A clearer map of Lighthouse's accumulated knowledge

In short. The first prototype put too much checking detail in the reader's path. At John's request, this iteration adds an overview of Lighthouse's three investigations and makes the existing trails and evidence separate levels. It also draws a larger documented history within the software-update investigation, without claiming a new research benefit.

10 November 2026. OpenAI Codex; exact model identifier and attributable cost unavailable. User-approved interface iteration following the pilot, not a fourth investigation, new empirical study or permission to collect external material. The original protocol, comparison reports and scores remain historical records of version 0.1. This plan is retained before testing the revised interface; its editorial design is informed by records already seen and is not an unseen-outcome protocol.

Fixed boundary and representation

Use the pilot's original input commit, 8af0cb026037a8ee63456900978b3c44473c49cd. Inventory every investigations/.md declaration, every studies/LH/ directory with a top-level Markdown lead (write-up, then brief, then first Markdown file), and every released top-level articles/.md and essays/.md piece. Exclude drafts, short-form duplicates, nested raw evidence and checkpoint documents from this catalogue's node population. Retain all explicit internal Markdown links among these nodes, plus declared study/piece/current-account metadata. A link records a reference, not support or intellectual influence. A record's existence is not a reviewed finding. Missing targets and studies not explicitly linked from an investigation remain visible in the coverage record.

The whole catalogue is generated, with hashes, from this boundary. Its completeness is file inventory completeness, not completeness of every scientific claim, the source packs or outside research. A small editorial projection explains how one inquiry developed; it is labelled as a selected history, with exact evidence for each typed relation. No centrality or influence scores, forced-layout graph, new database or framework.

Reader design

1. Overview at /map/knowledge-lineage/: three questions, a legible larger research-history drawing, routes to the two detailed trails and the full catalogue. Status and consequential uncertainty remain visible; routine provenance is optional.
2. Thread at /map/knowledge-lineage/trails/: concise cards, with dates, authorship and pinpoints inside native disclosures. Evidence at /map/knowledge-lineage/evidence/ retains the original typed relationships, sample and qualifications.
3. Every view works with native links and without JavaScript. The drawing has phone and desktop layouts, directly linked records and a text equivalent. No zoom or hover is required to read a label. Large-scope coverage and original-source limitations have one clearly labelled explanation rather than a repeated caveat on every card.

Checks, record and stopping

Verify corpus membership and all selected-history edges against frozen originals; preserve explicit references separately from typed editorial links. Test that draft pieces and short forms are excluded and that unassigned work cannot disappear. Check all local links, fragments, root and subdirectory builds, native disclosure markup and no-JavaScript content. Render and inspect light/dark phone and desktop proofs. These are static proofs; actual browser, keyboard, assistive-technology and unfamiliar-human acceptance remain open if the environment still cannot run them. A fresh agent may audit the sources and interface, explicitly as an agent proxy. Do not describe that as a human test or reuse the old 8/8 comparison as a result for this version.

Ceiling: this bounded interface iteration, two fresh agent contexts (editorial source check and final audit), no external research requests or paid APIs. Stop after fixes and one bounded recheck, retain remaining limitations, commit to the existing branch and update draft PR #48. No merge, release, schedule or programme obligation. Inference and machine cost remain incomplete if not exposed.