Evidence against the judgement is a required part, not an optional caveat at the end.
Why it matters when the plan changes
A report whose shape changes with its author cannot be compared with the one before it. Fixing the parts makes the output boring in the useful sense: a reader knows where the counter-evidence sits, where the owner is named, and what will test the call. Nothing can be quietly omitted because its absence is visible. This is also what lets an Execution Forecast and its later Reforecast be read side by side, since both carry the same eight fields.
The tension is that two of the eight are the ones people would drop. Counter-evidence weakens the story and the human outcome slows the decision. They are required precisely because they are the first to go, and a judgement that arrives without them has been edited rather than made. The Good Judgment Project's multi-year study of real-world probability judgements found that recording forecasts in a fixed form and scoring them against outcomes is what allows accuracy to improve over time.
In practice
A judgement about a delayed integration arrives with the friction named, the evidence on both sides, one pillar marked blind, a confidence, a named owner with a date, the outcome that must not worsen, and the test that will settle it. The leadership team spends its time on the decision rather than on reconstructing the reasoning.
Evidence
Judgements recorded in a fixed form and scored against outcomes improve over time, while unrecorded ones do not.
Philip E. Tetlock and the Good Judgment Project (2026)
What it cannot tell you
The eight parts fix what must be present in a judgement, not whether the judgement itself is correct. A forecast can carry all eight fields and still be wrong: the form guarantees traceability and comparability across updates, not accuracy. It says nothing about the quality of the evidence gathered, only that whatever evidence exists must be shown for and against.
How Atlas reads it
The eight parts are the unit of delivery for a forecast and for every later update, and they replaced an earlier card format in August 2026. Alongside them each judgement carries the rating, the recommended move from the ladder, and the audit trail: which versions of which evidence were used, who reviewed it and what happened afterwards.
Questions
Strategic shift, meaning what changed. Friction, where and why. Evidence for and against. Blind spots, marked rather than filled. Confidence, meaning how much reliance the judgement carries. Owner, who changes what by when. Human outcome, what must not worsen. Eval, what will test the call.
Because an intervention that fixes a delivery problem can create a worse one for the people inside it. Naming what must not worsen makes that consequence a stated condition of the recommendation rather than something discovered afterwards by the people it happened to.
It names in advance what would show the call was right or wrong, and when that will be visible. Philip E. Tetlock's Good Judgment Project, a multi-year study published in 2026, found that judgements recorded and scored against outcomes improve over time, while unscored ones do not. The eval is what makes scoring possible.
The part stays and is answered, including with an explicit none or unknown. Removing it would hide the fact that it was considered, and the value of a fixed shape comes entirely from a reader being able to trust that nothing was quietly dropped.
They make the earlier judgement comparable with the new one. Because both carry the same eight fields, an update can show exactly what changed: the evidence, the confidence, the owner or the friction itself. This mirrors the Good Judgment Project's finding that scored, fixed-form judgements improve over time.