#!/usr/bin/env node
/**
 * publish-data.mjs · MAKE THE PAGE'S OWN CLAIM TRUE.
 *
 * ═══════════════════════════════════════════════════════════════════════════════════════════════
 * ⛔⛔ WHY THIS EXISTS, AND IT IS NOT A NICE-TO-HAVE.
 * The reliability study shipped live saying, in three separate places:
 *     "the data here is released under CC BY 4.0"
 *     "The scripts and raw answers are published so the method can be re-run"
 *     "Our measurements, our scripts and our raw answers are published so these findings can be
 *      checked against us"
 * None of it was published. No public URL existed, and the page did not even link to the files.
 *   ★ A FALSIFIABLE FALSE CLAIM, INSIDE THE DISCLOSURE BLOCK, ON A PAGE AUDITING MEASUREMENT
 *     HONESTY. The one page in the estate that cannot afford it.
 *   ⚠️ It was not a lie anyone wrote on purpose. The sentence was true of the INTENT and nobody
 *     checked it against the filesystem, which is the same failure mode as a design doc written in
 *     the present tense. THE FIX IS A STEP IN THE CHAIN, NOT A REMINDER.
 *
 * WHAT IT PUBLISHES (everything needed to reproduce or refute the study):
 *   data/     the frozen report JSONs + the raw per-answer store, exactly as measured
 *   scripts/  every instrument, so the method can be re-run against models that still exist
 *   bank/     the question bank the answers were produced from
 *   README.md licence, manifest, method, and the stated limits
 *
 * ⛔ REFUSES rather than publishing a partial set: a data release missing its raw answers invites
 *   exactly the "trust us" posture the study argues against.
 *
 * usage: node variance/publish-data.mjs [--write]
 */
import fs from 'node:fs';
import path from 'node:path';

const HERE = import.meta.dirname;
const AUDIT = path.resolve(HERE, '..');
/* ⚠️ TWO levels up, not three. AUDIT is <site>/research/ai-recommendation-audit-2026, so '..','..'
   lands on the SITE root and 'research' re-enters it. The first version used three and wrote a
   complete, correct-looking release into clusters/01-core/research — outside any deployed site.
   ★ THE VERIFIER SAID 14/14 FILES LANDED AND IT WAS TELLING THE TRUTH. "The file exists" and "the
     file is where it will be served from" are different claims, and only the second one ships. */
const OUT = path.resolve(AUDIT, '..', '..', 'research', 'ai-recommendation-reliability-2026');
const WRITE = process.argv.includes('--write');

const REPORTS = [/^reliability-report-T0-.*\.json$/, /^model-mortality-.*\.json$/, /^temp-probe-.*\.json$/, /^size-at-temp0-.*\.json$/];
const SCRIPTS = ['reliability-study.mjs', 'model-mortality.mjs', 'temp-probe.mjs', 'size-at-temp0.mjs', 'build-study-page.mjs', 'page-shell.mjs', 'scaffold-audit.mjs', 'publish-data.mjs'];
const STORES = ['study-answers-cats-lattice-owned-T0.jsonl'];
const BANKS = [path.join(AUDIT, 'cats-lattice-owned.json')];

const newestMatching = (re) => fs.readdirSync(HERE).filter((f) => re.test(f)).sort().pop() || null;

const reportFiles = REPORTS.map(newestMatching);
const missingReports = reportFiles.filter((f) => !f);
if (missingReports.length) { console.error(`REFUSED: ${missingReports.length} report file(s) missing. Run the instruments before publishing.`); process.exit(2); }
for (const s of STORES) if (!fs.existsSync(path.join(HERE, s))) { console.error(`REFUSED: raw answer store ${s} is missing. A release without raw answers is a "trust us", which is what this study argues against.`); process.exit(2); }

const manifest = [];
const plan = [];
const add = (src, rel) => {
  const bytes = fs.statSync(src).size;
  plan.push({ src, rel, bytes });
  manifest.push({ file: rel, bytes });
};
/* ⛔ PUBLISH IN EXTENSIONS THE EDGE ACTUALLY SERVES, VERIFIED IN PRODUCTION (2026-09-23).
   The first release copied README.md and scripts/*.mjs faithfully, the deploy passed, and both
   returned 404 live:
     · `.md` is DELIBERATELY denied by the site worker's sensitive-root guard, which exists to stop
       CLAUDE-INSTRUCTIONS.md and friends leaking. That guard is right and must not be weakened for
       a convenience.
     · `.mjs` is not in the deny list but does not serve either; `.js`, `.json` and `.jsonl` all do.
   ★ THE FILE EXISTING IN THE DEPLOY DIRECTORY IS NOT THE SAME CLAIM AS THE FILE BEING REACHABLE.
     Audit as the crawler fetches, not as the filesystem lists. A data release nobody can download
     is the same broken promise as no release at all. */
for (const f of reportFiles) add(path.join(HERE, f), `data/${f}`);
for (const f of STORES) add(path.join(HERE, f), `data/${f}`);
for (const f of SCRIPTS) { const p = path.join(HERE, f); if (fs.existsSync(p)) add(p, `scripts/${f}.txt`); }
for (const p of BANKS) if (fs.existsSync(p)) add(p, `bank/${path.basename(p)}`);

const totalKB = (plan.reduce((a, f) => a + f.bytes, 0) / 1024).toFixed(0);
console.log(`publish plan: ${plan.length} files, ${totalKB} KB -> ${path.relative(process.cwd(), OUT)}`);
for (const f of plan) console.log(`  ${String((f.bytes / 1024).toFixed(1)).padStart(8)} KB  ${f.rel}`);

