Why this file exists
The Link Health audit on 2026-04-19 flagged 222 sources where the stored DOI resolves, but to a paper different from the stored title. Karen's recurring complaint — most recently the Chin-Chan 2015 citation on the Alzheimer's signature page — is that clicking a reference in the wiki can land on an unrelated paper. The renderer was not at fault. The stored DOI values are wrong.
Per §2a (Citation Integrity): "NEVER generate a DOI from memory." The only acceptable fix is verified lookup against an authoritative index (PubMed, Crossref). This file tracks that remediation.
Triage buckets
Extracting the `### DOI resolves — but to the wrong title` table from `wiki/analyses/link-health-2026-04-19.md`, bucketed by stored-title vs. Crossref-title similarity:
| Bucket | Count | Interpretation |
|---|---|---|
| Similarity = 0.00 | 67 | DOI is semantically wrong — resolves to an unrelated paper |
| Similarity 0.01–0.29 | 80 | Highly suspect; often wrong but occasionally a valid paper with a very different abbreviated title |
| Similarity ≥ 0.30 | 75 | Usually correct DOI, stored title is a shortened or authored alias |
The sim-0 bucket is the priority: these are the cases where clicking the link does not land on the stored-title paper. That is the specific bug Karen has raised three times.
Corrections applied this session (verified via PubMed)
| File | Old DOI | New DOI | PMID | Evidence |
|---|---|---|---|---|
| `chin-chan-2015-environmental-pollutants-ad-pd.md` | `10.3389/fncel.2015.00170` | `10.3389/fncel.2015.00124` | 25914621 | PubMed match; title identical |
| `ahlstrom-2019-nickel-allergy-review.md` | `10.1111/cod.13272` | `10.1111/cod.13327` | 31140194 | PubMed match; title + all authors |
| `capdevila-2024-bacterial-metallostasis-sensing-trafficking.md` | `10.1021/acs.chemrev.4c00132` | `10.1021/acs.chemrev.4c00264` | 39658019 | PubMed match; title identical |
| `chernikova-2022-brain-gut-microbiome-asd.md` | `10.3390/nu14071459` | `10.3390/nu13124497` | 34960049 | PubMed match; title identical (pub date Dec 2021) |
Each corrected source file now carries `pmid`, `doi_verified: 2026-04-19`, and `doi_verified_source: pubmed`.
Cases confirmed unresolvable via PubMed
Papers whose journals are not PubMed-indexed. These require Crossref title-search (currently egress-blocked from this sandbox) or manual lookup by Karen:
- `bergman-2016-low-nickel-diet-review.md` — Journal of Clinical and Experimental Dermatology Research
- `brock-2015-selenium-thyroid-autoimmunity.md` — Journal of Restorative Medicine
These should be flagged with `<!— NEEDS VERIFICATION —>` until Crossref access is restored.
Remaining work
Sixty-three sim-0 entries remain to be resolved, plus ~80 sim-0.01–0.29 entries that should be spot-verified. The resolution procedure is mechanical and can be continued in subsequent sessions:
- Read `/tmp/doi-zero-queries.json` (or regenerate from the link-health audit — see regeneration snippet below).
- For each entry, run `search_articles` on the PubMed MCP with `LastAuthor[Author] AND YEAR[pdat] AND KEYWORD[Title]` using 1–2 salient keywords from the stored title.
- Fetch metadata for any single PMID hit; verify title matches.
- Write back the new DOI along with `pmid`, `doi_verified: YYYY-MM-DD`, `doi_verified_source: pubmed`.
- Append the row to this file's corrections table.
- If a paper is not PubMed-indexed, log it in the "unresolvable via PubMed" section above.
Regeneration snippet
If the triage JSON is lost, regenerate with:
```bash node -e " const fs=require('fs'); const text=fs.readFileSync('wiki/analyses/link-health-2026-04-19.md','utf8'); const lines=text.split(/\r?\n/); const rows=[]; let inSec=false; for(const line of lines){ if(line.startsWith('### DOI resolves — but to the wrong title')) { inSec=true; continue; } if(inSec && line.startsWith('##')) break; if(!inSec) continue; const m=line.match(/^\| \`([^\`]+)\` \| \`([^\`]*)\` \| ([^|]+) \| ([^|]+) \| ([\d.]+) \|/); if(!m) continue; rows.push({file:m[1], doi:m[2], stored:m[3].trim(), crossref:m[4].trim(), sim:parseFloat(m[5])}); } const zero=rows.filter(r=>r.sim===0); fs.writeFileSync('/tmp/doi-triage.json', JSON.stringify({total:rows.length, zero, rows}, null, 2)); console.log('total:', rows.length, 'sim-zero:', zero.length); " ```
Tooling follow-up
The link-health audit currently flags the problem but stops there. A next-step improvement would be to extend `scripts/link-health-audit.cjs` with a `—resolve-missing` flag that:
- Accepts a list of files to attempt auto-resolution on
- Queries PubMed (and Crossref when egress is available) by author + year + title
- Writes verified DOIs back to source frontmatter with provenance
- Logs to a dated `wiki/analyses/doi-corrections-YYYY-MM-DD.md`
That is worth building once ≥20 manual corrections have confirmed the pattern reliably. At present the per-paper queries still need human judgment on ambiguous matches (shared author names, series papers with similar titles), so full automation is premature.
Systemic point
Once this pass is complete, the link-health audit and source ingest should be tightened so new sources cannot land in the vault with unverified DOIs. The ingest workflow already says "NEVER generate a DOI from memory" — the next enforcement step is to require `doi_verified` in frontmatter for any source whose DOI was added without a PMID attached, and to fail the build for any source page lacking both fields. That's a §2a rule that should be teeth, not a norm.