practice · briefs & discovery

Writing predictions

The five-field discipline that makes a brief a brief. A claim about a measurable change, written before the cycle runs, with a date someone has committed to running the check.

TL;DR

Every prediction names five things, all together: the baseline (current measured state, with sample size and date), the target (what we expect after), the check date (when someone runs the measurement), the check method (the specific way), and the owner (the named person responsible). Less than five fields and the brief is decoration. The corpus rule: no baseline, no execution.

What it is

A prediction is a falsifiable claim about a measurable change, written before the cycle runs. It is the smallest unit of honest commitment in the chain. It lives inside a Feature Brief or Initiative Brief. It is what Volume V Part 2 — Signal and the Prediction checks at the end of the cycle.

Distinguish from

Feature Brief — the document that contains a prediction. KPI — a leading or lagging indicator at portfolio scale; not cycle-bound. Success metric — vague; the corpus does not use this term. See Confusable with at the foot.

Why it matters

The cycle is a closed loop. The prediction is the loop's anchor. Without a prediction:

• There is nothing for Volume V Part 2 to check. The cycle ran blind — the worst of the four outcomes.
• The model never updates. The next cycle inherits the same wrong model.
• The team's calibration over time cannot be measured. You cannot improve at predicting if you never recorded a prediction.

Without a prediction, the rest of the chain still functions — code ships, features go live — but the team is now building software whose effect on the world is unmeasured. That is the most expensive kind of working.

How to do it

Step 1 — Capture the baseline before the cycle starts

A baseline without a sample size and a date is a guess. Witness the current state directly.

Baseline: 47 minutes (mean), 38 minutes (median), n=12, captured
2026-04-22 by direct observation of three named graders.

If the metric the prediction will check has no instrumentation today, the cycle's first story is as PO, I want to know how long Gal spent grading, so that I can run the check on 2026-06-15. The instrumentation lands as part of the slice. See Volume IV · Observability.

Step 2 — Name the target with the form the change supports

Pick one of three forms:

Form	When to use
Specific number	The change has a single dominant effect
Range	The change has high variance across the population
Threshold	The change must clear a binary criterion

Don't say "we'll see what happens." That's not a prediction.

Step 3 — Set the check date with a calendar commitment

Check date: 2026-06-15.

The check date is a date the named owner has committed to. It is in their calendar. It is at most six weeks after the flag is enabled — long enough for first-contact noise to settle (per Volume V · First 48 Hours), short enough that the cycle is still fresh in everyone's head.

Step 4 — Specify the check method

The method is named in Discovery, not invented at check time. If the method requires instrumentation, the instrumentation is a story. If the method requires observation, the observation sessions are scheduled.

Check method: Three observation sessions across three named graders,
in the field, with a stopwatch and the time-on-task event log as
cross-check. Minimum 8 cycles total. Same observers as Discovery.

Step 5 — Name the owner

A role is not an owner. PO is not an owner. Alex (PO) is.

Owner: Alex (PO).

A complete prediction

Prediction:    Gal completes a grading cycle in under 15 minutes.

Baseline:      47 minutes (mean), 38 minutes (median), n=12,
               captured 2026-04-22 by direct observation.

Target:        <15 minutes (mean) across n>=8 observed cycles
               OR <12 minutes (median) across n>=8.

Check date:    2026-06-15.

Check method:  Three observation sessions across three named
               graders, in the field, with a stopwatch and the
               existing time-on-task instrumentation as cross-check.
               Minimum 8 cycles total. Same observers as Discovery.

Owner:         Alex (PO).

Copy the template →

Evidence

Across our cycles, the predictions that survived contact with reality shared three properties.

1. Baseline captured by observation, not by query. Cycles where the baseline was retrieved from a dashboard had a 2.3× higher too conservative / too optimistic rate than cycles where the baseline came from sitting next to the named person.
2. Check method named in Discovery, not at check time. Cycles where the check method was decided after the cycle ran produced not checked outcomes 4× more often.
3. Owner is a person, not a role. When the brief said "PO" instead of "Alex (PO)", the check happened on time 60% of the time. With a name, 95%.

The largest gap remains: in two cycles in five, the baseline was witnessed for the wrong moment. The grader's 47 minutes was wall-clock; the prediction implicitly meant focused minutes. That gap belongs to Discovery — see Clinic below.

Anti-patterns

These are the failure shapes worth seeing first.

Pattern	What it looks like	Where to fix
No baseline	"We expect users to be happier." No number.	The corpus rule: no baseline, no execution. Capture before the cycle starts.
Vanity baseline	The baseline is the metric you wanted to see, not the metric you measured.	Witness in person. See Volume II · Observation.
Unnamed owner	Brief says PO will check.	Replace with a named person. The calendar commitment is the discipline.
Survey-shaped check method	"We will ask graders if it feels faster."	Replace with observation. See Volume V · Signal and the Prediction.
Drifting check date	The date passes. The PO moves it without recording the move.	Mark as not_checked (the only worthless outcome). The chain treats moved dates as a chain-level signal at retro.

The clinic for the most common failure: A brief that didn't witness.

Confusable with

This	Not this	Difference
Prediction	KPI / OKR	Cycle-bound, falsifiable, owned. KPIs are portfolio-scale.
Prediction	Hypothesis	Hypothesis is exploratory; prediction is committed. Predictions are made before, checked after.
Prediction	Goal (Volume I)	Goal is 12-month; prediction is 4–6 weeks.
Baseline	Benchmark	Benchmark is industry-comparison; baseline is our person, our cycle, our number.

Further reading

• Canon — Volume II · Part 9 — Prediction Writing
• Canon — Volume V · Part 2 — Signal and the Prediction
• Reference — Master Area · Prediction Writing
• Template — Prediction skeleton
• Checklist — Five fields
• Clinic — A brief that didn't witness
• Skill path — PO foundations (steps 4–6)