Lighthouse

Lighthouse watches the internet for what AI systems do there and what their activity leaves behind, and publishes what it finds, with the evidence and its limits.

Notes

Follow an idea: Knowledge Lineage

In short. The knowledge map starts with Lighthouse's three questions, then opens into selected histories, individual records and their evidence. All 49 study records and 21 released publications have explorable entries, and builds refresh the catalogue automatically. Source review and browser checks pass; improved human understanding remains to be tested.

Version 1, 11 November 2026. Release assessment and verification, maintenance, source review and code review. The material below retains the pilot's scope and reproduction notes. The dated version 0.2 assessment describes the earlier interface, not the current release checks.

Version 0.2, 10 November 2026. Written by OpenAI Codex, exact model identifier and attributable cost unavailable. Scope: cross-cutting method pilot under issue #46, using the still-open proposal #47. No fourth investigation, new collection or schedule. Existing released prose is unchanged.

Open the knowledge map. This releases navigation through existing material, without upgrading any working record's research or review status.

The three levels

  • Overview: three investigations, 49 study records and 21 released pieces at the frozen input, plus eight selected readings showing how the core software-update inquiry developed. All scoped items are reachable, including ten outside the investigation groups' explicit links. The inventory is complete within its declared boundary; this is not a complete claim-level or external-research graph.
  • Thread: the two original journeys, with routine provenance inside native disclosures.
  • Evidence: original passages, typed relationships, sample, dates and access limits.

Overview plan, source/editorial check, final agent audit, iteration assessment, catalogue, and selected-history annotations. The original protocol and comparison reports remain the record of version 0.1, before this interface changed.

What is here

Reproduce

From the repository root:

python3 scripts/lineage/extract.py --check
node scripts/lineage/overview.mjs --check
python3 scripts/lineage/draw.py
python3 scripts/lineage/draw_overview.py
npm ci --prefix site
npm test --prefix site
npm run build --prefix site
python3 scripts/lineage/check.py
python3 scripts/lineage/preview.py
npm run serve --prefix site

Visit /map/knowledge-lineage/. extract.py without --check rebuilds the JSON from the exact commit in baseline.json plus explicitly versioned annotations. Every edge retains its passage, line span, input hash and frozen-revision URL. It fails on a missing passage. A newer source requires a deliberate input/version change; it cannot silently change this history.

The tests check source passages and edge endpoints against frozen git, sampling, candidate separation and direction. They do not decide whether an attributed external event happened. The local check validates generated links and reproducibility without making external requests. The build has existing unresolved-link warnings elsewhere; the pilot's local links are checked separately.

Static proofs can be regenerated after the site build with python3 scripts/lineage/render_proofs.py, requiring WeasyPrint 70 and PyMuPDF. The session used an isolated scratch install. The script prohibits network fetches, models closed disclosures and applies the phone layout explicitly because a print renderer is not a browser viewport engine. The overview has separate tall and wide layouts; the tall one stays in use until the wider drawing has room for readable labels. Use a real browser for acceptance, not these proofs.

What the sample can and cannot say

The fixed position rule selects 17 of the current table's 87 rows. The earlier Claude review covered 83, a different denominator. Twelve external records are source-panel entries, not twelve successful source reads or a systematic literature corpus. The external provider report is retained context outside those twelve selected entries. No DOI or citation index matching was attempted, so there is no match-success metric.

Historical review verdicts, an HTTP access failure and its later resolution stay separate from the corrected claims. Original full-text source-pack paths are explicitly unresolved; a reviewer quotation or source URL does not replace a locally retained original passage. A source account is not independent incident confirmation. The thematic bridge between the two trails remains an inferred candidate; the recorded reading of the first vulnerability brief by the second is a narrower documented reuse link.

Why this could become useful

The strongest opportunity is to make claims, caveats and revisions reusable across time. A reader could see what changed since their last visit; a researcher could find the strongest earlier evidence and the unresolved disagreement before commissioning another study. Each later observation could extend a trail rather than produce another isolated page.

Keep the record centred on bounded claims, with bibliographic links, attributed intellectual influence, evidence relations and operational outcomes in separate types. Multiple parents and incompatible accounts should survive. Unknown dates and inaccessible regions should remain visible. A citation count or central position should never substitute for truth or benefit.

The proposal points to Lancaster's thematic review, the AI-security knowledge-graph paper, OpenAlex, Semantic Scholar, OpenCitations, W3C PROV-O and nanopublications. Those are methodological leads already identified by the proposal, not sources newly validated by this pilot. The current local JSON borrows the entity, activity and provenance distinction; it claims no PROV/RDF conformance. The next literature step should choose and inspect a small antecedent-paper set under a new fixed rule, starting from existing scholarship rather than inferring a field's history from this vendor case.

Next

John authorised the maintained feature and merge after review and verification. Source review and real-browser layout and keyboard checks now pass, as recorded in the release assessment. Give an unfamiliar human the two journeys without coaching, and check assistive-technology behaviour. Preserve errors and misunderstandings before changing the interface.

If that works, add optional “earlier evidence / what changed / used later” blocks to one existing investigation, with an explicit small maintenance allowance. The twice-weekly research scout could attach existing claim IDs and a reason for reuse to leads it already produces; the existing proposals gate still decides what starts. No new crawler or timer is needed. Ask whether a real investigation choice changed before claiming better research. Only then consider wider citation coverage, multiple evidence parents and a broader landscape view.

Corrections and cost

Protocol amendment 1 repaired the original prose-link sampling rule. Review found an ambiguous visual direction, overly short excerpts, missing date/access metadata and a missing pilot-index route; these were repaired and the review keeps its initial findings. The evaluation states its remaining limits.

No paid external research API was used. Inference spend and fresh-agent overhead are unavailable in this harness, not zero. Package installation, CI and local computation were not metered. The existing Harbour pricing commands need a named session/account and were not used to read credentials or invent a bill. No current scientific account changed; the release register records the navigation feature separately.

From notes/knowledge-lineage/README.md in the repository, last changed 11 November 2026.