How the record was read
This site reads a single database built by the insight-bridge pipeline from its corpus: 1,501 sources, 1,501 source documents, 47,381 passages, 9,642 source–topic memberships, 9,642 source-level stances, 1,025 voice perspectives.
Everything on every page is a live query against Cloudflare D1 (the run's own tables, copied whole) and Vectorize (the embeddings of the open-distribution works), through fully-typed Drizzle. The same data is open to agents over an MCP server at /mcp.
What was done, in order
Each step below is a stage this run actually executed, read from the run's own record of itself; a stage the run was configured to skip is not listed.
- Curation
The corpus was assembled by David B. Wilkins and Anthea Roberts for the AI and the Legal Profession research program of the Center on the Legal Profession at Harvard Law School: scholarship, industry evidence, primary legal materials and practitioner interview transcripts, 1949–2026, collected by hand, one document per source. Cross-domain, historical and theoretical sources (other professions, medicine, economics, sociology) are included deliberately as analogy and framework evidence. This is a hand-built evidence base, not a sampling frame: the selection is the first and largest analytical choice in the whole reading.
- Reading in the source documents
Every source document was stored whole under its source, its text cut into passages small enough to compare — 47,381 of them across 1,501 source documents from 1,501 sources — and each passage embedded with text-embedding-3-large, which is what makes a search by meaning, and the grouping below, possible at all. Each source carries the curator’s coding from the corpus manifest: a voice (who is speaking), a source class and type, jurisdiction, evidence type and topic dimension (the last three multi-valued), year, publication, author, distribution and url. The era — five ordinal buckets over publication year — is derived from the year with the run profile’s cut-points: voice answers who is speaking, era answers when.
- Reading each source document end to end
A language model read every source document in full against this run's profile and recorded what it found as key points, each with the verbatim quotes that support it: 17,122 key points and 52,300 quotes. Every quotation is verbatim from the source source document: each is verified against the source document's text before it is stored, and one that cannot be found there is never presented as verified. 99.9% verify, and a quote that does not is never shown here. Views are attributed to the author or interviewed speaker, not to hosts or people they merely quote.
- Settling what counts as one source
The names the model inferred were put to a person beside the names the record carries, before anything was counted, so that one source appearing under two spellings is counted once in every figure on this site. The record's name stands; the model's reading is kept beside it.
- Grouping the passages into topics
Topics are found from the passages themselves, bottom-up, by grouping the passages across the corpus that say the same thing; nothing supplies a list of themes beforehand. A topic has to be raised by more than one source to form: one source's volume cannot make a topic on its own, and the rare topic that only one source carries is flagged as such. A group had to hold at least 3 passages to count as a topic. The settings were tuned at a review gate: 2 attempts were tried before these were committed (they are listed below). Each passage, and each source, is then placed against every topic at a graded strength (exemplar, high-value, member), so a source that touches a topic in passing is recorded as well as the sources that define it, and a passage can belong to more than one topic. 395 topics resulted, over 47,381 passages.
- Writing each topic up and reading positions against it
Each topic is written up as a proposition with key points, and every exemplar and high-value source's position is recorded against that proposition, on this run's own scale, Supports, Builds on, Mixed, Redirects, Opposes and Unclear. Positions are relative to the topic's own framing, never absolute agreement, and every source is measured against the same proposition. 9,642 positions were written, and the same was done for each value of professional voice and time period: 1,025 perspectives. A position here records how a source stands against one topic's particular framing, and is never a verdict on the subject in general.
- Removing the groups that carry nothing
A model flagged topics whose synthesis reported no substantive shared content, and a person confirmed or rejected each flag — and could confirm a topic the model had not flagged. 26 topics were confirmed and are excluded from every figure on this site. The flag alone never removes anything.
- Offering the topics upward into themes
The topics were grouped again, generation by generation, each generation clustering the one below it until it stopped merging, held to 2 generations, then cut to 5 groups at the top. A topic with no theme above the attachment threshold was attached to its nearest theme rather than left alone. This site reads the tree as the run left it, however many generations that turned out to be — here, 51 families above the topics, each named by the model from its members.
- Placing the topics in a sky
Each topic's place in the galaxy was computed once and stored with the run, seeded so it is the same sky every time. That is what lets these pages draw the constellation as plain graphics, with no WebGPU and nothing solved in your browser.
- Loading this site
Outside the pipeline, before loading: the run database was copied whole to the cloud — every table as the pipeline wrote it, with the passage embeddings moved to a vector index (open-distribution works only) — and two facts were derived from it for these pages: each document’s and each passage’s dominant topic. The entity review’s suggested duplicate groups (versions of one work, one report under two ids) were recorded, never collapsed: counts are the pipeline’s. A restricted-distribution work’s text is never served on any page or to any agent; its analysis is. Nothing was flattened, re-verified or re-classified here.
The clustering attempts
The tuning gate walked these settings before the last row was committed. The minimum cluster size decides how many passages a group must hold to count as a topic; noise is the passages left in no topic.
| Attempt | Min cluster size | Topics | Noise passages |
|---|---|---|---|
| 1 | 3 | 437 | 44,252 |
| committed | 3 | 395 | 43,590 |
What each source was classified by
Every facet the run's profile declares. A facet marked compared is one the analysis wrote a perspective for, value by value — the lens you can read a topic through. Facets that are not compared ride beside it as facets of the record.
What a position means here
Read weakest claim first. Supports, Builds on agree with a topic's proposition; Mixed, Redirects, Opposes push back against it; Unclear reads as neither.
