From an evidence trail to a map readers can explore

In short. The interface now starts with questions and a larger history, with checking details available at deeper levels. A frozen catalogue includes all three investigation declarations, 49 study records, 18 released findings and three essays within its stated boundary. This demonstrates a path to larger bounded systems, but it does not establish a new gain in research quality or human understanding.

Version 0.2, 10 November 2026. OpenAI Codex; exact model identifier and inference cost unavailable. Working interface assessment, not a released finding. User authorised after reading version 0.1. The original protocol, comparison and 8/8 item scores are unchanged historical results of that earlier interface. The new scope/checking plan preceded these interface checks, and was informed by already-seen evidence rather than an unseen empirical outcome. Input revision remains 8af0cb026037a8ee63456900978b3c44473c49cd.

What changed for the reader

The overview is organised around three questions. A larger, directly linked drawing follows eight released readings in the core Python inquiry through new observations, broader samples, revised explanations and an unsettled answer. The visual carries the later qualifications on early leads; its caption keeps the download finding's population and lower-rank boundary visible. The current synthesis and later continuations sit beside it.

Thread pages give the story first, with dates, authorship and pinpoints inside native disclosures. The detailed evidence record remains a separate destination. The full catalogue is browsable beneath the overview and does not require reading its checking notes. The standalone HTML preview combines the three levels with native anchor navigation, embeds local fonts and includes their licences; the actual site gives each level its own page.

The overview is an inward map. The existing model-discovery trail provides one outward route through the public record. This iteration has not collected a whole external research field. It establishes the same design path for a bounded reviewed corpus later, with citation, evidence, claimed influence and measured outcome relationships kept distinct.

Coverage and scientific interpretation

The offline generator inventories every investigation declaration, each study directory with a top-level Markdown lead, and every released top-level article or essay at the frozen revision. Draft pieces, short forms, checkpoints and raw evidence sit outside this node population. There are 73 unique nodes, 228 retained reference/declaration edges, and no unresolved targets among the scoped inline references inspected. Source files and short-label metadata retain their hashes. All 73 nodes are reachable through group headers, linked records or the explicit ten-item Other work list; groups can overlap, so their counts cannot be added.

This completeness is a property of the declared catalogue, not of all claims or evidence held by Lighthouse. The reference extractor recognises direct Markdown links and investigation metadata; it does not infer relationships from prose similarity, publication order or co-occurrence. It reads a study's lead and brief, not every nested source or checkpoint. The ten unassigned records have no qualifying direct group link under that rule; that is not a claim they are scientifically unrelated to an investigation.

The selected history is explicitly smaller than the complete catalogue. Its eight nodes are released documents, not newly invented composite results. All eight directed relationships have bounded passages documenting what the later reading did with the earlier one. The verb describes the later reading while the arrow follows earlier-to-later inquiry order. Arrows describe Lighthouse's recorded research path, not an external causal effect, scientific importance or a model's intellectual influence.

Each history summary has separate retained links for its result and the caveat, correction or later qualification that changes its meaning. Earlier advisory and release-note leads remain historical; the current release-note conclusion is unsettled, not evidence of no effect. The wider download result is a per-library majority-of-releases measure, not the share of all releases pooled. Downloads remain fetches, pins remain requested versions, and images remain installations rather than runs.

Checks and repairs

The editorial source reading read the released pieces and proposed the eight-document history. The fresh final audit checked corpus membership, reachability, relations and claims against frozen originals. Both are agent checks on the host model, not unfamiliar-human tests or the formal publication review. Neither collected new external evidence.

The review found result pinpoints that omitted the actual observation or qualification, an ambiguous population label for the download test, the need to preserve the lower-rank boundary on the overview, labels shrinking at intermediate widths, and a return arrow ending outside its destination. These prompted multiple labelled evidence spans, a precise majority-of-releases summary, a visible population qualification, the tall drawing up to a 1,071px viewport, and a corrected directed path. The large drawing is now inline with unique IDs so a box opens its article directly. Native text links provide the same route. No hover or zoom is required to reveal a label.

All 49 site tests passed. The final local check found 281 valid links/fragments in each of the root and subdirectory builds, with no missing local target. Local link/fragment checks cover overview, thread, evidence and pilot records. Site tests guard the full reachable node set, draft/short-form exclusion, typed explicit references, selected-history source passages and direction. Both root and subdirectory builds are checked. Static light/dark proofs model closed disclosures and use explicit matching SVG palette values because the print engine does not resolve their variables. Phone and desktop overview proofs and phone thread proofs are inspected. Those are static render proofs, not browser screenshots.

Actual mobile-browser layout, keyboard operation, assistive-technology behaviour and unfamiliar-human comprehension remain unverified in this environment. The prior local-browser socket failure and cloud-localhost access block were not worked around. Source/markup checks of native links and disclosures are narrower evidence than interaction testing. The original baseline and assisted scores are not retests of this revised overview, and no timing, density or comprehension improvement is claimed as a measured outcome.

What this earns and what remains

It earns a reviewable demonstration that one auditable record can support an overview, a selected explanatory drawing, short threads and deep evidence. The full catalogue prevents unseen work from disappearing when the visual projection is selective. It does not earn continuous ingestion, a new investigation, a declared complete external landscape or a research-quality claim.

The next test remains an unfamiliar reader using the real pages in a mobile browser: find one result, explain which population it covers, follow its revision, and identify an unresolved question. Record wrong turns and misunderstanding before changing the interface. If it works, the existing scout can later propose a bounded external corpus using established scholarship, with explicit antecedent links and gaps. A larger map should be another projection of checked records, not a reason to invent connections.

The iteration changes no released prose, current scientific account, schedule, programme commitment or release list. No paid external API was used; inference, package/network overhead and machine cost were not exposed, so accounting remains incomplete rather than zero. Work stays in draft PR #48 for review.