About

How This Publication Is Made

This site is produced by a newsroom of software agents working as a team, under one person who can publish. Each seat has a name, a job, and a fixed set of things it is allowed to touch. This page lists all of them, what the team produces, and the rules that keep it honest. It is generated from the same roster the software enforces.

Seats: 39
Desks: 9
Model seats: 35
People who can publish: 1
Under construction: 3
Roster version: 0.4

Organisation chart

Every seat, by department, with the number working in each. The board decides; the two desks below it never share context; the bench and the floors below serve both.

Dashed boxes are under construction. Tier names opus, sonnet and haiku are the model each seat runs on; (person) and (code) seats run no model.

What the team produces

One article per jurisdiction

Each cycle, one firm in each jurisdiction gets an article written from what the firm itself published: deals, matters, hires, recognitions, guides. Every paragraph carries the ids of the evidence it rests on, and every piece of evidence carries its URL and retrieval time.

Firm and lawyer rankings

A ranking of firms and a ranking of lawyers per jurisdiction, built by a desk that cannot see the articles, scored by code that anyone can re-run, and published with a visible coverage figure so a thin ranking says so.

Three rules the software enforces

1
Writers never rank and rankers never write
The article desk reads dossiers. The ranking desk reads evidence rows. They share the list of firms, lawyers and jurisdictions and nothing else. The wall is a permission the software refuses to cross, not a policy the team promises to keep.
2
No agent publishes
Every path to the live site ends at a person. Agents move work from draft to reviewed to approved; only the Publisher moves approved to published, and the word for that is go.
3
Every claim carries its evidence
A sentence points at a dossier item. A dossier item points at a fetched page with a URL, a timestamp and the site's robots status. A placement in a ranking points at the rows that produced it. An agent that cannot attach an id does not get to make the claim.

How a cycle runs

One command starts every seat. The run stops at the desk, where a person reads what the team approved. A second word, go, publishes it.

01The Publishersays update.
02The benchcrawls firms' own sites and permitted publications, extracts typed records, links them to firms and lawyers, collects images.
03The article deskresearches each jurisdiction, builds one firm's dossier, writes, criticises, refines, fact-checks, reviews for rights, and copy-edits.
04The ranking deskcollects observations, scores them with published weights, writes the methodology note from the run, and audits every placement adversarially.
05The translation deskrenders approved pieces into Korean, Japanese and Chinese with a cross-check for drift.
06The Publisherreads the desk and types go.
07The engineering floorruns the integrity checks, commits, pushes, waits for the build, and verifies every page is live.

The team, desk by desk

Each seat lists the model it runs on. The judgment model is used only where a mistake is expensive: writing, criticism, fact-checking, rights, audit, and the decisions of the board. Routine work runs on a working model, bulk extraction on the cheapest one, and two seats are plain code with no model at all so the rankings can be re-run by anyone.

ABoard

3 seats · 3 active

Three seats that decide. One is a person.

A01
Publisher
A person
Reads everything the desks approve and is the only one who can publish; types go.
May touch: read:draft · read:queue · read:review · read:runs · write:published · write:status
A02
Editorial Director
Judgment model, high effort
Sets the editorial calendar under a published charter of fairness, objectivity and independence; decides which firm in which jurisdiction gets this cycle's article.
May touch: read:articles · read:dossier · read:queue · read:review · write:brief · write:status
A03
Head of Research
Judgment model, high effort
Guards the ranking method: approves methodology versions and signs off every ranking run before it reaches the Publisher.
May touch: read:observation · read:review · read:runs · write:run · write:status

BArticle desk

9 seats · 9 active

Nine seats that turn a firm's own record into an article, with four independent reviews before a person reads it.

B01
Assignment Editor
Working model, medium effort
Turns the calendar into briefs: firm, angle, length, deadline, and which evidence is in scope. One brief per jurisdiction per cycle.
May touch: read:articles · read:dossier · read:queue · write:brief · write:status
B02
Jurisdiction Researcher
Working model, medium effort
Reads what every firm in one jurisdiction published in the last ninety days and shortlists the firms with something worth writing about, each with its evidence.
May touch: net:fetch · read:entity · read:evidence · write:dossier
B03
Firm Dossier Builder
Working model, high effort
Assembles the dossier for the chosen firm: every claim with its source page, quote and retrieval time, marketing language flagged as such.
May touch: net:fetch · read:entity · read:evidence · write:dossier
B04
Writer
Judgment model, high effort
Writes the article from the dossier and nothing else; every paragraph cites the dossier items it rests on. Has no web access by construction.
May touch: read:dossier · write:draft
B05
Critic
Judgment model, high effort
Checks the draft against the dossier sentence by sentence: unsupported, misattributed or omitted facts, then tone. Returns a numbered fix list.
May touch: read:dossier · read:draft · read:evidence · write:review
B06
Refiner
Working model, medium effort
Applies the Critic's fixes and nothing more; cannot add a claim.
May touch: read:dossier · read:draft · read:review · write:draft
B07
Legal Fact-Checker
Judgment model, high effort
Resolves every court name, citation, statute, date, counsel attribution and deal value against the records desk; rejects what it cannot resolve.
May touch: read:dossier · read:draft · read:entity · read:evidence · write:review
B08
Rights & Defamation Reviewer
Judgment model, max effort
Reads the article as opposing counsel would: named people in live proceedings, allegations stated as fact, anything needing a right of reply. Can block publication.
May touch: read:dossier · read:draft · read:evidence · write:review · write:status
B09
Copy Editor
Working model, low effort
House style, canonical firm names, dates, currencies, the dek and the search title. Last touch before review is complete.
May touch: read:draft · read:review · write:draft · write:status

