We've had think tanks that argue, journalists who chase the press release, audits that come 3 years late, and government dashboards that measure their own homework.
We never had a citizen-held, append-only, binary scoreboard that exists before the outcome.
That's the missing layer.
Think of the stack:
For 10 years we tried to jump from 2 to 4 and wondered why nothing closed. You can't hold power accountable for a test that was never written down.
The register doesn't need to win the argument about Kamloops, or LNG, or immigration levels, or dual-use with China. It just needs to force:
"Here's what you said would count. Here's who you said owns it. Here's when we will check a primary source. If it's not there, it's FAILED, and here's the layer that blocked it."
Once that's public and fingerprinted, three things change that didn't before:
It's not the house. It's the foundation no one poured.
As the last line of The unclosed file puts it:
"...until someone starts writing down the names and the dates."
That's what this is.
The register is a ledger of public commitments and whether they were kept by the date they were tested. This page sets out how a promise gets in, how its test is set, how its scoring date is chosen, how hard cases are ruled, and how the register relates to the desk's essays.
A promise qualifies when all three hold:
Every row records why it qualified.
Each quarter the desk also adds a sample of routine commitments from outside its usual themes, including ones that look likely to pass. This is a check on the desk's own selection: if the register only held promises the desk expected to fail, its failure count would say more about the desk than about the government.
The register is not a census of all promises. It is a record of the commitments the desk chose to test, under the rule above, and every score should be read that way.
Each new test carries a short note explaining why the bar sits where it does.
Each new test also carries a looser and a tighter variant. Only the main test is the score. The two variants are scored for information only, so a reader can see how much the result depends on where the bar was set.
A row that fails the main test but passes the looser one says so on the row.
The scoring date is chosen in this order:
Any exception to that order is justified in writing on the row.
Where the dates come from now. Of the 397 rows in the method side-file on 6 Nov 2026, 212 (53.4 per cent) carry a desk-default date, 120 (30.2 per cent) a government date, 30 (7.6 per cent) a statutory date, and 35 (8.8 per cent) a date basis that is not recorded. These counts come from each row's date_basis label in data/row-method.csv, which is built mechanically from the date source on the row. None of the 212 desk dates in this file are marked as an exception to the 12-month default.
When a case falls between the rules, the desk makes a ruling. Each ruling is logged once, in a public precedents log, with its date and the rows it was first applied to, and is applied the same way every time after.
A change to a ruling goes in as a new dated entry. It is never applied quietly to old rows: rows already scored keep the score they were given under the ruling in force at the time, and any later view is added as a dated note.
The register records the scores. The essays argue the desk's reading of them.
The scores stand without the essays. A reader who disagrees with every essay the desk has written can still check each row against its own test, its own date and its own primary source.
Method adopted 1 Nov 2026. Counts on this page were computed on 6 Nov 2026 from data/row-method.csv (397 rows).