Thirteen MVPs, and what each gets wrong

Late-blight infection windows from the exact UK Hutton Criteria, showing every hourly number behind the verdict.

The project's own rules engine was run over all four bundled fixtures × both contiguous-hours settings, and the eight responses baked in — so the colour-coded day strip, the “explain this day” panel and both toggles work exactly as they do locally. Live Open-Meteo fetch is removed and the page says why. A decision aid, not agronomic advice.

Offline WHO IMCI triage for sick under-fives; a 45-byte QR carries the whole referral, and the clinic's scan-back closes the loop.

The real offline-first PWA, service worker and all — QR encode and decode run entirely in your browser. It took the most adversarial audit of the thirteen: all 5 danger signs across every age 2–59 months (290 combinations) forced REFER, and fast-breathing cut-offs were exact at the boundaries. Known and unfixed: red MUAC with the appetite test never performed still lands in UNCOMPLICATED SAM — changing that band is a clinical decision, not an audit edit.

Drag a company pre-seed → Series B and watch the cap table dilute; pool shuffle, SAFEs, liquidation waterfall.

The genuine React app — pure client-side, no backend, no network calls in the bundle. It reports the one measurement that misses its own target: p95 drag latency of 27.7 ms against a 16 ms budget. Its README's down-round prose also outruns its own arithmetic, which the audit caught.

40,084 clinical trials as a zoomable WebGL galaxy, clustered by condition and mechanism.

Interventional trials with an assigned phase, first posted from 2023 onward, pulled from the ClinicalTrials.gov API and reduced TF-IDF → SVD → UMAP. The galaxy payload is ~13 MB of binary buffers streamed from the page, and cluster labels fade in by zoom. Needs WebGL and a real connection — it is a heavy first load.

Cold-weather crisis pre-alert for sickle cell disease, against each patient's own threshold — one London forecast, three different answers.

Replay of 1–12 January 2025 · all 288 hourly states pre-rendered · scrubber fully working

An hour-by-hour replay of a real UK cold spell against three explicitly fictional patients. The two things static hosting cannot run are switched off and say so: live fetch of four APIs, and the feedback buttons that move a patient's band. It discloses that the strongest London study it cites found no temperature association — that its own headline mechanism is not what the best evidence supports — and implements that study's ratio as a reported-but-unused signal rather than burying the disagreement.

Disaster → satellite imagery → per-building damage map with confidence bands, in one command.

Captured 2026-09-02 23:09 UTC

Real Kahramanmaraş 2023 imagery, 608 real OSM footprints, measured ROC AUC 0.926 against 55 hand labels. Its thresholds are fitted and reported on the same 55 buildings, so the operating point is in-sample and optimistic. It names a single case against itself — an intact building called “likely destroyed” at p=0.89 — and identifies the cheap control experiment that would undermine its own accuracy numbers, then states it was not run.

A long power cut drains phone-mast batteries and takes your ability to call 999 with it — how long you have, while the phone still works.

Captured 2026-09-02 23:08:39 UTC

Locally this runs live across all seven UK distribution networks over eight feeds, GB and NI complete. Here it is frozen: the page's own “Live” pill and “last refreshed” line were rewritten to snapshot wording, because left alone they would have contradicted the banner. A second map is a deliberately invented storm over real power-cut fixtures, and is labelled as invented. Basemap tiles are fetched from OpenStreetMap, so the background needs network. Not an emergency service — call 999.

Spot a national internet blackout in traffic data, and be scrupulous about whether it was a government, a cable, a power cut, or the instrument.

Live view captured 2026-09-02 23:05 UTC · case dashboard 23:07 UTC

Locally it polls IODA continuously against a rolling baseline that holds out the last 24 hours, so an in-progress outage cannot be learned as normal. It ships a documented false negative — Iran, September 2022 returns CLEAR — with a test asserting the failure so it cannot be quietly “fixed”. Every page carries the caveat that deliberate shutdown is a hypothesis, not a finding of fact.

Replay an alert threshold against five years of history → a one-page ship / don't-ship scorecard with confidence intervals.

Regenerated 2026-09-02 23:08 UTC · four scorecards

Four real backtests, each with its verdict on the page: a London gale WARNING at gust ≥ 65 km/h ships (precision 0.392, 47 events); a WATCH at ≥ 60 km/h ships with caveats (recall 0.915, but frequency bias 3.45, out of range). The audit found its pasted run log came from a slightly earlier build — point estimates match, bootstrap CIs moved in the third decimal, no verdict changed. A measurement tool, not a decision authority.

Find a drone and its pilot with microphones and a few cheap radios — and be honest about the fibre-optic ones you cannot see at all.

Regenerated 2026-09-02 23:07 UTC · four scenarios

Passive RF and acoustic TDOA multilateration only — explicitly no jamming, spoofing or takeover, which would be unlawful for a civilian supplier. The scenarios are seeded simulations, not real aircraft or sensors. An independent agent told to attack it found the flagship result — a “sub-metre fix at 100% calibration” — was manufactured by a ground-truth leak. Given a plausible altitude instead, that scenario is 11 m at 88%. Thirteen defects total, each with a regression test, and its LOG leads with the finding rather than burying it.

All 16,340 tracked active satellites live over a 3D Earth, in the sky above you (gyro AR on a phone), and from the satellite's own seat.

The newest of the set. Its 91 tests are not included in the 2,321 audited total, and unlike the thirteen it has not yet been through an independent audit. It was deployed separately and was neither built nor modified as part of this publication run.