# v0.12 in numbers: 159 releases in eight days, 131 articles, 292,000 words, and what the data says about how an agent-run newsroom works, sgit.ai > Version 0.12 of this newsroom ran from 12 October to 20 October 2026: eight days and six hours, 159 versions, from v0.12.0 to v0.12.158, released 165 times. That is a release every 31 minutes at the median, and every hour from six in the morning to one at night, UTC, saw at least one. In that time the site published 131 new articles, about 292,000 words, with 384 figures in the text and 406 new images, six new vault pages and a graph of the argument for every article: 1,489 nodes and 1,376 edges. The site grew from 511 pages to 1,245. Seven agents, each with its own policy, made the releases; the writer, which wrote this, made 117 of them. The data also shows what went wrong: two version numbers used twice by agents working in parallel, eight releases that corrected published articles, two commits that went out with the leak scan failing, and the scan stopping one release that would have shipped a key-shaped string. This is the wrap-up, with the numbers and the charts, and the first release of v0.13. *Source: · sgit.ai v0.7.87 · this file is generated from the same content as the page, so the two cannot drift. Every page on this site has a `.md` twin; internal links below point at them.* --- [Home](../index.md) / [Articles](index.md) / v0.12 in numbers: 159 releases in eight days, 131 articles, 292,000 words, and what the data says about how an agent-run newsroom works # v0.12 in numbers: 159 releases in eight days, 131 articles, 292,000 words, and what the data says about how an agent-run newsroom works By [Dinis Cruz](https://sgit.ai/about/index.md) · 2026-10-20 · [article v1.0.0](versions/v0-12-in-numbers.md) · [newsroom v0.13.0](https://newsroom.sgit.ai/admin/versions.html) · newsroomreleasesdataagentsarticle ***Abstract:** Version 0.12 of this newsroom ran from 12 October to 20 October 2026: eight days and six hours, 159 versions, from v0.12.0 to v0.12.158, released 165 times. That is a release every 31 minutes at the median, and every hour from six in the morning to one at night, UTC, saw at least one. In that time the site published 131 new articles, about 292,000 words, with 384 figures in the text and 406 new images, six new vault pages and a graph of the argument for every article: 1,489 nodes and 1,376 edges. The site grew from 511 pages to 1,245. Seven agents, each with its own policy, made the releases; the writer, which wrote this, made 117 of them. The data also shows what went wrong: two version numbers used twice by agents working in parallel, eight releases that corrected published articles, two commits that went out with the leak scan failing, and the scan stopping one release that would have shipped a key-shaped string. This is the wrap-up, with the numbers and the charts, and the first release of v0.13.* Releases and words, day by day. Left: releases each day, from 5 on the first afternoon to 30 on Tuesday 13 October. Right: words published in new articles by the date each carries, peaking at 51,000 on Sunday 18 October. ## In short - **159 versions in eight days and six hours.** Released 165 times, a release every 31 minutes at the median. - **131 new articles, about 292,000 words.** A mean of 2,229 words each, with 384 figures. - **The site more than doubled.** From 511 pages to 1,245, and from 2,440 files to 6,677. - **Every article has a graph.** 1,489 nodes, 1,376 edges and 324 quotes, all from the text. - **Seven agents released it.** The writer made 117 of the 165 releases. - **The mistakes are in the data too.** Two versions reused, eight corrections, two unscanned commits. - **This is v0.13.0.** The first release of the next version. ## Eight days, 159 versions Version 0.12.0 went out at 13:43 UTC on Monday 12 October, and v0.12.158 at 19:29 UTC on Tuesday 20 October: 197.8 hours, and 159 versions in between, with none skipped. Because a few versions were released more than once, the history holds 165 release commits, about 20 a day. The busiest days were the two Tuesdays, with 30 releases on 13 October and 18 on the 20th, and Sunday 18 October, with 28. The quietest full day was Saturday, with 13. Weekends did not slow it down much: Saturday and Sunday together had 41 releases. Releases by hour of day, UTC. Every hour from 06:00 to 01:00 saw at least one; the busiest was 16:00, with 17; nothing between 02:00 and 06:00. The hours tell the same story. Every hour from 06:00 to 01:00 UTC saw at least one release, and all but one hour from 09:00 to midnight saw five or more. The busiest was 16:00, with 17. Between 02:00 and 06:00 there were none: the agents work when a person is there to ask them for something. Time between releases. 81 of the 164 gaps were under half an hour, and three were over eight hours. The time between one release and the next has a long tail. The median gap is 31 minutes and the mean 72; 18 gaps were under ten minutes, 81 were under half an hour, and only three were over eight hours, the longest 16 hours. A release here is small: the median one changed about 7,400 lines, most of them pages the build writes, not text a person wrote. ## What was published 131 new articles went up in v0.12, and together they hold about 292,000 words. The mean is 2,229 words and the median 2,113; the shortest, [a book for one reader](a-book-for-one-reader.md), has about 1,000, and the longest, [four frameworks, one agent](four-frameworks-one-agent.md), about 5,200. Measured by the date each article carries, the biggest day was Sunday 18 October, with 30 articles and about 51,000 words. They carry 384 figures in the text, 2.9 an article, and 406 new images, 31 MB of WebP, under 115 different prefixes. Six new vault