The workings
How this instrument keeps itself honest
The machinery that regenerates these figures, the checks that run against them, and a complete list of everything this instrument has got wrong and corrected. An instrument that cannot admit its own miscalibration cannot be trusted about anything else.
This instrument has been running for 91 days.
Room 09 of 10
The clockwork
What runs, and how often
Every figure on this instrument is read from the database when the page is built. Nothing is assembled in your browser, and nothing is injected into a stored file. Scheduled work keeps the underlying rollups current; the pages themselves need no schedule to be correct.
The truth spine
The checks that run against us
A set of automated checks that compare what this instrument publishes against what the database actually holds. Their value is entirely in what they catch, so what they caught is published below.
What the spine can and cannot do
In plain words
The checks can tell us when a number on the page disagrees with the database. They could not tell us when a number on the page had quietly stopped being read from the database at all, because from the outside a stale number looks exactly like a fresh one.
Precisely
The previous build injected live values into a static file by pattern substitution. A pattern that stopped matching left the old value in place and returned no error. Four such figures were found on 7 August 2026. The checks now assert that each figure was actually read at generation, rather than only that it looks plausible.
Corrections
Everything we have got wrong
In full, dated, in the main text and not in a footnote. Where a fault is not yet fixed, it is listed as outstanding rather than omitted.
| Date | What was wrong | What it says now | Status |
|---|---|---|---|
| 7 Aug 2026 | The page was a static file frozen on 19 June 2026 with live values injected by pattern substitution. Patterns that stopped matching left June figures on the page, indistinguishable from live ones. | Every figure is read from the database when the page is built. The substitution layer is gone. | Fixed |
| 7 Aug 2026 | Four different figures were all described as "successful", meaning four different things. | Each has a single name and a single rule, set out in The Story and The Dictionary. | Fixed |
| 7 Aug 2026 | Recorded sessions were published as "human sessions". The session table classifies every row it has ever written as human, so the label could not be earned. | Published as recorded sessions. Only the AI-referred subset is described as people. | Fixed |
| 8 Aug 2026 | The tracker wrote down a status code predicted from the URL path instead of the one the server returned, and the prediction overrode reality. Refusals were recorded as successful reads, passed the served-read filter, and inflated citation counts. On the strength of them a blocked crawler ranked first by read volume. | The real status wins; the prediction survives only as a fallback where no real status exists. 16,475 affected rows are identified and counted in The Story. They are not rewritten, because editing a log to make an old claim true would be the worse fault, and they are excluded from every total by the withdrawn-from-credit rule. | Outstanding — read counts published before 8 August 2026 are overstated, and totals spanning that date are not comparable across it. |
| 7 Aug 2026 | Amazonbot was described as blocked while our own server was still returning content to it. | The block is real and dual-mechanism, effective from 30 Jun 2026, with a server-level refusal added on 3 Aug 2026 as belt and braces. The number of times it was served beforehand is published in The Machines and its whole history stays visible. | Outstanding — the edge still answers some Amazonbot requests from cache without reaching our server, so we cannot yet claim the refusal is total. |
| 7 Aug 2026 | Megabyte figures for images had been baked into the June file and had stopped updating. | Read live from the asset-knock table, which is the only table that records images. | Fixed |
| 7 Aug 2026 | A day count of 57 was hardcoded in two places and never incremented. | Computed from the launch date at page generation. No template holds a day count. | Fixed |
| 7 Aug 2026 | 2,518 public pages and 82,632 image records were published as constants that no query returned. | Counted from the content tables at generation. Where the old figure cannot be reproduced, that is stated. | Fixed |
| 7 Aug 2026 | All five lifetime columns claimed a window starting 10 May 2026. Three of the five source tables did not exist until June. | Each row states the real first row of its own table. | Fixed |
| 7 Aug 2026 | Our own verification probes appeared among the most frequent questions asked, unlabelled. | Separated into their own column and subtracted from every demand figure. | Fixed |
| 7 Aug 2026 | Per-agent outcome splits were drawn from a rollup that ingests every citation row twice, so splits exceeded the counts they were splitting. | Every per-agent column is drawn from the citation table itself. The double-counting rollup is not read anywhere on this instrument. | Fixed |
| 7 Aug 2026 | Sealed daily study rows recorded about two and a half hours of each day and were published as closed-day totals. | The sealed totals are no longer used for any figure. Live source-table figures are published instead. | Outstanding — the sealed chain itself has not been re-sealed, because rewriting an append-only record needs a decision first. |
| 7 Aug 2026 | Door-knock capture was gated to AI surfaces, so ordinary crawler volume was never recorded. Googlebot appears at about one per cent of its true request count. | The gate is open. Historic figures remain undercounted and every affected table says so. | Outstanding — data before and after 7 August 2026 is not comparable. |
| 7 Aug 2026 | Image capture stopped after 20 June 2026 and has run only sporadically since. | Published as a floor, with the days on which capture ran listed in full. | Outstanding — capture has run on 13 of 63 days. |
| 7 Aug 2026 | Only 33 of 139 real orders reach the event tables the commerce columns are built from. | No conversion rate is published anywhere, and the order column carries the caveat. | Outstanding — the event pipeline needs rebuilding. |
Every entry states its date in the first column. Corrections are never removed once published.
JSON mirrors
The same figures, for machines
Every long table on this instrument defaults to its first fifteen rows and offers the rest from a JSON mirror. The mirrors carry the same figures under the same rules, and their paths never change even when the prose around them does.
Machine surface paths are a contract. Prose and labels on these pages may be rewritten at any time; the paths a machine has learned are not. Every mirror carries the same four attributes as the page: the source it came from, the window it covers, the rule that produced it and whether it is live or frozen.
The route authority in The Doors lists every published surface with its expected response and the result of its most recent probe.