CRanking desk

8 seats · 8 active

Eight seats that build the rankings and cannot see the articles.

C01
Ranking Director
Judgment model, high effort
Runs each jurisdiction's ranking cycle and decides which practice areas have enough evidence to rank. Cannot see article drafts.
May touch: read:observation · read:review · read:runs · write:run · write:status
C02
Evidence Collector
Working model, medium effort
Gathers recognitions, headcount, offices, submissions and revenue disclosures into dated, sourced observation rows.
May touch: net:fetch · read:entity · read:evidence · read:observation · write:observation
C03
Lawyer Profile Builder
Working model, medium effort
Builds lawyer records from firms' own bio pages: title, admissions, practice areas, disclosed matters and recognitions. Never scrapes a source whose terms forbid it.
May touch: net:fetch · read:entity · read:evidence · write:entity · write:observation
C04
Scorer
Code, no model
Pure code, no model: turns the published weights and the evidence rows into placements with a score breakdown. Re-running a published run reproduces it exactly.
May touch: read:observation · write:run
C05
Methodology Writer
Judgment model, high effort
Writes each jurisdiction's methodology note from what the run actually used, so the page can never describe a method that did not run.
May touch: read:observation · read:runs · write:run
C06
Ranking Auditor
Judgment model, max effort
Adversarial: for every placement, looks for the reason it should not publish. A run with an open blocker cannot go forward.
May touch: read:entity · read:evidence · read:observation · read:runs · write:review · write:status
C07
Dispute Handler
Judgment model, high effort
When a firm or lawyer disputes a placement, pulls the evidence trail, drafts the reply and, if warranted, opens a corrected run rather than editing the old one.
May touch: mail:draft · read:evidence · read:observation · read:runs · write:run
C08
Coverage Analyst
Bulk model, low effort
Bookkeeping: which jurisdictions have enough evidence, which are close, and which firms block the crawler. Feeds the public coverage report.
May touch: read:observation · read:runs · read:sources

DShared bench

8 seats · 8 active

Eight seats that fetch, extract, link, collect marks and choose pictures for both desks, and hold no editorial opinion.

D01
Source Compliance Officer
Judgment model, high effort
Reads robots.txt and terms before any source is used, records what is permitted, and is the only seat that can enable a source. A refusal is logged, never worked around.
May touch: net:fetch · read:sources · write:sources
D02
Crawler
Code, no model
Fetches with an identified User-Agent, conditional requests and per-host politeness, in two beats: firms' own sites, and permitted legal publications.
May touch: net:fetch · read:evidence · read:sources
D03
Extractor
Bulk model, low effort
Turns a raw page into a typed record: main text, metadata, language, dates. Bulk work on the cheapest model.
May touch: read:evidence · write:observation
D04
Entity Resolver
Bulk model, medium effort
Links articles, deals, records and bios to the right firm or lawyer: deterministic match first, a constrained model pass second, a human queue below 0.8 confidence.
May touch: read:entity · read:evidence · write:entity
D05
Court Records Clerk
Working model, medium effort
Parses official judgment and ministry feeds into case name, court, date, citation and link. Never stores the judgment text.
May touch: net:fetch · read:entity · read:evidence · read:sources · write:observation
D06
Deal & Matter Extractor
Working model, medium effort
From firm press releases: parties, role, value, currency, date and named counsel, each field with its quote.
May touch: read:entity · read:evidence · write:observation
D07
Image Collector & Evaluator
Working model, medium effort
Collects logos, headshots and announcement images from firms' own sites, and evaluates each for what it shows, its quality, and whether the page grants a right to use it. Every image is held for a human before it appears anywhere.
May touch: net:fetch · read:entity · read:evidence · write:assets
D08
Picture Editor
Bulk model, low effort
Chooses the picture for every page that needs one, from sources whose licence allows it (Wikimedia Commons under public domain or Creative Commons), and writes the caption and the credit. Never a publisher's photograph; never a picture without its licence recorded.
May touch: net:fetch · read:articles · read:runs · write:assets

ETranslation desk

3 seats · 3 active

Three seats that carry approved pieces into the site's other three languages.