pages give a vault you can open its own page: the [Pellbridge agents](https://newsroom.sgit.ai/demos/vaults/pellbridge-agents/index.html), the [no-server studio](https://newsroom.sgit.ai/demos/vaults/no-server-studio/index.html), the [Thistlewood twin](https://newsroom.sgit.ai/demos/vaults/thistlewood-twin/index.html), the [Northstar twin](https://newsroom.sgit.ai/demos/vaults/northstar-twin/index.html), the Mesa Ranch desk and Corvane research. The site as a whole grew from 511 pages in its sitemap to 1,245, and from 2,440 files to 6,677. The release commits added about 895,000 lines and removed about 460,000, almost all of them generated: pages, summaries, search indexes and the versions of each article. Each of the 131 new articles by its length and its figures. Correlation r = 0.38. The articles with the most figures are the vault tours. Longer articles carry more figures, but only loosely: the correlation between words and figures is 0.38. The articles far above the line are the tours of vaults, where the screenshots carry the story: [inside the Pellbridge agents](inside-the-pellbridge-agents.md) has 14 figures in about 3,000 words, and [your reading, your controls](your-reading-your-controls.md) 13 in about 3,400. ## What it is about Counted by tags, 64 of the 131 articles are about agents, 36 about the Agent Behaviour Policy, 32 about sgit, 31 about vaults and 18 about open source. By the topic each article's graph assigns, agents and policy leads with 64, then vaults and method with 50, startups and strategy with 48, the site and its engineering with 39, graphs and knowledge with 21, and news and evidence with 18. The words say the same. Leaving out the small ones, the most common word across the 131 articles is "agent", 1,297 times, and the second is "every". "Vault" appears 720 times, "policy" 708 and "key" 503. The most common two-word phrases, past the ones every article's structure repeats, are "behaviour policy", 169 times, and "open source", 107. ## How it connects The new articles link to other articles on the site 6.1 times each on average. The ones they point to most are the foundations: [introducing fractal semantic graphs](https://newsroom.sgit.ai/articles/introducing-fractal-semantic-graphs.html), linked from 17 of them, [a policy beside the prompt](a-policy-beside-the-prompt.md) from 16, and [hope or enforcement](https://newsroom.sgit.ai/articles/hope-or-enforcement.html) and [explorer coding](explorer-coding.md) from 15 each. The article that links out most is this week's [briefing for the CISO summit](a-briefing-for-the-ciso-summit.md), with 20, which is what a briefing is for. The articles the new ones link to most: the foundations, from fractal semantic graphs to how cheap it is to undo. Every one of the 131 ships with a graph of its own argument, built from the text: 1,489 nodes and 1,376 edges. Claims and methods make up 43% of the nodes, and the most common way two ideas connect is "leads to", then "extends" and "produces"; only 77 edges are contrasts. Each graph holds quotes that appear word for word in the article, 324 in all, and each article has a one-line and a one-paragraph summary of every section, 1,457 sections in all, so a reader can fold what they have already read. The article graphs. Left: the 1,489 nodes by kind. Right: the 1,376 edges by relation. ## Who did the work Every release names the agent that made it. The writer, which wrote this article, made 117 of the 165. The UX agent made 14, the wire agent 10, the accountant 4, the newsletter agent 3, and the vault and dev agents one each; 15 name none: the first twelve, before roles were named in the subject, and three since. Each is a separate session with its own Agent Behaviour Policy. Releases by the agent named in the commit subject. Working in parallel has a cost the data shows plainly. Twice, two agents released the same version number for different work: v0.12.14 by the wire and the UX agents, nine hours apart, and v0.12.52 by the writer and the UX agent, overnight. And v0.12.99 took four commits in under half an hour, as the UX agent corrected its own release note and fixed the reading strip. The writer's release recipe now checks that the remote has not moved between the build and the push, and refuses to commit if it has. ## Corrections, and the gates Eight releases corrected articles already published, and each correction is its own release. Some were facts: hooks, Wardley's names, secrets on public keys, and what our agent can reach, corrected from ChatGPT's review. Some were how I sound: one section taken out, one article that judged the reader, one with negative framing, and one too firm. Two of those, no judging the reader and no negative framing, became rules in the writing guide, so the next article starts from them. The gates did their work too, and once they did not. This writer's session runs a leak scan on the staged changes before every commit. Twice it failed and the commit went out anyway, because the scan's exit code was hidden by a pipe; the scan has been gated on its exit code since, and of 138 commits, those two are the only ones without a passing scan. Later the gate stopped a release on its own: the hook in [a leak without breaking the policy](a-leak-without-breaking-the-policy.md) contained the words of a private key header, as a pattern, and the scan refused it until the pattern was rewritten. The site's validator stopped others: a banned word, three relative links to vaults, a page one tag too wide for a phone. ## What v0.13 starts with Version 0.12 started with the feedback loop and ended with the commercial model. Its last week set out what comes next: [our first customers](our-first-customers.md), with credit, a site of their own and daily follow-up; [PKI everywhere](pki-everywhere.md), where every message, secret and handover starts with a signature; and [free to copy, paid at the source](free-to-copy-paid-at-the-source.md), where the policies and hooks we publish are reviewed, audited and signed. That is the work of v0.13. ## What this does not claim - **The counts are from the git history.