forum 2016-17.
4. Narayanan, Sudha, and Upasak Das. "Women Participation and Rationing in the Employment Guarantee Scheme." Economic & Political Weekly 49, no. 46 (2014).

The library reported 867 references where it holds 942.

The affected set is not random

Six India studies (MGNREGA, UPI, DBT/JAM, Kerala model, PLI, Swachh Bharat), two Bangladesh (BRAC, garments), two Nepal (community forestry, remittances), Pakistan (BISP), Sri Lanka (2022 crisis), Bhutan (GNH), the Maldives (climate), and a regional stunting study.

That is essentially the whole South Asia set, reading as uncited on a site whose README calls it an evidence library.

The warning pointed the wrong way

npm run validate printed, for each of them:

⚠ india-mgnrega.json: no parseable References section

That reads as a content gap. It was a heading the regex did not know. Acting on the warning as written would have meant adding citations to studies that already have them. This is why I checked one of the files before touching the parser.

Verified across all 204: no study carries both headings, so widening the regex counts nothing twice.

Result

	before	after
total references	867	942
studies with 0 refs	15	0
⚠ warnings on --check	15	0

All 15 move from refs: 0 to refs: 5. master-list.json, stats.json and sitemap.xml regenerated.

One finding recorded, not patched

Fixing the count exposes something about the evidence field that is worth stating plainly.

evidenceTier() returns strong when refs >= 4. The reference distribution across all 204 studies is:

  4 refs: 78 studies
  5 refs: 126 studies

Nothing has fewer than four. So after this fix the function returns strong for all 204, and the moderate and emerging branches are unreachable. The hasRCT and hasMeta tests are short-circuited by the refs >= 4 branch before they matter.

Before the fix, 8 studies read emerging. All 8 were among the 15 whose references were not being counted. The tier was reporting the parser bug, not the evidence.

I have not changed it. What counts as strong evidence is an editorial judgement about this library, not a bug to patch on someone's behalf. The honest options:

1. Raise the threshold above the range the corpus actually occupies.
2. Weight hasRCT / hasMeta rather than letting refs >= 4 short-circuit them.
3. Drop the field, and stop implying a distinction the data does not carry.

CLAUDE.md states the position and the options so it is not rediscovered as a surprise.

Also in this PR

The link check moves out of ci.yml into its own daily link-check.yml. Two reasons: it carried fail: false on pull requests, so a contributor got a result they could neither act on nor be blocked by; and it shared a workflow run with the data validation, where a job cancelled by its own timeout-minutes cancels the whole run — a slow third-party host could have taken the validation down with it. On the schedule it fails, because external link rot needs no commit to happen.

Adds CLAUDE.md, which this repository did not have. It records the two-template split, the sitemap's query-string URLs (a path-based comparison reports all 204 as missing pages), and that the absent lockfile is correct: package.json declares no dependencies at all.

Checks

$ npm run validate
Validating case-study data…

✓ 204 studies processed cleanly (check mode).

No warnings, where there were 15.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MrvR2NXsJFVRCeJZFCPuNL

---

Generated by Claude Code