/* The raw store is JSONL of every answer INCLUDING failures, because the failures are part of the
   method: a reader can see which cells were rate-limited and recompute coverage themselves. */
const storeRows = fs.readFileSync(path.join(HERE, STORES[0]), 'utf8').split('\n').filter(Boolean).length;
const rel = JSON.parse(fs.readFileSync(path.join(HERE, reportFiles[0]), 'utf8'));
const mort = JSON.parse(fs.readFileSync(path.join(HERE, reportFiles[1]), 'utf8'));

const README = `# AI recommendation reliability, 2026

Raw measurements, instruments and question bank behind
**[We audited our own AI visibility index](https://lucreya.com/articles/ai-recommendation-reliability-2026)**.

**Licence: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).** Use it, republish it, argue with it.
Attribution: Vincent Wesley Couey, Lucreya (Lattice network).

## What is here

| path | what it is |
|---|---|
| \`data/reliability-report-T0-*.json\` | self-agreement floors and between-model agreement, bootstrap CIs over questions |
| \`data/model-mortality-*.json\` | live availability probe of every model this estate has published with, incl. the positive control |
| \`data/temp-probe-*.json\` | self-agreement at temperature 0 vs 0.3 |
| \`data/size-at-temp0-*.json\` | the size effect isolated against a measured noise floor |
| \`data/study-answers-*.jsonl\` | **every raw answer**, one JSON object per line, including failed and rate-limited cells |
| \`scripts/*.mjs\` | the instruments, runnable as-is |
| \`bank/cats-lattice-owned.json\` | the question bank the answers were produced from |

${storeRows} raw answer rows. ${rel.questions} distinct questions. ${mort.roster} models probed for availability, ${mort.retired} retired.

## Reproducing it

\`\`\`
node scripts/reliability-study.mjs --plan                       # what it would do, no API calls
node scripts/reliability-study.mjs --run --questions 40 --repeats 5
node scripts/reliability-study.mjs --report
node scripts/model-mortality.mjs --write
\`\`\`

Each instrument needs a provider key on disk (\`~/.groq-api-key\`, \`~/.gemini-api-keys\`, \`~/.hf-token\`).
Models absent from your provider will be reported as retired, which is the point of the third finding.

## Stated limits

- Agreement is Jaccard overlap over canonicalised company names.
- Confidence intervals are bootstrap over **questions**, not over pairs, because pairs inside one question are not independent.
- Any model with fewer than 5 questions carrying repeats is reported as insufficient data rather than given a number.
- Several models are served through one inference host; a host-level effect would be common to them.
- "Retired" means a provider no longer serves that model id on the probe date. The weights may exist elsewhere.
- This is one estate's roster on one date. A lower bound on the problem, not a census.

## Conflict of interest

Lucreya is part of the Lattice network, which sells disclosed placements on pages that AI engines cite.
That is a commercial interest in this subject. These files are published so the findings can be
checked against us.

_Generated by \`variance/publish-data.mjs\`. Regenerated whenever the measurements are._
`;

if (!WRITE) { console.log('\nDRY RUN — pass --write'); process.exit(0); }

fs.mkdirSync(path.join(OUT, 'data'), { recursive: true });
fs.mkdirSync(path.join(OUT, 'scripts'), { recursive: true });
fs.mkdirSync(path.join(OUT, 'bank'), { recursive: true });
for (const f of plan) fs.copyFileSync(f.src, path.join(OUT, f.rel));
fs.writeFileSync(path.join(OUT, 'README.md'), README);

