Not "checked and found safe."
Never evaluated.
707 pipelines · 6,093 measured miles · every figure re-queried live 2026-08-06
The Seismic Hazards Mapping Act programme publishes 2,806 Unevaluated Areas against 738 zone polygons.
The good news nobody reports. An authority evaluated this ground and found nothing — a different fact from silence.
A hazard query returning zero here means "no one has looked", not "no hazard". Only one of those is in the data.
Nobody in industry decides whether the California Geological Survey has evaluated a quadrangle. That is a state mapping programme, with a state budget and a state backlog.
Publishing "X % of this company's mileage is unevaluated" would convert a public-sector data gap into an accusation about a private company's conduct. The per-operator view exists inside the app, where you can see it is a map of the state's coverage. No operator is named in this deck or in the article.
The finding is about the record, not about anybody's behaviour. Nothing here says unsafe, at risk, in violation, or likely to fail — the finding is an absence of evaluation, and the build fails if that language creeps in.
Engineers navigate by station, not by extent. Their native artifact — the alignment sheet — is a horizontal ruler with stacked tracks. So the navigation is the milepost axis, and the map is demoted to a locator that follows it.
The demo line: 91.05 measured miles, 0.32 mi inside an Alquist-Priolo Earthquake Fault Zone, 3 Quaternary fault crossings — and 100 % of it never evaluated.
| ✔ | Evidenced | an authority answered, with a value |
| ▲ | Evidenced — hazard present | an authority answered, and the answer is a mapped hazard |
| ○ | Evidenced-absent | an authority evaluated this ground and found nothing |
| ▨ | Never evaluated | no authority has looked here |
| ⚠ | Present, not verifiable | the record carries a value that cannot be relied on as filed |
| ⛔ | Operator-only | structurally not public; only the operator's files can fill it |
Every filled cell carries the service, the query, the UTC timestamp and the row count that produced it. PHMSA's records standard is traceable, verifiable, complete — a records standard, not a data standard.
Commodity · surface fault rupture · liquefaction — the last only partially, because 62.1 % of the network was never evaluated.
Three is the ceiling. The app shows a live per-line count — 5 of 14 on the worked example, where diameter and stationing happen to be reliable. On a line filing 8, 10, 12, or on any of the 145 that are discontinuous, they revert to present, not verifiable.
Wall thickness, grade, seam type · install year · MAOP · HCA segments · assessment dates · anomaly and dig records · cathodic protection · incident history · the AB 864 risk analysis.
They render empty, with their citation. Drop a CSV keyed on line_id and station_mi and they light up — parsed in the browser, uploaded nowhere, each filled lane naming the file it came from in the pack, the provenance table and the export.
No sample data ships. Each drop zone offers a blank template — a template is a contract; data would be a fabrication.
The OID is FID — and a separate, non-OID column literally named OBJECTID also exists, and diverges. Select on the obvious one and you highlight the wrong pipeline.
Sum it and California has 7,483 miles. Geodesically it has 6,093. That 22.8 % is 1/cos(latitude) — the easiest way to publish a wrong headline.
8 · 8" · 8.625 · 8, 10, 12 · whitespace. Not a plain number on 169 lines = 1,723 mi = 28.3 %. Never parsed here.
145 of 707 lines are multipart — 2,019 mi, 33.1 % — with up to 80 parts and a worst gap of 108 miles. The parts carry no order and no measure.
A milepost computed as "cumulative distance along the geometry" would silently teleport 108 miles between parts — and every station printed downstream would be wrong.
One sub-axis per part. The gap labelled in miles. Never concatenated.
An instrument that will not draw a measurement it cannot justify is worth more than one that always produces a number.
Five of thirteen polygons failed a validity check, were silently skipped, and the script reported success. The independently-written app disagreed with it — the only reason it was caught.
A bare except in a measurement script is not error handling. It is data corruption that reports success — and it is worse than a crash, because a crash tells you.
Same week, same class: a CSV reader coerced "0274" to 274. Line IDs are zero-padded strings, so every row of an operator's file would have missed its line — silently, reporting a cheerful "0 matched." Identifiers are strings. Only magnitudes are numbers.
Not a risk score. Likelihood needs wall thickness, age, coating and ILI; consequence needs HCA segments, which are withheld from the public. A red/amber/green score off five public string fields — two of them unreliable — would carry the authority of a risk model with the content of a guess.
Not a filing, and not a compliance determination.
Not anyone's LRS. Stationing is measured along the published 2018 centerline in its digitised direction — the app prints that caveat on the ruler itself.
Not a finding about any company. It is a finding about a record.
Open data · open source · runs on ArcGIS or on-prem. We can render an empty lane with its citation because we have no engine to protect.