The Zero That Stopped Moving

The check, and everything in it that is a choice rather than a fact

Is the data actually current?

An HTTP 200 with a well formed body and a recent Last-Modified header is not freshness. Two sources in this wave were caught returning exactly that over data that stopped years ago. So every feed here is judged on the newest timestamp inside the data, and the page refuses rather than reports if any of them is older than its own stated cadence allows.

What the HTTP headers say, and why we do not use them. Three of the four NOAA endpoints this page reads send no Last-Modified header at all, and the tide data API explicitly sends Cache-Control: no-cache with an expiry in 1970. Only the sea level trends CSV carries a real one. So the shared kit reports "cannot be determined" for the rest, which is the honest answer, and this page keeps its own clock from the newest row in the data and labels it as ours.

The free choices, named

• Whole hours only, nothing interpolated. NOAA publishes predictions hourly and observations every six minutes. Interpolating an hourly tide curve costs a couple of centimetres near the turn of the tide, which is the size of the effect being measured, so the join is only at exact whole hours and unmatched hours are dropped.
• The quality flags, read properly. NOAA's four flags for preliminary water level are, in order, a count of one second samples outside a three sigma band, then flat, rate of change, and out of range. Only the last three are faults. Rejecting every row whose flag string is not "0,0,0,0" throws away about a ninth of Boston's record for no reason, and the count of what that rule would have discarded is printed above the chart.
• Preliminary data gets revised. Six minute samples arrive marked preliminary and are replaced later by verified ones, sometimes with different values. The record file stores the preliminary reading and, on a later run, the verified one, so revisions are visible instead of quietly overwritten.
• Season matters, and the headline is chosen not to hide it. The 48 hour figure at the top and the August figure in the table both run several centimetres above the twelve month mean, because sea level at Boston peaks in autumn. The headline is a full twelve months for that reason, and the last complete calendar year is printed beside it because it is a different window and gives a different answer.
• Missing and inferred months. Where NOAA has not published a monthly mean it is simply absent from the average rather than filled in, and the count of absent and inferred months is stated wherever an average is shown.
• The epoch is read where it is computed, and hardcoded where it is not. Every number on this page, and the body prose around it, is generated from the live epoch field: the dek above, the epoch sentence, the anchor, the control and both forward claims all change if NOAA changes that field. Six surfaces do not, because they have to exist before any JavaScript runs: the <title>, the meta description, the Open Graph and Twitter cards, the JSON-LD block, and this layer's entry on the ground. Those name 1983 to 2001 as literal text. On the day NOAA moves the field, those six go stale until someone edits the file, and pretending otherwise would be the exact failure this page is about. So the verifier compares the live field against the value written into them on every run and goes red when they disagree, which is the only mechanism that actually forces the edit.
• The planted control marks what it doctored. NOS asks that its information not be modified in content and then presented as official government material. When the plant is on, the numbers in the headline panel are modified NOAA content, so the panel carries a banner, the table's source column stops reading "measured" and reads the planted amount instead, and a screenshot of the panel cannot be mistaken for the real one.

This is not a trend estimate, and here is the arithmetic anyway

.

No line is fitted anywhere on this page. The headline is one published number subtracted from another. The rate above is that difference divided by a time, printed because a reader will do the division anyway, and it is not the same quantity as a regression over a century of data. For what happens when you do start fitting lines to records like this one, see Every Trend You Could Draw.

The datum does not matter, which is the surprising part

Switch NOAA's datum parameter from MLLW to MSL, STND or NAVD88 and both the prediction and the observation move by the same constant, so observed minus predicted is unchanged and everything here survives. The trap is the name: NOAA's "MSL" datum is itself the frozen epoch value, so a reader who assumes it means present mean sea level will get the sign of all of this backwards. That is the same mistake as the fourth radio button above, wearing a label that sounds authoritative.

The one thing this page rests on that your browser cannot check

Everything above depends on a single sentence: that the constant term of NOAA's published prediction curve is mean sea level minus mean lower low water over the epoch. This page reads both of those from NOAA's datums file and it reads the constituent file to show that no constant lives there. Neither of those touches the prediction curve itself. The direct test is to average a whole year of NOAA's own published hourly predictions, over which the periodic terms cancel, and see whether the constant falls out.

The browser cannot do it. This site's worker allowlist caps the tide feed at 48 hours of the curve per request, and 48 hours is not a whole number of tidal cycles, so a mean over it is not the constant. Rather than assert the sentence and move on, the verifier fetches a full year of NOAA's published hourly predictions directly, on its own run and not on yours, and checks two things with it: that the mean of the curve equals the published zero, and that the amplitude NOAA actually synthesises for the principal lunar term is different in two years nine apart, which is the 18.6 year nodal modulation the constituent file does not contain. Both numbers are printed by the verifier and stored in the record on every run. It is the one load-bearing fact this page takes from an instrument other than itself, and it is named here rather than buried.

Running it yourself

node research/the-zero-that-stopped-moving/verify-the-zero-that-stopped-moving.mjs

It imports this page's zero.js unmodified, recomputes every number above from the same feeds, asserts the refusal paths against deliberately broken and deliberately stale fixtures, and appends a line to research/the-zero-that-stopped-moving/record.jsonl, which is append only and carries the score of every forward claim this page has made.