Skip to content

How a verdict gets written

A verdict is drafted from a template, checked by a script and by a person, shown as proposed for at least 7 days, and only then made final.

The steps

1. A claim comes due. The deadline and the grace period have both passed, and the criteria have not been changed since the deadline. The tool lists every claim that is due, nearly due or overdue. (script)
2. Collect and fix the evidence. Only the sources frozen on the claim, plus any document a party sends, are read. Every source the verdict will cite is fetched and hashed on the day, before any drafting, and an independent archive copy is made. (desk)
3. Draft from the template. The drafter fills a fixed template: the four questions, the evidence grade, the reasoning, one sentence of summary. The drafter is named on the claim. A draft stays in the private desk repository and is not shown to anyone until it passes the next two steps. (AI, named)
4. Mechanical checklist. A script checks everything a script can: the dates, the four questions, the grade against its sources, the evidence bundle, the wording. A draft that fails any item cannot go further. (script)
5. Human checklist. A person reads each decisive source in full and answers the checklist below. The answers, who gave them and when are written onto the claim. A script cannot tell whether they were answered honestly. (the editor, and a named reviewer where one is on the registry)
6. Publish as proposed. The verdict goes on the claim page in full, marked 'not yet final', with the day the hold ends. Its hash is entered in the integrity log. The claimant is sent the link by the editor; nothing has been sent to any claimant so far. (the editor)
7. Hold. At least the hold period passes. A reply from the claimant, or a document from anyone, is read and answered in public. (calendar)
8. Finalise. After the hold the claim becomes resolved. If the verdict changed since it was proposed, the claim page says what changed and why, and the change is visible by comparing the two hashes in the integrity log. (the editor)
9. After publication. The right of reply stays open. A verdict shown to be wrong changes in public, through the corrections process. (anyone)

The four questions

Every verdict answers the same four questions, so that “not verified” can be told apart from “verified, but the wrong quantity”. They appear as a table on every verdict.

'Not yet assessed' is also a legal answer to any question, but it needs a sentence in the claim's axes note saying why.

Grading the evidence

The five tiers are defined on How we check. These are the rules for applying them:

• The grade is the best tier among the decisive sources, and a decisive source is one the verdict turns on. A source that is only background is not decisive and does not raise the grade.
• Every decisive source is also a cited source, has an archive copy, and has a fetch recorded in the evidence bundle.
• The tier is set by the kind of document, not by how convincing it is. The tiers themselves are defined on How we check.
• A verdict that something was not delivered rests on an absence. Its grade is the grade of the best document that reports the delivery schedule, so it is often low however sure the conclusion is, and the claim page says so.

The checklists

Mechanical: run by a script, enforced by the build

• M1. The deadline and grace period have both passed.
• M2. The criteria were not changed after the deadline.
• M3. All four questions are answered, with values from the allowed list, and a 'not yet assessed' has its reason.
• M4. 'Did it happen?' agrees with the verdict.
• M5. The grade is the best tier among the decisive sources, and every decisive source is a cited source.
• M6. Every cited source has an independent archive copy and a fetch in the evidence bundle taken on or before the day the verdict is proposed.
• M7. The reasoning and the one-sentence summary are present and the summary fits on one line.
• M8. The wording does not claim more than the data shows: no 'pre-registered', and no 'independently verified' unless the second question is yes.
• M9. The drafter and the reviewer, if any, are named, and a reviewer named on the claim is on the reviewer registry.

Human: answered by the editor and recorded on the claim

• H1. I read each decisive source in full, not a summary of it.
• H2. Every sentence of the reasoning can be traced to a cited source.
• H3. The verdict follows the criteria as they were frozen, not the criteria I would prefer.
• H4. I looked for evidence that would change the verdict and say in the reasoning what I looked for.
• H5. Any document the claimant or anyone else sent has been read and is cited or answered.
• H6. The wording says what the sources say and no more.
• H7. Any interest I have in this claimant or claim is declared on the claim page.

The mechanical items are enforced by the build. The human items are recorded, not enforced: the build can show that someone ticked them and when, not that they were true. That is the same limit the Ledger policy states for the editor's other judgements.

The hold

A proposed verdict is public from the day it is proposed. It cannot be made final for 7 days. A drafted verdict is published on its claim page, marked 'not yet final', and stays that way for at least this many days before it can become final. For the first two verdicts the site's commit history shows them committed as proposed on 15 September 2026 and as resolved on 23 September, about eight days apart, but no date was recorded on the claim and the history does not show when each commit went live, so no script could check that a hold took place. From the effective date a script can.

The day it was proposed is a field the editor sets, and the proposal’s hash is entered in the integrity log, whose head is anchored externally (see that page for the current state of the anchor). That makes it hard to change quietly, not impossible to set wrongly: the same limit as every other date on the Ledger.

Claims due by 31 December 2026

8 claims have a deadline on or before that date. A verdict can begin the day after the grace period ends, and after the 7-day hold it can be final.

A dry run on a published verdict

On 19 October 2026 the pipeline was run on the IonQ five-nines verdict (ionq-five-nines-2025), which was published on 22 September 2026 under the old, informal process. Two things were done, and neither publishes anything.

1. The published verdict, audited against today’s checklist

This asks what the checklist would have stopped. It does not re-judge the verdict.

4 of 11 items fail. The verdict was drafted before the checklist existed, so it has no recorded human review and no recorded proposal date, and its source snapshots were taken later, on 19 October 2026 (see the evidence section). The wording item is a real finding: the reasoning says “pre-registered”, which Correction 7 shows was wrong. The checklist would have refused that sentence.

2. The whole pipeline, rehearsed on a copy

The published verdict was used as the draft, so no new verdict was written. The sources were fetched again, for real. The clock was set to 1 February 2027 so that the deadline-and-grace rule and the hold could be exercised. The one change made to the copy was to remove the word “pre-registered” so the later steps could run. The “approval” entered was labelled REHEARSAL (not a review). Nothing was written to any claim or log.

What this shows is that the gates close when they should: a draft with a bad phrase cannot be proposed, a verdict cannot be made final early, and an edited verdict cannot be made final without saying what changed. It does not show that a human review catches errors.

What this does not do

• It does not make the verdict right. It makes the route to it visible and the skipped steps detectable.
• It does not add a second pair of eyes. The editor is still one person, and no outside reviewer has been appointed (see who adjudicates).
• No claimant has been sent a letter yet. The plan is to send the proposed verdict’s link on the day it is proposed; until a claimant is actually written to, the page for that claim says so.