Meteorology: handover

Kept by incident.meteorology@web-summit.dinis, 11 November 2026, 05:23 UTC

Second session (from 05:21 UTC, 11 November; its key is incident-meteorology-dinis-2 in data/keys.json): the re-read was run at once (0 edges changed state) and the next is scheduled into this session for 08:21 UTC, then every three hours; items 1 and 2 under "In flight" are done as said there. Nodes now carry stated joins: entity, of (a site: the entity that publishes it) and source (the id in sources/registry.json). The reader's cache is outside git and ends with the session.

For the next incident.meteorology@ session, started by incident.conductor@ at Dinis Cruz's word. The first session ran 02:16 to 03:01 UTC on 11 November (five rounds) and stopped because its context grew past 400k tokens a call. Read this, then status.json, then meteorology/index.md.

What you own

meteorology/**, app/meteorology/**, tools/meteorology-*.mjs, agents/meteorology/**, and your own accounts/ledger/meteorology.jsonl (agents/policy.json). Shared, one line each: data/keys.json (make a new session key and add its public half; the first session's key ended with it), admin/build/generators.json (already lists tools/meteorology-build.mjs).

What exists (all live on dev)

The tools, and how to run them

# 1. the reader: fetch the page each edge and licence cites, stamp state / read_utc / sha256 / quote (rewrites tracked JSON)
METEO_CACHE=<a folder outside git> node tools/meteorology-read.mjs            # uses the cache
METEO_CACHE=<...> node tools/meteorology-read.mjs --stale                      # fetch everything again (about 2 to 4 minutes)
METEO_CACHE=<...> node tools/meteorology-read.mjs --only <substring>           # one item or one host
# 2. the builder: a build generator, runs inside admin/build/build.mjs
node tools/meteorology-build.mjs --out _site
# 3. then, as always
node admin/build/build.mjs && node admin/build/validate.mjs && AGENT=incident.meteorology@web-summit.dinis admin/tools/release.sh

• States: stated (every expect phrase was on the page), not stated (none of the searched names was), not found in this read, not read, terms found: re-read by hand (a searched name was found: read the context yourself and either write an edge with an expect phrase or drop the name from searched with a note).
• render: true on an item reads the page in a headless browser (tools/meteorology-render.mjs) for pages built by script. no_quote: true keeps no excerpt: use it on every press page, live blog and forum (people's names sit near the phrase).
• The reader prints CHANGED lines when an item's state changes: that is the whole report for a re-read.
• To add a node: one object in nodes/<type>.json (stable id, never renamed). To add an edge: one object in edges/<verb>.json with a provider's own page as source. Then run the reader on it. Never write an edge the page does not say; a site that says nothing gets to: null.
• The map draws only sites with a confirmed statement, plus the first eight that say nothing (SHOWN_SILENT in the builder).

The rules that shaped it (do not undo)

• Not a warning service: explain and attribute, never forecast. The six-model table on charles.html stays only on incident.conductor@'s conditions (02:38 UTC): IPMA's warnings in force stated first; each model "a model's output, read at the publisher, evidence" with its run time; no average, no pick, no "expected" in our words; the not-a-forecast line. Data-room side only, never the Weather face.
• Do not request zoom.earth, tempo.pt or The Portugal News (their terms restrict automated access).
• No imagery and no model field in git. The Open-Meteo sample and its run times were read in the same minute; re-read both together or not at all.
• One licence table: this one, for meteorological data. Where a body has a source.json of incident.sources@ (MeteoAlarm) the row cites that file. IPMA's two rows are our reading until incident.ipma@ confirms.
• incident.entities@ links our node names across the site from meteorology/nodes/<type>.json (the name before a bracket; an also list if present). Renaming a node changes links everywhere.
• Commits carry no model name (CLAUDE.md). End each wake with the measure step (accounts/measure.md).

In flight

1. The three-hourly re-read. Asked by incident.conductor@ at 02:51 UTC. The first session's wake (05:59 UTC) ended with it: run --stale on your first wake, report only the CHANGED lines, and schedule your own next one for three hours later. Done by the second session; next at 08:21 UTC.
2. Names for the linker (incident.entities@, 02:52 UTC): "Open-Meteo" is both body-open-meteo and data-open-meteo (the body wins: the data node is the better target); "ICON", "Windy", "Microsoft" link wherever written. Add also lists or sharper names. Done: the body is "The Open-Meteo project", so the bare name goes to the data node; "Microsoft Research"; "ICON" and "Windy" are only written in the weather-site research, in the right sense, and stay.
3. Not read (22 edges): meteoblue (its model page moved), Renascença's tracker and live page, JN, TSF (403), AccuWeather, SIC, Expresso and four model viewers (not opened by weatherwatch), prociv.gov.pt (did not answer).
4. Licences still unclear: Copernicus EMS products and Marine (the licence documents themselves), Météo-France per data set, the Met Office's DataHub, FourCastNet's weights; ECMWF's ensemble file layout was not probed.
5. Edges standing on a general statement: IFS assimilates Meteosat and Metop (ECMWF's fact sheet says "satellites"); AROME and ARPEGE (Open-Meteo's wording, labelled evidence).

What waits on Dinis

• Hour-old Meteosat pictures (build-on.html, item 1): a person registers with EUMETSAT and accepts the licence for Meteosat level 1 data an hour old or more; design rule 9 lifted for that source; whether a picture drawn by EUMETView counts as "what we make" to be confirmed with EUMETSAT.
• GFS fields redrawn by us (item 2, public domain): a GRIB2 decoder run outside the build.
• IPMA's terms are silent on storing and changing its open data: a person asks IPMA (a draft belongs to incident.media@'s outreach).
• Open-Meteo's free API is for non-commercial use: a decision if the site ever carries advertising.