practice · checks & signal

Model update

The corpus's amendment. What the cycle learned that the next cycle should inherit. Close at least one assumption. Add at least one. Sharpen at least one template, checklist, or glossary entry. Without it, the next cycle inherits the same wrong model.

TL;DR

The model update is one hour, written by the PO, signed by the trio, run the same week as the retro. It closes at least one assumption from the Initiative Brief, adds at least one new assumption, and sharpens at least one template / checklist / glossary entry the next cycle will read. The model update is a PR to the corpus, not a Slack post.

What it is

The model update is named in After We Build · The Model Update. It is the artefact that makes the cycle compound. Without it, every cycle starts from the same assumptions; with it, the corpus sharpens.

Distinguish from

Retrospective — the team's change. Model update — the corpus's amendment. Postmortem — written after an incident; structurally similar but bound to a specific failure. Documentation edit — a non-cycle change to docs. See Confusable with at the foot.

Why it matters

Without the model update:

• Assumptions never close. The Initiative Brief lists biggest unvalidated assumption — and nine cycles later it is still listed as unvalidated even though three cycles tested it.
• Templates never sharpen. The cycle learned that the Feature Brief needs an observation scheduling slot. Nobody amended the template. Next cycle's brief has the same gap.
• Knowledge is in heads, not in artefacts. The team's senior people remember; new hires don't. Tribal-knowledge debt accumulates.
• The next cycle inherits the same wrong model.

The corpus's most distinctive discipline is that we update the model after every cycle. The model update is how.

How to do it

Step 1 — Open with the retro's one change and the signal reading

The model update reads what produced it. The retro's named change is on the page. The signal reading is on the page. Together they tell the model update what the cycle taught.

Step 2 — Close at least one assumption

Open the Initiative Brief's assumption section. For each unvalidated assumption, ask: did this cycle test it? If yes, close it.

Closed assumptions (from 2026-Q2 Initiative Brief):

  ASSUMPTION (now closed):
    "Graders skip afternoon batches because of fatigue from
    morning friction, not because afternoon is when they teach."
  STATUS: VALIDATED.
  EVIDENCE: Signal reading 2026-06-15 measured 13.2 min mean
  per cycle (down from 47); voice-of-customer named three
  graders requesting afternoon batches enabled by default.
  The mechanism the bet relied on (focused-time recovery)
  held.

If the cycle did not test an assumption that was listed, that is also a model update — name the assumption as still-unvalidated, and the next cycle is now scheduled to test it.

Step 3 — Add at least one new assumption

Every cycle that produces a real signal reading also surfaces a new question. Name it.

New unvalidated assumption (added 2026-Q2):

  "Graders will adopt the new keyboard shortcut without
  prompting. (2 of 5 first-48h tickets were about shortcut
  discoverability.)"
  Owner: Alex (PO).
  Next cycle: in-app coachmark for the shortcut; measure
  adoption at H+48 of next cycle's release.

Step 4 — Sharpen one corpus artefact

The cycle learned that the Feature Brief template should have an observation scheduling slot. The model update is the act of changing the template.

# templates/feature-brief.md

  ## Linked artefacts
  - Initiative Brief: [/briefs/initiative/...]
  - Technical Design Brief: [/briefs/tdb/...]
  - Journey map: [link]
+ - Observation sessions scheduled: [dates · named graders]
  - Observation notes: [link]

This is the change that next cycle's brief will use. The discipline is the PR to the corpus — not a Slack note, not a comment in the retro doc. A real edit to a real template.

Step 5 — Update the glossary if needed

If the cycle taught a new term, or a new precision for an existing term, edit the glossary.

Step 6 — Sign and publish

The trio signs the model update. It is published the same week as the retro. The PR merges. The next cycle's brief inherits the sharpened template.

A complete model update

# Model update — 2026-Q2 cycle · Hebrew-name grading flow

## Inputs
- Signal reading 2026-06-15 (linked)
- Retro 2026-06-17 (linked)

## Assumptions closed
- "Graders skip afternoon batches because of morning fatigue,
  not because of teaching time."
  STATUS: VALIDATED.
  EVIDENCE: signal reading; three graders requested afternoon
  default.

## Assumptions added
- "Graders adopt the new keyboard shortcut without prompting."
  STATUS: UNVALIDATED.
  TO TEST: H+48 of 2026-Q3 cycle's release.

## Corpus changes
- templates/feature-brief.md: added "Observation sessions
  scheduled" line under Linked artefacts. PR #341.
- glossary.md: clarified "focused-grading time" definition
  to exclude alt-tab time. PR #342.
- /clinics/a-brief-that-didnt-witness.md: added the
  alt-tab/wall-clock distinction to point 2. PR #343.

## Sign-off
PO:        Alex (PO) · 2026-06-19
Designer:  Maya (Designer) · 2026-06-19
Tech Lead: [TL] · 2026-06-19

Evidence

Across our cycles, model updates that produced compounding learning shared three properties.

1. At least one template was changed. Model updates that produced only assumption-closures (no template change) had a 2× higher rate of the same problem next cycle. Templates are where the learning sticks.
2. The model update merged within one week of the retro. Updates delayed beyond one week were merged in 40% of cases; the rest decayed.
3. The PR was to the corpus, not to a private doc. Model updates that lived in Confluence were read by the team that wrote them; ones that landed in the corpus were read by the next team.

Anti-patterns

Pattern	What it looks like	Where to fix
Retro happens; model update doesn't	The team agreed something; nothing landed	Schedule the model update the same week as the retro
Assumptions never close	Initiative Brief lists the same 4 unvalidated assumptions every cycle	Close one per cycle — or admit the cycle did not test any
No template changed	The learning is in heads	Edit a template; the PR is the discipline
Update is a Slack post	"Just FYI we learned…"	The model update is a corpus PR
Updates accumulate	Three retros without a model update	Stop. The next cycle starts blind. Catch up before opening the next cycle.

Confusable with

This	Not this	Difference
Model update	Retro	Retro = team's change. Model update = corpus's amendment.
Model update	Postmortem	Postmortem = bound to a specific incident; model update = bound to a cycle.
Model update	Documentation edit	The model update is driven by what reality answered; doc edits can come from anywhere.
Closed assumption	Validated theory	An assumption is closed when the cycle tested it; it is validated if the test confirmed it. (Closed-and-refuted is also closed.)

Further reading

• Canon — After We Build · The Model Update · The Retrospective
• Practice — Signal reading · Retrospective
• Skill path — PO foundations · Step 9
• Principle — Compounding small changes · Predictions over plans
• Reference — Area · Model Update