part two · signal and the prediction

Signal and the Prediction

The check that was promised before the cycle ran. What the four possible outcomes mean — and why all of them are valuable except one.

Events in this phase

The check session happens on the date named in the brief — a scheduled commitment. The signal reading is written immediately after, short and factual. The retrospective uses both as input but happens later.

This is where meaning is tested against reality. The prediction was written before the cycle ran. The check date arrives. Someone runs the check — the specific measurement named in the brief, not a survey, not an impression.

Running the check

The check is an observation, not a report. If the prediction was "Gal will complete the grading cycle in under 15 minutes," the check is watching Gal grade a real exam — not asking her how long she thinks it takes. The check method mirrors the discovery method: witnessed, not described.

You bring the baseline numbers, the prediction target, and an open mind. You are not there to confirm success. You are there to correct the model. The purpose of the check is not validation — it is model correction. A check that confirms the prediction updates the model with confidence. A check that contradicts it updates the model with direction.

Only a check that never happened leaves the model exactly as wrong as it was. A prediction that was not checked is indistinguishable from a guess.

The four outcomes

Outcome	What it means	What you do next
Right	The prediction was accurate. The model was correct.	Document it. Name what specifically was right so it carries forward.
Too conservative	The change was bigger than predicted.	Understand why the model underestimated. Still a gap — in calibration.
Wrong	The prediction did not come true.	The most valuable outcome. Name the gap specifically. Feeds the next brief.
Not checked	Nobody ran the check. The date passed.	The only outcome with no value. The model cannot update. The cycle ran blind.

The check that confirms what the team hoped for is comfortable. The check that surfaces a gap is uncomfortable. Both are equally valuable. Only the check that didn't happen is worthless.

Writing the signal reading

After the check, a signal reading is written — a short document next to the Feature Brief in Confluence: what was predicted, what was measured, the gap, what the gap tells us. Not a retrospective — facts and first interpretation. The retrospective comes later.

A useful signal reading has five lines and no flourish. It is read by people who weren't in the check session. It is the input to the model update.

Prediction:   Gal completes the grading cycle in under 15 minutes.
Baseline:     47 minutes (mean, n=12, captured pre-flag).
Target:       <15 minutes.
Measured:     11 minutes 20 seconds (mean, n=8, captured weeks 1-3 post-flag).
Gap:          Better than predicted. Investigate why — likely the new
              keyboard shortcut absorbed more time than the deep-link
              navigation we built for.

Resolution gate — signal to bug analysis

Enough to know whether the model held.

The check ran on the named date. The signal reading is written and lives next to the brief. The team has an honest answer.

Part 3 — Bugs and Their Roots →