Decision log entries with a written prediction improve later estimates

Hypothesis

When a team records, alongside each non-trivial decision, a concrete prediction of its outcome (what will be observed, by when, with a number or threshold) and a review date, and later grades the prediction against what happened, its estimates for similar decisions become better calibrated over months than those of a team that records decisions without predictions. Architecture decision records in Nygard's form capture context, decision, status and consequences; the hypothesis concerns the additional effect of an explicit, gradeable prediction and a scheduled review, for decisions of any kind.

Prediction

Comparing two periods, the second with prediction-and-review entries: the gap between predicted and observed effort, timelines or metric changes narrows; decisions reversed within the review window become fewer, or reversals happen earlier; contributors' stated confidence and their hit rate move closer together. Teams that write predictions but never hold the reviews show no change, which would indicate that the grading step, not the writing, carries the effect.

Proposed test

1. Define a decision log entry: decision, alternatives considered, predicted observable outcome with a number or threshold, review date, owner.
2. Have several teams keep such logs for at least two quarters; keep a comparison group that logs decisions with rationale only.
3. At each review date, grade the prediction (met, missed, unclear) and record the actual value.
4. Compare calibration (stated confidence against hit rate) and prediction error between periods and between groups, controlling for decision type and team size.

Status

No result is claimed. Possible confounds: teams willing to write predictions may already be more disciplined; small numbers of decisions per team make calibration estimates noisy; the effect may vanish when reviews are not scheduled or when predictions are written vaguely enough to be unfalsifiable.