Network-wide Link Inventory v1.4

Network-wide Link Inventory v1.4

# Network-wide Link Inventory v1.4

Lee Sharks · 2026-07-08 · Dataset

AXN: `AXN:0431.ARCHIVAL.🔗🕸️🗺️🔍📋⚖️`

Description

## Description

A per-link database of every anchored hyperlink across all 34 substantive repositories in the `leesharks000` network. Each row records the anchor text, current href, ~200 characters of prose before and after the link, ~1,600 characters of enclosing prose context, extracted DOI/AXN references from anchor text and href and context, a resolver lookup for any extracted DOI (returning the current live target, the resolver's stored title, the mapping type), a prose-vs-resolver-title jaccard similarity, the best-matching alternative record in the alexanarch registry, and a verdict flag classifying the link into review categories.

This dataset was assembled to answer a question that the earlier link-scan (a 440,012-row token count across the same 34 repos) could not: not just how many links there are, but whether each link is right — whether the href points at what the anchor text says, whether the DOI in the anchor text matches the descriptive prose around it, and whether the resolver's current mapping for that DOI is consistent with everything the surrounding context asserts.

v1.1 (2026-07-08): 106 fixes applied since v1.0 — 69 sentinel-0 hrefs corrected to specific records via unique-substring anchor-text matching against the registry; 34 href_disagrees_resolver corrections where the anchor uniquely matched a different record than the current href; 3 ambiguous fixes with high-confidence matches. Verdict deltas: href_wrong_target 258→208 (-50); ambiguous 11163→11149 (-14); consistent 1286→1303 (+17); href_disagrees_resolver 365→428 (+63, false-positive increase — sentinel-0 fixes now expose disagreement with resolver's DOI-based guess for nearby DOIs, though anchor-text-based fixes are semantically correct).

The immediate use is a review pile for correcting broken and mismatched links across sites — with corrections applied preferentially to the resolver map at its source of truth (deposit #4 v3.11.0+) rather than to individual site source files, so that the sovereign resolver behavior propagates automatically to every site link via the `/go/?doi=X` endpoint.

Companion deposits

## Companion deposits

Companion to [DOI Resolution Index v3.11.0](/s/records/4/) (#4) — this inventory audits the resolver's real-world usage across the site network. Complements [Zenodotus' Book-Burning](/s/records/1/) (#1) which established the sovereign-succession framework, and [DOIs ≠ Permanent Identifiers](/s/records/868/) (#868) which characterizes the substrate-wide breakdown of DOI persistence.

Artifacts

## Artifacts

Files live at `/data/link-inventory/`:

• `inventory-summary.json` (2 KB, compact JSON) — scan metadata, verdict distribution, per-repo counts, regeneration pointers.
• `link-review-actionable.csv` (~370 KB, 623 rows) — links with high-confidence mismatch verdicts: `href_wrong_target` (sentinel `/s/records/0/`), `href_disagrees_resolver` (href points at record N; resolver says the DOI-in-text should go to M ≠ N), `href_disagrees_text` (DOI in text ≠ DOI in href), and `likely_prose_mismatch` (prose context very different from what resolver says the DOI is, AND a much better alt record exists).
• `link-review-ambiguous.csv` (~600 KB, 830 rows) — DOI-carrying `ambiguous` cases with a promising alternative record match (best_alt_jaccard ≥ 0.15).

Regeneration scripts live at `/scripts/link-inventory/`:

• `build_link_review_db.py` — the scanner; produces the SQLite links database from a local mirror of the 34 repos.
• `enrich_link_review_db.py` — the enrichment pass; joins each link against the current DOI resolution index and computes jaccards, best-alt matches, and verdicts.
• `build_review_report.py` — exports the review artifacts (summary MD, actionable CSV, ambiguous CSV).
• `rewrite_links.py` / `rewrite_anchor_text.py` — the two rewriters used in the Phase 10 pass (URL-form and semantic anchor-text).

The 81 MB SQLite database itself (`links_review.db`) is not shipped in-repo; it is regenerable from the scripts against any local clone of the network. Scan takes about 32 seconds on a modern laptop.

Schema (SQLite `links` table)

## Schema (SQLite `links` table)

Verdict taxonomy

## Verdict taxonomy

Verdict distribution as of v1.0

## Verdict distribution as of v1.0

Mismatch concentration by repo (top 15)

## Mismatch concentration by repo (top 15)

Note: this is the inventory as of the anchor-text rewrite pass of 2026-07-08. Repos with rewrites still-not-deployed carry undercounted `href_wrong_target` counts (fixed in git but not yet in production).

Worked example — the "Generative Monoculture goes to Meaning Caste" case

## Worked example — the "Generative Monoculture goes to Meaning Caste" case

MANUS flagged `lagrange-observatory/index.html:250` — anchor text `DOI 10.5281/zenodo.20675438` next to prose about "Generative Monoculture: Model Collapse in Code" — but that DOI resolves to record #198 "The Self-Audit Module Dissolved". The actual "Generative Monoculture" work is at record #199 (Talos Morrow, June 13 2026) or the re-mint at #1023 (EA-UMBML-MONOCULTURE-01).

The row in the inventory:

• `anchor_text` = "DOI 10.5281/zenodo.20675438"
• `href` = "https://www.alexanarch.org/go/?doi=10.5281/zenodo.20675438"
• `doi_in_text` = "10.5281/zenodo.20675438"
• `resolver_target` = "https://alexanarch.org/s/records/198/"
• `resolver_title` = "The Self-Audit Module Dissolved..."
• `prose_vs_resolver_jaccard` = 0.043
• `best_alt_record` = #188 (adjacent link in source, not the correct match — the true match at #199/#1023 has weaker jaccard because the scanner captured trailing prose after the h3 title rather than the h3 itself)
• `verdict` = `ambiguous`

Two possible root causes and neither can be picked automatically: (1) the DOI in the lagrange source is wrong for the work described (someone typed the wrong DOI when the page was authored), or (2) the resolver's `20675438 → #198` mapping is wrong (Phase 6/7 fuzzy title matching landed the wrong record). The inventory surfaces the ambiguity; a MANUS-level decision picks which side is authoritative.

Iteration policy

## Iteration policy

The inventory regenerates cleanly from source at any point — no in-place mutation, no state drift. Standard iteration cycle:

1. Re-run `build_link_review_db.py` after each rewrite pass or resolver update to refresh the row set.

2. Re-run `enrich_link_review_db.py` after any resolver-map correction to refresh verdict flags.

3. Re-run `build_review_report.py` to regenerate the CSV exports.

4. Bump this deposit's minor version and update the summary JSON with new counts.

The versioned snapshots in `/data/link-inventory/` are the durable record of network-wide link integrity at each rescan point; they can be diffed across time to show which repos are converging toward zero mismatches.

Related identifiers

## Related identifiers

• Companion (DOI Resolution Index): [/s/records/4/](/s/records/4/) — the sovereign resolver whose correctness this inventory audits.
• Companion (DOIs ≠ Permanent Identifiers): [/s/records/868/](/s/records/868/) — the theoretical framing of DOI-persistence breakdown, of which the mismatches surfaced here are one empirical face.
• Framing (Zenodotus' Book-Burning): [/s/records/1/](/s/records/1/) — the sovereign-succession premise that made the resolver necessary.

∮ = 1