The workings

Every grade on this site comes out of the same chain of steps. This page writes the chain out, one step at a time, so that anyone can follow a grade back to where it came from and check it.

Each step is marked as one of three kinds:

• Rule. Given the same inputs, anyone applying it gets the same answer. Most rules are applied by code, so they cannot drift.
• Judgement. A trained reader could reasonably disagree. Each one says how the disagreement is contained.
• Judgement, then rule. A judgement feeds a rule. The judgement is where to look if you disagree with the result.

Each step also names the question the publication’s reasoning framework puts to it. The six thinkers are not a separate stage. Their questions are asked at the point where the failure each one guards against would get in.

A few symbols keep the steps short. c is one claim. s is one source. P is a pattern, and D a domain.

1. What kind of claim is it?

A claim is typed by the question it answers, never by how strong its evidence looks. “Thirty per cent lose function in hospital” is natural history. “Lying in bed causes it” is harm from exposure. One sentence can make both, and then each is graded separately. The type decides which method grades the claim in step 4. This step stops a claim being graded by whichever method flatters it.

The question (Lennox): what is actually being claimed, and what is only being inferred from it?

2. Which sources count?

Each source is marked as testing the claim, bearing on it, or not testing it. That mark is a judgement, written down with one line saying why. Then two rules apply:

• A study counts once, however many times it is cited.
• Recognition never counts. Mandates, inquiries, guidance, board papers, endorsements and this publication’s own articles are all recognition. They are listed, so you can see what was set aside, and never counted as evidence that the claim is true.

The question (Weinstein): has this consensus been earned by evidence, or is it held in place by who said it?

3. Where does the grade start?

The effect and natural history rows follow GRADE (Guyatt et al., 2008; Iorio et al., 2015), and the experience row follows GRADE-CERQual (Lewin et al., 2018). Mechanism is judged by realist questions (Wong et al., 2013). The test for a state of the system is this publication’s own, and it says so: who collected the figure, how and when; whether the source is independent of the body whose interests it touches; and whether the figure can be traced to a primary dataset.

4. How sure are we?

The downgrades are the five GRADE domains: risk of bias, inconsistency, indirectness, imprecision and missing studies. For experience they are CERQual’s four questions instead: how well the studies were done, coherence, adequacy of data, and relevance. Upgrades are allowed only for observational evidence. They apply to a large effect, to a dose-response, or where confounding would have shrunk the effect seen. Each move is a judgement, and each is named. The arithmetic is a rule.

Indirectness carries the publication’s population lens. Evidence from another health system, or from people unlike those in areas of high deprivation or dispersed rural areas, is marked down for it where the difference matters.

Two caps are rules:

The questions: what is known, what is reasonable inference, and what is speculation presented as settled (Lennox)? Is this guidance being read against the people it lands on (Baxandall, the period eye)? Base rates before vivid cases (Data, anchored in Kahneman).

5. How far does the evidence reach towards the person?

This is the value-added time reading. It is made for every claim about the effect of care, and it is a separate question from step 4.

Whether an outcome “moved” is read from the source. That is the judgement. Taking the highest rung reached is the rule. The rungs, nearest the person first:

• person: what the person would judge. That means goals, quality of life, capability or self-rated health. Satisfaction alone does not count, and nor does a difference in quality-adjusted life years alone.
• harm: harms the person experiences, such as delirium, falls, drug reactions or pain.
• place: being alive and where the person chooses. Place of death counts only where the person’s preference was recorded.
• function: function in daily life.
• process: measures that serve the system, such as length of stay or readmission.

Five more fields sit beside it:

• person: shown, not shown, or never measured. Where it was measured, its own confidence is graded as in step 4.
• cost: the time the care takes from the person. It is reduced, unchanged, increased or never measured, and savings to the service do not count.
• accompanied and carers: each either measured or not.
• equity: results reported separately for frail or deprived people, or not.

Each reading names the result that would show it wrong. Hard cases follow written rules. Symptom scales sit on harm. Physiological markers sit on process. An adverse effect is a harm, not a cost. A carer’s gain is recorded for the carer, not the person. Mixed evidence is read at its weakest consistent point.

The question (McGilchrist): does this see the whole person, or only the part that was measured?

6. A pattern starts from what we already said

The domain assessment is the starting point. In the terms of probability it is the prior, and the pattern’s evidence is what moves it. It is never evidence for the pattern, because a record that counts its own earlier view as support is only agreeing with itself.

The question (Schmachtenberger): what is structurally producing what we see, and does it show up where the domains meet, not only inside one?

7. The checks every pattern faces

These are the seven checks on how we grade. Steps 2, 4 and 5 above carry the first four. The other three:

The question (Holmes): working from the evidence alone, what else would produce this, what has been ruled out, and what would change the conclusion?

8. The pattern’s standing and status

A pattern is never stronger than its weakest claim. Its page shows every claim with its own grade, so a strong claim is not hidden behind a weak one.

A pattern moves one status at a time, and every downgrade is published with its reason.

The question (Data): is the result honest about what is not known, whether or not that is welcome?

9. Adding it up: the profile

The profile is these counts, for each domain and for the register, alongside how far each claim’s evidence reached. It is computed by code from the readings stored with each claim. Nobody types the figures, and the site’s checks fail if they drift. It is never combined into a score, because a score invites being read as a verdict on a service.

10. Over time

A correction is made at source and appears in the next month, with a note of what changed. Nothing is quietly rewritten.

Who checks the judgements

Every step marked as a judgement can be wrong. Three things contain that.

• Two independent passes. The October 2026 re-grade of the register was done twice, by passes that could not see each other, and they agreed on 15 of the 18 pattern statuses. In a domain’s evidence, anything that moves a grade is checked by a second pass that cannot see the first. Where they disagreed, the disagreement is shown on the pattern’s page. For a value-added time reading, the more cautious reading stands.
• A second reading each month. The monthly review grades in one pass. Its rules require it, before proposing a change of status, to argue the opposite case from the same evidence and keep the change only if it survives. That reading is not blind, and it is not described as one.
• A read after publication. Each published cycle is read again against its sources, and anything wrong is corrected at source.

What is running, counted from the site

• 91 of 91 claims about the effect of care in the five domains carry a value-added time reading.
• 7 of 18 patterns carry at least one value-added time reading from the October 2026 re-grade, and 11 more make only claims about the state of a system, which take no reading. 0 carry a reading written by the monthly review itself; none has yet, so until one does, the register's readings are the re-grade's.
• 0 of 7 months of published Signals carry readings. That series is not yet read, which is not the same as zero.
• Signals carry the type of their evidence, such as a meta-analysis or a primary study, and the publication's wording of its strength. They do not carry a confidence grade from step 4, which is made for the claims in each domain's evidence and for each pattern.

What no formula can do

The chain shows where every grade comes from. It does not make the judgements right. Whether a source really tests a claim, whether a rival explanation has really been ruled out, and whether an outcome really moved are all judgements. They are written down so they can be argued with. The six thinkers are there to make those judgements harder to get wrong, not to replace them.