/* ⭐ index.html IS THE ONLY ENTRY POINT THAT RELIABLY SERVES. Markdown is denied at the edge, so a
   release whose front door is a README is a release with no front door. This page is what a reader,
   a crawler and a sceptic all actually land on, so it carries the manifest, the licence and the
   limits rather than pointing at a file that 404s. */
const esc = (t) => String(t).replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;');
const rows = manifest.map((m) => `<tr><td><a href="${m.file}">${esc(m.file)}</a></td><td class="n">${(m.bytes / 1024).toFixed(1)} KB</td></tr>`).join('\n');
const indexHtml = `<!DOCTYPE html>
<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<title>Data release: AI recommendation reliability 2026 | Lucreya</title>
<meta name="description" content="Raw answers, reports and instruments behind the AI recommendation reliability study. ${storeRows} raw answer rows, CC BY 4.0.">
<link rel="canonical" href="https://lucreya.com/research/ai-recommendation-reliability-2026/">
<link rel="stylesheet" href="/styles.css?v=9f477032">
<script type="application/ld+json">${JSON.stringify({
  '@context': 'https://schema.org', '@type': 'Dataset',
  name: 'AI recommendation reliability measurements 2026',
  description: 'Raw per-answer records, agreement reports and model-availability probes behind the AI recommendation reliability study.',
  url: 'https://lucreya.com/research/ai-recommendation-reliability-2026/',
  license: 'https://creativecommons.org/licenses/by/4.0/',
  isAccessibleForFree: true, temporalCoverage: new Date().toISOString().slice(0, 10),
  measurementTechnique: 'Repeated prompting of named models at stated temperature; Jaccard agreement over canonicalised company names; bootstrap CIs over questions; provider availability probed with a positive control.',
  creator: { '@type': 'Person', name: 'Vincent Wesley Couey', url: 'https://lucreya.com/author/vincent-couey', sameAs: ['https://orcid.org/0009-0005-6869-308X'] },
  distribution: manifest.map((m) => ({ '@type': 'DataDownload', contentUrl: `https://lucreya.com/research/ai-recommendation-reliability-2026/${m.file}`, contentSize: String(m.bytes) })),
})}</script>
<style>body{max-width:820px;margin:0 auto;padding:2rem 1rem 4rem;font:16px/1.65 Georgia,serif}
h1{font-size:1.9rem;line-height:1.15;margin:.2rem 0 .6rem}
table{width:100%;border-collapse:collapse;margin:1.2rem 0;font:14px/1.5 ui-sans-serif,system-ui}
th,td{text-align:left;padding:.5rem .6rem;border-bottom:1px solid #e3e6ec}
td.n{text-align:right;font-variant-numeric:tabular-nums;white-space:nowrap}
.lede{font-size:1.06rem;color:#5d6470}
.box{border:1px solid #e3e6ec;border-radius:10px;padding:1rem 1.15rem;margin:1.4rem 0}
code{background:#f4f5f7;padding:.1rem .3rem;border-radius:3px;font-size:.9em}</style>
</head><body>
<p><a href="/">Lucreya</a> &rsaquo; <a href="/research/">Research</a> &rsaquo; Data release</p>
<h1>AI recommendation reliability, 2026: the data</h1>
<p class="lede">Everything behind <a href="/articles/ai-recommendation-reliability-2026">We audited our own AI visibility index</a>: every raw answer, every report, every instrument. Licensed <a href="https://creativecommons.org/licenses/by/4.0/">CC BY 4.0</a>. Attribution: Vincent Wesley Couey, Lucreya.</p>
<p><strong>${storeRows}</strong> raw answer rows across <strong>${rel.questions}</strong> questions. <strong>${mort.roster}</strong> models probed for availability, <strong>${mort.retired}</strong> retired.</p>
<h2>Files</h2>
<table><thead><tr><th>file</th><th class="n">size</th></tr></thead><tbody>
${rows}
<tr><td><a href="manifest.json">manifest.json</a></td><td class="n">n/a</td></tr>
</tbody></table>
<p>Instruments are published with a <code>.txt</code> extension because the edge does not serve <code>.mjs</code>. Rename to <code>.mjs</code> to run them.</p>
<div class="box"><h2 style="margin-top:0">Stated limits</h2>
<ul>
<li>Agreement is Jaccard overlap over canonicalised company names.</li>
<li>Confidence intervals are bootstrap over <em>questions</em>, not pairs, because pairs inside one question are not independent.</li>
<li>Any model with fewer than 5 questions carrying repeats is reported as insufficient data, not given a number.</li>
<li>Several models share one inference host; a host-level effect would be common to them.</li>
<li>"Retired" means a provider no longer serves that model id on the probe date. The weights may exist elsewhere.</li>
<li>One estate's roster on one date. A lower bound, not a census.</li>
<li>The raw store includes failed and rate-limited cells on purpose, so coverage can be recomputed independently.</li>
</ul></div>
<div class="box"><strong>Conflict of interest.</strong> Lucreya is part of the Lattice network, which sells disclosed placements on pages that AI engines cite. That is a commercial interest in this subject. These files are published so the findings can be checked against us.</div>
</body></html>
`;
fs.writeFileSync(path.join(OUT, 'index.html'), indexHtml);
fs.writeFileSync(path.join(OUT, 'manifest.json'), JSON.stringify({
  _contract: 'Every file behind the AI recommendation reliability study, released CC BY 4.0 so the findings can be checked against their author.',
  study: 'https://lucreya.com/articles/ai-recommendation-reliability-2026',
  licence: 'https://creativecommons.org/licenses/by/4.0/',
  generatedAt: new Date().toISOString(), rawAnswerRows: storeRows, files: manifest,
}, null, 1) + '\n');

/* ⛔ VERIFY THE ARTIFACT, NOT THE EXIT CODE. A publish step that copies zero files and exits 0 is
   the failure this estate keeps paying for. */
const written = plan.filter((f) => fs.existsSync(path.join(OUT, f.rel))).length;
console.log(`\nWROTE ${written}/${plan.length} files + README.md + manifest.json -> ${OUT}`);
if (written !== plan.length) { console.error('REFUSED: not every planned file landed.'); process.exit(3); }