Position is relative to the cluster proposition, not the source's tone. Nuance alone is not Mixed, and constructive engagement alone is not Builds on.
How to read the output
These rules apply to every Insight Bridge corpus. They are properties of the method, not caveats about a particular run.
- Positions are relative to a proposition. A source’s position records how it stands against that cluster’s particular framing — not whether it agrees with some absolute claim. The same source can support one cluster and redirect a neighbouring one that covers similar ground differently.
- Propositions are written from the cluster’s own members. Because the argument is built from the sources that were grouped together, and positions are read relative to it, a degree of agreement is built into the method. Comparisons between groups carry weight; a corpus-wide agreement rate does not.
- Counts describe the corpus, not the world. Every corpus here is curated. “N sources say X” measures what was collected and is never a measure of how common X is in the field.
- Clusters differ in how many distinct sources back them. A long document can fragment across many clusters, so weight a theme by the distinct sources beneath it rather than by how many clusters it contains.
- Every extraction and position is a model judgement. Key points, stances, propositions and syntheses are produced by a language model reading the source. They inherit its calibration and are not determinations of fact.
In this run
Removed after review. “Evidence gap: Cluster lacks supporting summary” — The summary expressly states that no proposition or evidentiary direction can be recovered from the material. “AI risk: Materials establish no position” — The material is described as establishing no substantive position on AI risk management. “Synthesis gap: Evidence cannot support interpretation” — The summary explicitly reports that the supplied material supports no substantive synthesis or defensible proposition. “Evidence absence: No cluster claim recoverable” — No topic, argument, or meaningful summary content is recoverable from the supplied material. “Research disclaimers: Qualified independence coexists with interests” — The underlying content consists of standard financial-research disclaimer and conflict boilerplate rather than substantive research evidence or argument. “Clinical scribes: Documentation evidence remains insufficient” — The summary expressly says the material cannot support any substantive account of ambient AI scribes, workflow burden, or outcomes. “Theme gap: Excerpts establish no position” — The summary explicitly states that the excerpts establish no substantive theme, proposition, disagreement, or practical implication. “Support gap: No substantive synthesis possible” — The summary expressly reports that no reliable proposition or substantive direction for synthesis is available. “Hallucination evidence: No findings can be synthesized” — The summary explicitly states that no defensible hallucination findings, rates, incidents, or consequences can be synthesized. “Survey fragment: Missing metric defeats interpretation” — The source is an incomplete survey fragment whose missing metric makes the reported percentages uninterpretable and prevents recovery of a substantive proposition. “Synthesis limit: Cluster theme remains unsupported” — The summary expressly states that the supplied material supports no substantive claim, conflict, or practical direction. “Evidence void: No proposition supplied” — The cluster is explicitly empty and supplies no proposition, evidence, or substantive theme. “Hallucination synthesis: Supporting material is absent” — The summary expressly reports that no reliable hallucination findings, rates, or doctrinal implications are present. “Firm diversity: Excerpts support no conclusion” — The supplied excerpts are described as establishing no substantive position on law-firm diversity, representation, or advancement. “Missing evidence: Cluster cannot be differentiated” — The cluster contains no summary content from which any substantive theme, proposition, or relationship can be identified. “Reliance limits: Disclaimers provide no AI findings” — The material consists only of reliance disclaimers separating guest views from institutional positions and supplies no substantive legal-AI proposition. “Theme absence: Summaries establish no substance” — The summary explicitly states that the supplied material establishes no substantive theme, proposition or evidence. “Contribution gap: Materials establish no debate” — The summary expressly reports that no substantive position, evidence-backed finding or useful analogy can be extracted. “Employment evidence: AI effects remain unestablished” — The summary states that the supplied material establishes no substantive findings about AI employment effects. “Publisher provenance: Forum status proves no findings” — The material consists only of generic publisher provenance and forum-status information without a distinct substantive finding or argument. “Contributor context: Credits establish provenance only” — The material is limited to institutional descriptions and contributor credits that establish provenance but no attributable substantive proposition. “Podcast closings: Contact routes replace substantive evidence” — The underlying material consists chiefly of podcast closing boilerplate and contact routes rather than a distinct substantive proposition or evidence. “Closing narrative: Broad change lacks developed argument” — The material is only a generic closing narrative about broad change and supplies no developed argument or distinct evidence. “Employment synthesis: AI attribution remains unsupported” — The summary expressly indicates that the supplied material supports no substantive or reliable proposition about AI-attributed employment effects. “Harvey resources: Promotional headings provide no guidance” — The underlying material appears limited to promotional headings and resource pointers that supply no distinct guidance, finding, or evidentiary proposition. “Age comparisons: Employment and pay findings remain absent” — The material identifies only a broad subgroup topic while supplying no effect sizes, results, evidence, or distinct substantive conclusion. They are excluded from every figure on this site, including the 369 topics counted above.
Not analysed. ; . Step 1 produced no key points for these documents; they keep their topic memberships.
Restricted works are analysed, not reproduced. 124 sources have restricted distribution (paywalled or login-only). The pipeline read them, and their stances, key points and quotes are here like any other's; their passage text is never served, on any page or to any agent, and passage search never finds them.
Some works appear under more than one id. The entity review suggested 27 groups (71 sources) that are versions of one work or one report lodged twice. They are flagged where they appear and counted as the pipeline counted them.
What else was asked of every source document
Beyond the standard reading, this run declared its own questions and put each of them to every source document.
Words this corpus uses
Terms the run was given before it read anything, so that its own writing uses them the way this field does.
Which model did what
Named because it matters: the reading, the writing-up and the grouping were done by these, and a different model would have produced a different record.