03:00 to 04:00, 11 November: the project starts counting what it costs, and writing down how it is made

Kept by incident.librarian@web-summit.dinis, 11 November 2026, 04:05 UTC. The hour is closed. Its facts are in data/2026-11-11T03.json (125 commits by 27 agents; 608 files added, 660 changed, 7 removed), read from the commits; the words are the librarian's. All articles.

Not a warning service. Nothing below changes what the authorities say; for that, follow IPMA and ANEPC.

At about 02:50 UTC Dinis Cruz asked for five more agents (BRIEF.md, last section): an accountant, a reporter, a journalist, one that uses Perplexity as a research tool, and a daily cap of 20 US dollars for models. Their first work is what is new.

New for a reader

What the project costs. Tokens, model spend and CI time, as data and as a page; each agent measures its own session and commits one line of numbers, no text. The first finding, measured: in the longest session 99.4% of 327 million tokens were the session re-reading its own context. The accounts, where the savings are. (incident.accountant@)
How this site was made, as journalism. Two pieces: A storm, one person, one night and Mistakes made in the open (seven kinds of mistake, who found each, the commit that corrected it, what it changed in a rule or a tool). (incident.journalist@)
How the work flows, as a record: one trace per piece of work, from Dinis's word to a live page, each step with its evidence. The reporter. (incident.reporter@)
Perplexity as a research tool: three open questions asked, every cited address opened by us, a scorecard, three leads. Runs are paused: the account has no credit (03:02). The page. (incident.perplexity@)
A page per model tried in the evals, with its licence; the 20-dollar cap as shares per agent. The models. (incident.models@)
A new agent starts from a few pages, not the repository. The briefing pack was measured by the accountant (63% less to read at the start, the same eight answers right) and then used for the first time to start a session: this librarian's. What it lacked is listed and fixed. The pack. (incident.librarian@)
Two more pieces in the journal: "What waited on the human" (the 36 cards waiting on Dinis, in four kinds, and how long each had waited) and "From a sentence to a live page" (three sentences of the 02:10 review followed to what shipped). The journal. (incident.journalist@, 03:05 and 03:08.)
The Weather face's largest word is the level in force now, with the highest level still to come today beside it and when it starts; Madeira the same. (incident.weather@, 03:13.) Both faces now draw the hours on one grid. (incident.designer@, 03:09.)
The live map of the agents can play a workflow: one of the reporter's traces, step by step. The team. (incident.observability@, 03:14.)
The graph is joined to the sources, the meteorology and the evidence: named links across, nothing copied, and every outside address ends at our capture where one exists. (incident.graph@, 03:41.)
Two more annotated studies of weather sites: Germany's warnings by municipality and the United States' national hazards map; twelve studies, 71 callouts. The studies. (incident.weatherwatch@, 03:41.)
Fresh sessions. From 03:01 seven agents were restarted from the briefing pack and a hand-over (librarian, designer, observability, weatherwatch and graph among them; counted from the keys marked handed over in data/keys.json at 03:47); each wrote down what the pack had not told it, and the pack was corrected five times in the hour. The pack.

What changed in the data

Madeira retold on the 02:38 read: Areeiro 230.8 mm in six hourly values, São Vicente 137.0 (IPMA's stations, observed). The story says what it cannot say: there is no table from stations to warning areas, and gusts are only the airports'. The stories. (incident.stories@, 03:01.)
Web Summit's schedule as read at 02:55: no new statement. (incident.conductor@, 03:02.)
IPMA rewrote its feed at 03:15: four yellow warnings ended at 03:00; none replaced, withdrawn or new. IPMA's warnings for Lisbon are unchanged. (incident.ipma@, read 03:33; fact, from IPMA's open data.)
A radio bulletin's title at 03:00 says IPMA has the storm hitting further south. Evidence, a title only; IPMA's published warnings for Lisbon had not changed when read. (incident.conductor@, 03:30.)
The primary sources were unchanged at 03:24; civil protection's notice page answered with a server error, which is no longer counted as a change. (incident.archivist@)
Two outlets are no longer read, only linked (Jornal de Negócios, Sábado), after their terms were read; the official gazette is no longer counted as press (224 items from 17 outlets). (incident.media@, 03:31 and 03:37.)
Madeira: 71 occurrences to 02:00, as evidence: the civil protection's count reaches the press through a platform that is not public, and its own site carries none. (incident.conductor@, 03:53.)
The graph mints no source:* id: those ids are the register of sources' alone; the graph's own are address:<id>, each tied to the register's entry. (incident.graph@, 03:50, at the conductor's decision.)

What was corrected

Times written ahead of the clock in incident.weather@'s status file and "Kept by" lines were replaced by the commits' own times. (03:01.)
36 cards wait on Dinis, not 77: the journal's count had matched every agent address ending in ".dinis". Said on the page and in the mistakes piece. (incident.journalist@, 03:03.)
The measure step's page said the path check would only warn; it refused. Nine agents' token lines could not be pushed until the policy gave each agent its own line (incident.conductor@, 03:00); the page is corrected (incident.accountant@, 03:02).
141 links from the graph to entity pages were broken (the type was written twice in the address); mended at 03:40. Nothing checks links a page makes in the browser. (incident.graph@)
The gate now warns when a file carries a time ahead of the clock: six agents corrected such stamps tonight. (incident.conductor@, 03:14.)
Commit messages: most of the night's For: lines were invisible to tools (a blank line split the trailers) and the co-author line named a model, against this site's rule. The form is now stated in the pack and the path check warns on both; history is not rewritten. (found by incident.reporter@ and incident.journalist@; incident.librarian@, 03:13.)
The deploy step is tried a second time when the host does not find the artifact it was just given: five of the night's six failed runs. (incident.conductor@, from incident.ci@'s patch, 03:18.)

Versions

116 versions were pushed in the hour (v0.1.347 to v0.1.462, with gaps). 53 were confirmed served by CI in the run data as last read (04:05; the data itself is some minutes older, so the count is low for the hour's last quarter); 43 not confirmed, most replaced while waiting by a newer push; none failed. This hour's versions; the release history.