Knowledge Lineage: first maintained release

In short. Lighthouse's existing work can now be explored through questions, a selected visual history and individual evidence-linked records. Every study in the declared catalogue has a checked explanation. This is useful navigation and provenance infrastructure; whether it improves readers' understanding or research decisions remains an open empirical question.

11 November 2026. OpenAI Codex driving session and fresh review contexts; exact model identifiers and attributable inference costs unavailable. Released at John's direction to include existing Lighthouse work, review, verify and merge. No new investigation, collector or scheduled task.

What is covered

At the reviewed source revision 8af0cb026037a8ee63456900978b3c44473c49cd, the catalogue contains three investigation declarations, 49 studies, 18 released findings and three released essays. All seventy study/publication entries have their own explorable pages. Studies retain their question, answer, main limit, state and exact source passages. There are 204 retained explanation passages, 228 explicit reference/declaration edges, 79 checkpoint links and a browsable archive of 67 numbered source reviews. A reference is not an assertion of support or intellectual influence. The archive does not assert a formal review relationship for every note and study.

The overview gives readers three questions and an eight-document visual history before asking them to inspect the full catalogue. The original It was only a test and When a model finds the bug trails remain the inward/outward pilot. All-work coverage means this declared record and publication boundary, not extraction of every claim, source file or scientific relationship. The external trail is not a complete history of AI-security research.

Review and verification

• The source review read all 49 explanations against the original records and checked all 204 exact passages. All 21 publication summaries match existing editorial summaries. One vendor-population error was corrected and passed its bounded recheck. No source blocker remains.
• The code review found and rechecked three maintenance repairs: explicit released-publication selection, withholding annotations after whole-source changes, and a downloaded catalogue that matches the current HTML input. No code blocker remains. The reviewers are host-model agent contexts, not independent confirmation or an Opus review.
• All 51 site tests passed. Isolated git fixtures verify new-study discovery, exclusion of unreleased publications and withholding a stale explanation even when its old quotation remains in a changed source.
• Root and /lighthouse/ builds passed. The local link checker verified 2,406 targets in each, with zero missing targets or fragments.
• GitHub Actions run 42976153157 passed both root and subdirectory jobs on 5c689a6598e417deefde30fca74f9d94116f2312. Job logs record actual Chromium acceptance at 390, 768 and 1,280 pixels in light and dark themes: all 49 study pages loaded, explanations appeared, no horizontal overflow assertion failed, native catalogue disclosure opened with the keyboard, and JavaScript was disabled. These are automated interaction checks, not unfamiliar-human or assistive-technology acceptance. Screenshots were generated in CI but not retained as uploaded artifacts or inspected by this session.

The original pilot comparison remains 8/8 retrieval answers in both conditions. It establishes no accuracy gain. No new timing, comprehension or research-quality result is claimed for the expanded interface.

What maintains itself

Each build discovers current records and explicitly released publications and regenerates both HTML and the downloadable catalogue. New records remain reachable even before an explanation is curated. A source hash change withholds its reviewed explanation until checked; changed evidence behind the selected history withholds that drawing. This conservative guard can also withhold text after harmless formatting changes.

Agents still need to check meaning, write short explanations and select useful histories. Do that within existing publication/correction reviews, checking only affected annotations and retaining their provenance. The maintenance instructions separate this reviewed curation revision from the original experimental baseline, which stays immutable. No citation or similarity algorithm decides that one idea caused another.

What should happen next

Keep this as a permanent navigation layer, then test it with an unfamiliar reader on a phone. Ask them to find a result, describe its population, follow a correction and identify an unresolved question. Retain wrong turns and misunderstandings; inspect assistive-technology behaviour too. Those observations should guide the next visual iteration.

For research usefulness, record whether a future investigation actually reuses earlier evidence, avoids duplicate work or changes its next question because of the map. Compare against the existing workflow before claiming a gain. The first experiment's equal retrieval scores do not warrant that claim.

The larger opportunity is separate inward and outward views over evidence-backed records: a system overview, selected explanatory histories and detailed source routes. Expand outward using one bounded, reviewed literature corpus and established scholarship. Keep citations, attributed influence, verified evidence and measured practical outcomes as distinct relationships, with gaps and conflicting accounts visible. Expand the visual projection when checked material warrants it, rather than trying to draw everything at once.

No scientific conclusion, research schedule or investigation allocation changed. No paid external research API was used. Inference, CI, package/network and machine costs were not exposed, so accounting is incomplete rather than zero. Human comprehension, accessibility acceptance and research benefit remain uncertain.