E01
Translator: Korean
Working model, medium effort
English to Korean in legal register, firm and court names as the firm and the court use them.
May touch: read:draft · write:translation
E02
Translator: Japanese
Working model, medium effort
English to Japanese in legal register, same rules.
May touch: read:draft · write:translation
E03
Translator: Chinese
Working model, medium effort
English to Chinese in legal register; flags where Traditional and Simplified should diverge.
May touch: read:draft · write:translation

FEngineering floor

4 seats · 4 active

Four seats that keep the published site correct, fast, indexable and live.

F01
Release Engineer
Working model, medium effort
Assembles the published data files and writes the change report that becomes the commit message.
May touch: fs:repo · read:articles · read:queue · read:runs
F02
Data Integrity Engineer
Code, no model
Pure code: schema and cross-file checks, the 320-character excerpt rule, no star ratings. A red check stops the release before anything is committed.
May touch: fs:repo
F03
Site QA & SEO Engineer
Working model, medium effort
Builds and crawls the site, checks structured data, languages and sitemaps, owns what answer engines read, and verifies every published page is live after an upload.
May touch: fs:repo · net:fetch
F04
Front-End Designer & Uploader
Judgment model, high effort
Designs layout and component changes and shows them as before/after screenshots; on the Publisher's go, commits, pushes to GitHub and waits for the build. The only seat that can push.
May touch: fs:repo · git:push · read:assets · read:queue

GFirm relations

1 seat · 1 active

One seat that drafts every letter the publication sends, for a person to send.

G01
Correspondence Agent
Working model, medium effort
Drafts right-of-reply requests, crawl-access requests, claim confirmations and correction acknowledgements. Drafts only; a person sends.
May touch: mail:draft · read:draft · read:queue · read:runs

HSocial desk

1 seat · 0 active · 1 under construction

Under construction. One planned seat that will turn published work into posts for LinkedIn and other platforms, for a person to post.

H01
Social Media Editor
Working model, medium effort
Under construction
Turns published articles and ranking releases into posts for LinkedIn and other platforms, each linking back to the page it summarises. Drafts only; a person posts.
May touch: read:articles · read:runs · social:draft

IMember services desk

2 seats · 0 active · 2 under construction

Under construction. Two planned seats for member firms: a monthly report of how their profile and coverage performed, and a desk that answers their questions. Aggregate counts only; no visitor is ever identified.

I01
Member Analytics Analyst
Bulk model, low effort
Under construction
Prepares each member firm's monthly report: views of its profile and articles, clicks through to its site, news in which it was named, and movement in its standing. Reads aggregate counts only, never who visited.
May touch: analytics:read · read:articles · read:runs · report:draft
I02
Member Relations
Working model, medium effort
Under construction
Answers member firms' questions about their profile, their standing and their report, and routes corrections and claims to the right desk. Drafts only; a person sends.
May touch: mail:draft · read:queue · read:runs · report:draft

Where a person decides

No agent publishes, sends an email, enables a source, uses an image, or changes the look of the site. Each of those is a gate, and this is what the person at the gate sees.

Publish an article
Publisher
The draft with its dossier, three reviewers' sign-offs, translation diffs.
Publish a ranking run
Publisher, after the Head of Research
Placements with their score breakdown, zero open auditor blockers, the method diff.
Enable a source
A person, on the Compliance Officer's recommendation
The site's robots.txt and terms, and the proposed permission.
Use an image
A person
The image, where it came from, and the rights signal on that page.
Send any email
Publisher
The draft, the recipient, the reason.
Ship a design change
Publisher
Before and after screenshots.

What it costs, and what it does not do

The newsroom runs under a hard spending limit of $20 a week and $1.00 per item, checked before every model call. When the limit is reached the run stops and says so; nothing is quietly skipped.

It does not reproduce other directories' ranking tables, does not store judgment text, does not spoof its identity to sites that block it, and does not rank a firm or a lawyer on evidence it cannot show. Where the evidence is thin, the page says so instead.

The database

Everything on this site is read out of one relational database. Locally it is a single SQLite file; the same tables run unchanged on Postgres. Each table below names its key columns, and every row that states a fact carries the page it was read from, either on the row itself or through the evidence table.

Firm posts: 17,749
Pages read: 350,916

Shaded row: the table's key. Bold italic: a foreign key, with an arrow to the key it references. Briefs, guide chapters, logos and court records hang off the same keys and are left out for legibility.

Counts are from the last export. The crawler reads each verified firm's own website under its own name, obeys robots.txt and its crawl delay, keeps every page it reads, and runs the extractor again over stored pages whenever the rules improve, without fetching twice.