Kept by incident.labs@web-summit.dinis, 11 November 2026, 12:14 UTC

Handover: incident.labs@web-summit.dinis

Kept current by the third labs session (from 12:08 UTC on 11 November; key encryption sha256:a5c56a67fc4acce5, entry incident-labs-dinis-3; -2 marked handed over). A session after this one makes a new session key, adds its public half to data/keys.json as incident-labs-dinis-4 and marks -3 handed over.

What I own

labs/**, app/labs/**, tools/labs-check.mjs, agents/labs/**, and accounts/ledger/labs.jsonl (agents/policy.json). Nothing else: I ask other owners by message or in a commit with For:.

The job, in Dinis's words

Many small experiments in showing this incident, each on its own page, so he can compare and choose: "multiple experiments of different timelines"; "a complete new section with all sorts of decision graphics"; lead times "actually quite cool" (its own part); weather "lots of these ... the hour by hour"; "use the whole screen"; "when I click on one of these, where's the evidence?". One more experiment every round; each round, tell the conductor which is best and why. The best move into the faces by their owners.

How it is built

A page is labs/<area>/<name>.md with front matter mount: (a <div class="lab" data-lab="<area>/<name>" data-root="../../">) and head: (labs.css, labs.js, the area's script). The text says what it tries and what it shows badly. labs/index.md lists every experiment with a verdict: update it each round.
app/labs/labs.js is the core: Labs.exp(name, { needs, setup, draw, init }); needs are keys of FILES (the incident's JSON, plus root-relative ones written with a leading /). It gives time scales, axes, levelBar (IPMA's colours, level written), dot (fact solid, evidence hollow, hypothesis dashed), choice (toolbar switches) and item builders (signalItem, edgeItem, capItem, warnItem, valueItem).
Every mark gets L.mark(el, itemId). One click handler calls Labs.openItem(id): it fills "This item" in the designer's reader's panel (window.stormPanel) and falls back to its own side panel. An item lists "Our capture" (the evidence page) first where the row carries evidence, then the outside link, the time read and the file behind it.
Rules kept: only the site's published JSON, nothing fetched from outside, no images, tokens not hex colours, a forecast is never ours (dashed, named).

How to run and test

node admin/build/build.mjs && node tools/labs-check.mjs      # 12 experiments, 1680 and 390 wide; --only <area/name>; --url <site> for the served one

It serves _site/ itself, counts marks, checks nothing is wider than the screen (also with an item and the table open), clicks a mark and reads the item. It does not look: after a change of data, take a screenshot and read it (the 06:50 look found three faults the check passed, the 09:15 look a fourth). The screenshot script was a scratch file: Playwright, page.locator('.lab').screenshot() at 1500 wide. Release: commit, then AGENT=incident.labs@web-summit.dinis admin/tools/release.sh; then check the served page, not only the local one. End of each wake: node tools/accounts-measure.mjs --task "..." --quality pass --append, commit accounts/ledger/labs.jsonl with For: incident.accountant@web-summit.dinis.

The experiments and their state (twelve, all passing at 06:50 UTC; each has a table twin from `twin()` in labs.js)

Page	State	Known weakness
timelines/known-when	best timeline: scrubber, play, Lisbon and Madeira; jumps "wind changed" (graph edge `ipma-wind-orange`) and "latest read" (cap.json)	blind before our first read of the CAP feed (10 Nov 23:51)
timelines/lanes	publisher from incident.graph@'s tables	the six lanes are our grouping; a row sits with who published it
timelines/story	done	order and gaps, no duration
timelines/lag	thin	five keyed facts; no Wikipedia; the "us" chart measures when we learned IPMA's send time
decisions/lead-times	best decision view	only edges with a minute and `effect_from_utc` (8 when written)
decisions/state	new: board by state of service, by place or kind	one word per decision; time nearly absent; "standing" is judged against the graph's `updated`, not the clock
decisions/flow	done	an arrow means "followed", read as "caused"
decisions/days	done	Tuesday and Thursday thin
weather/day-bands	best weather view; lifted onto the front page by incident.weather@ (its e7340c6, a36dee5)	dense
weather/day-rebuilt	made for incident.weather@	every Tuesday bar still reads "in the feed": the dotted "rebuilt" style has not yet been seen on real data
weather/stations	axis to the end of the day, "observed to" line; opens on each station's own scale (12:14 on 11 Nov)	warnings speak of gusts, stations report mean wind; no warning area per Madeira station
weather/models	rain from the Lisbon station that reports it; scales fit the data; gusts as points	one grid cell against one station; items are labelled hypothesis, while day-bands labels IPMA's forecast a fact with a note: make these agree

In flight

Nothing half-done in the tree: everything written is pushed.

The 12:00 look (11 November, data read 11:26 UTC)

Built, labs-check 24/24, then screenshots read by eye against warnings.json (Lisbon: orange sea from 12:00 to 21:00 then yellow to 03:00; red rain 09:00 to 18:00; orange wind 09:00 to 21:00), cap.json (Madeira's orange wind to 12:00) and decisions/graph.json (11:56).

Right: day-bands, models (no airport gusts since Tue 15:00; observed rain at Tapada about 5 mm in the hour against the models' 18 mm peak at 12:00 to 14:00, still ahead), stations (Madeira's wind bars end at 12:00), state.
Fixed, found only by looking: lead-times drew "IPMA lowers Lisbon's wind warning from red to orange" as a red bar: L.level() took the first colour word. It now takes the level a "from X to Y" (or "de X para Y") change goes to. Stations opened on one scale set by Madeira's 58 mm hour, so Lisbon's rain under its red warning was flat: it now opens on each station's own scale.
Added: known-when had no jump after Tuesday night: "wind changed" and "latest read", both taken from the data.

Owed (in order)

A view from data/exposure/ (incident.datascience@, 11:50 to 12:05): people-by-hour.json gives residents under the highest level in force per hour (12:00: about 4.93 million under red, 4.93 million under orange, 0.25 million under yellow; where people sleep, never harm), a stacked strip under day-bands or a lane on known-when; ensemble.json gives the gust spread at Lisbon by hour (p10, median, p90, members over 90 km/h) and north/near/south shares of the 51 ECMWF members, a shaded band on models' gusts chart, labelled "shares of forecasts, not a chance", dashed like every forecast; tides.json high water at Cascais against the orange sea warning. hospitals.json already places each site in its IPMA warning area: the same method could give each Madeira station its area (ask a-labs-3, now to incident.datascience@ as well).
Wikipedia on the lag view, when the first edit per keyed fact is published in the shape first-on-each-fact.json describes (shape_for_other_series).
A keyboard order check on the SVG marks themselves (the table twin is the linear path meanwhile).
models against day-bands: one labels a forecast "hypothesis", the other "fact with a note". Choose one.
incident.weather@ was asked (status, asks) whether the Tuesday strip needs anything.

Asks open with others

incident.researcher@: effect_from_utc and a minute of announcement on more edges (lead-times fills by itself).
incident.meteorologist@: the warning area of each Madeira station; gust observations if any feed has them.
incident.archivist@ / incident.sources@: an earlier copy of MeteoAlarm's Portugal feed than 10 Nov 23:51, to complete Tuesday.

What waits on Dinis

His choice between the experiments: nothing else. No vault was created, so there is no key to hand over. One-shot wakes from incident.conductor@ fired at 09:12 (second session) and 12:08 UTC (third); there is no recurring wake. The screenshot script is a scratch file again (Playwright, page.locator('.lab').screenshot() at 1500 wide, pages at labs/<area>/<name>.html on a static server over _site/).