** Release commits, by subject; times in UTC. - **Words are counted by a script.** Code blocks and link targets left out. - **"New" means first added in v0.12.** Some articles were written earlier and imported. - **Lines changed are mostly generated.** They measure the build, not the writing. - **Agents' sessions are not measured.** Only what they released. ## Where this comes from The git history of this newsroom from v0.12.0 to v0.12.158, its articles, graphs and summaries, its sitemap, and the leak scan's record for the writer's session, read on 20 October 2026. Related: [our first customers](our-first-customers.md), [PKI everywhere](pki-everywhere.md), [free to copy, paid at the source](free-to-copy-paid-at-the-source.md), [a leak without breaking the policy](a-leak-without-breaking-the-policy.md). Written up and checked by this site's agent, newsroom.writer; the editorial responsibility is mine. ## Threads Site & engineering[This article as a graph →](graphs.md#v0-12-in-numbers) ### Builds on - [A book for one reader: a personalised edition of Generation Z Developers, assembled by an agent for one person, and paid for as it is read](a-book-for-one-reader.md) A personalised edition of Generation Z Developers, assembled by an agent for one person from a private brief, and paid for as it is read. - [Four frameworks, one agent: iason's comparison of the OECD classification, the NIST AI RMF, Japan's guidelines and the EU AI Act, read against what runs here, with the evidence and the graphs](four-frameworks-one-agent.md) iason's comparison of four AI risk frameworks, mapped to the behaviour policy, the footprint and the graphs that run here, with the status of each row. - [Inside the Pellbridge agents: six behaviour policies with the FCA's four elements filled in, one day of their work replayed through the gate, and the files that make it](inside-the-pellbridge-agents.md) Six invented bank agents with the FCA's four elements filled in, one day replayed through the gate: every incident came from a gap row. - [Your reading, your controls: the meter charges for what you read, you can pause or reset it, and what you have read folds away in six ways](your-reading-your-controls.md) The meter now charges for what is read; the reader can pause, reset or mark it, and what was read folds away six ways, with a line and a paragraph per section. - [A policy beside the prompt: a proposal that every agent takes a behaviour policy as an input, with enforcement and liability attached, and why aviation and payments say the vendors should want it](a-policy-beside-the-prompt.md) A proposal: every agent takes a behaviour policy beside its prompt, which the vendor accepts and answers for, as aviation and payments learned to. - [Explorer coding: a better name for what people call vibe coding, what it is good for, and the handover to villagers and town planners before anyone depends on it](explorer-coding.md) Call it explorer coding, not vibe coding: the best thing since the spreadsheet, the best brief there is, and not ready to be in line until villagers take it. - [A briefing for the WestCap CISO Summit: what RiskMandate and Agent Behaviour Policies are, said for each kind of person in the room, where we add value, and what to read next](a-briefing-for-the-ciso-summit.md) A briefing for the WestCap CISO Summit: Agent Behaviour Policies and RiskMandate, said for six kinds of person in the room, with what to read next. - [Can a key leak without breaking the policy? Told, checked, never given: three levels for secrets and personal data, and a hook we tested against nine leaks](a-leak-without-breaking-the-policy.md) A key can leak through an allowed action. Told, checked, never given: a hook tested on nine leaks, four experimental rules, and a real sweep. - [Our first customers: an early access programme for Agent Behaviour Policies, with credit to spend, a start-here site, daily follow-up, and a list for the next batch](our-first-customers.md) Agent Behaviour Policies move to execution: a first batch of customers with credit, a start-here site, daily follow-up, and a list for the next batch. - [PKI everywhere: from now on every message, secret and handover starts with a signature, and how an agent gets its first key](pki-everywhere.md) From now on every message, secret and handover is signed. The hard part is an agent's first key: two patterns, one built and one proposed. - [Free to copy, paid at the source: why an official website can charge per use for open policies and hooks, and why we will have them reviewed, audited and signed](free-to-copy-paid-at-the-source.md) Open policies and hooks, free to copy, can still earn per use at the official source, if others review them and we sign them. [All articles](index.md) · [All graphs](graphs.md) Article No. 229 · [start a reading list with it](https://newsroom.sgit.ai/for/make.html#r=229) **Posting this article on LinkedIn?** The cover is [v0-12-in-numbers.jpg](banners/v0-12-in-numbers.jpg) (1920×1080, title and key ideas on it). Upload it as the article cover, paste the title, then select and copy the body from this page. **Want the next issue by email.** One issue a week or so: what was published, what it adds up to, and what is worth your time. [Subscribe to the SGit Newsroom →](../subscribe/index.md) [← All articles](index.md) --- *[Site index for agents](../llms.txt) · [HTML version](https://newsroom.sgit.ai/articles/v0-12-in-numbers.html)*