01 / xReview
Start with the decision the report exists to support.
A review starts with the decision a report is meant to support and the person who owns it — not with the dataset underneath it.
A runnable review prototype exists: eight screens, an annotated specimen, and a scoring engine whose integer output is superseded by the band contract. It has no backend, its data is mocked, and it has never been run against a real client report.
Status as at 2026-09-11 · Sources: existing-code inventory · prototype reuse matrix · xScore contract
The intake
What a review looks at.
Four questions, asked in this order, before anything is scored.
- 01
The decision, and who owns it
Which decision the report exists to support, and the person who is personally accountable for acting on it. A report nobody has decided anything from in recent memory is a finding about the report, not a reason to skip this question.
- 02
What each number is defined to mean
The same label can carry two different calculations in two different reports. A review writes the definition down before it argues about the visual.
- 03
The model and source underneath
Where the numbers actually come from — the relationships, the grain, the keys — checked, not assumed, because a clean visual can sit on top of an unreliable model.
- 04
Whether anyone has acted on it
A report with no decision behind it in recent memory is treated as evidence in its own right, not as an audience problem to route around.
The output
What comes back.
Six things, produced in this order — never a single number standing alone.
- 01
Capability manifest
What was actually examined and what was out of scope, published first, before any finding, so the boundaries of the review are on the record from the start.
- 02
Report context, sourced never inferred
The decision, the owner, and the audience — taken from the person who answered, not guessed from the report's own layout.
- 03
Executive verdict
A short, plain statement of where the report stands, written for the decision owner, not for the person who built the model.
- 04
Dimension profile
A read across each quality dimension a report is checked against — never averaged into one figure without saying which tier and how much evidence supports it.
- 05
Top findings, maximum five, each with its warrant
Every finding names the exact thing that is checkable behind it — a definition, a relationship, a key, an unanswered question — capped so the report is not buried under its own review.
- 06
Refusals, printed not buried
Anywhere the evidence does not support a conclusion, that is stated plainly, in the same document, not left as a silent gap.
The public form
How a result is spoken about.
The band, not a number
A band, with its evidence tier and its coverage.
When the evidence supports it, a result is described as a band, together with its evidence tier and how much of the report that tier actually covers — never as a single figure that invites false precision.
While a critical finding is open, no composite is issued at all. The band exists so a result can never claim more certainty than the evidence that produced it.
76–84
at Tier 2 (indicative), coverage 0.71
Never “83/100”. A band always travels with its tier and its coverage, and while a critical finding is open there is no band at all.
The limits
What a review will not conclude.
Some things a review can measure. Some it cannot. Saying which is which is what makes the rest worth trusting.
Refusal
A review does not conclude that a screenshot establishes correctness or accessibility, that an unflagged report is a good report, or that a review is reproducible without the decision contract behind it.
These limits are stated here so the rest of a review can be trusted without them being repeated everywhere else.
Start with the question
Request a review.
Name the report, and name the decision it is supposed to support. The rest follows from those two answers.