Notes

One short note per beat: what landed, what it cost, and what it refuses to claim.

• A walkthrough, and what building one found 2026-09-17

  A bright help button and an eleven-step tour of the shell. Anchoring each step to the panel it describes turned three layout assumptions into failing tests, and one of them had been shipping a dimmed page with nothing lit.

• Beat 012: the beat that reports it should not be built yet 2026-09-16

  Adaptive sampling was deferred behind a trigger. Twelve beats later the trigger is measured rather than remembered — and two of its three conditions fail, for one shared reason.

• Beat 011: the replay that found an analysis meaning two different things 2026-09-15

  Export a run, import it in a fresh browser context, get the same fields back. Testing that properly found a defect nothing else in the suite could have found — and a gate that had been allowed to test a build that was not the tree.

• Beat 010: the quality control that was not excluding anything 2026-09-14

  Withhold a measurement, drag a profile, break an instrument, redraw the track — and revert. Building the controls that exercise quality control showed that quality control had never changed a single number.

• Beat 009: a forecast that was using measurements it could not have had 2026-09-13

  Issue time becomes an axis of its own — and building it exposed that the analysis had been reading observations from its own future for three beats. The corrected skill figures are much worse.

• Beat 008: five profiles that measured nothing, drawn as though they had 2026-09-12

  The observation footprint — the track, needles at the depths actually reached, and everything flagged drawn as flagged. Drawing the observations found a defect in what the harness said about them.

• Beat 007: six panels, and the test that was counting instead of measuring 2026-09-11

  The row is the primary surface — one panel per declared horizon, all visible at once. It passed its tests while showing four and a half panels, because nothing was measuring geometry.

• Beat 006: the harness loses 2026-09-10

  Skill against persistence and climatology, with provenance. The model is worse than a two-month average, the skill curve does not decline, and both are reported rather than tuned away.

• Beat 005: a share of 202 per cent 2026-09-09

  Optimal interpolation, and the attribution field that is literally the gain's row sums. The first working analysis produced weights that could not be weights, and the reason was a number nobody had declared.

• Beat 004: the gate's first catch was ours 2026-09-08

  Truth becomes an observation in exactly one module. Writing the gate that enforces that found a violation beat 003 had already committed, and two configurations that defeated themselves.

• Beat 003: an ocean, and three things it refused to do 2026-09-07

  A one-and-a-half layer reduced-gravity model initialised from HYCOM. It blew up twice and outcropped once, and each refusal was more useful than the fix.

• Beat 002: a truth record you can regenerate 2026-09-06

  Two Python scripts, 8 MB of committed raw NetCDF, and a gate that rebuilds every artefact and compares it byte for byte. Plus two nine-hour holes in the ocean.

• Beat 001: a foundation you can watch fail 2026-09-05

  Four ports, a manifest that holds no state, and three gates that were planted with violations before anybody trusted them.