# Assumption map

Every plan rests on beliefs. This artifact separates the beliefs that would be expensive to
get wrong from the ones that would not, and buys evidence only where the answer would change
a decision.

Source: lesson 0040 · Facilitate Discovery and Brainstorming.

---

## Blank

Place each belief on two axes: **impact if false** and **evidence strength**. Test the
high-impact, weak-evidence corner first. Everything else waits.

| Assumption | Type | Impact if false | Evidence today | Test | Owner | By | Result |
| --- | --- | --- | --- | --- | --- | --- | --- |
| `[stated as a falsifiable belief, not a question]` | value / usability / feasibility / viability | high / medium / low, and *what breaks* | none / anecdote / sample / measurement | the cheapest thing that could disprove it | `[name]` | `[date]` | `[fill in after]` |

One line of guidance per column:

- **Assumption** — write it so it can be false. "Customers want self-service" is a slogan;
  "at least half of eligible customers complete a return without contacting support" is a belief.
- **Type** — *value* (will they choose it?), *usability* (can they complete it?),
  *feasibility* (can we build and run it?), *viability* (can the business support the policy,
  cost, and operation?).
- **Impact if false** — name the decision or artifact that would have to change, not a severity word alone.
- **Evidence today** — be honest about vendor claims and internal opinion; both are "none".
- **Test** — the smallest thing that could change your mind. A sample, five interviews,
  a four-hour spike. Not "research".
- **Owner / By** — one person, one date. An unowned test is a wish.
- **Result** — the number or finding, and which artifact it changed.

**Anti-pattern.** A wall of sticky notes with no owners. The board is not the output; the
tests and their dates are.

---

## Completed — returns resolution pilot

Eight beliefs. Three got a budget, one was converted into a rollout guardrail, and two were
deliberately left untested — which is a recorded decision, not an omission. A map where every
row has a test is usually a map nobody costed.

| Id | Assumption | Type | Impact if false | Evidence before | Test chosen (cost) | Owner | Result |
| --- | --- | --- | --- | --- | --- | --- | --- |
| A1 | At least 90% of eligible requests match a codifiable policy rule | feasibility | High — the 50/day manual-queue guardrail breaks and the automation premise weakens | Weak: one estimate from the policy owner, no sample | Classify 200 closed agent decisions from June–July as codifiable / needs one new rule / judgment call. 2 analyst-days | Priya | **Falsified.** 63% codifiable today, 22% need the opened-item photo rule, 15% are judgment calls. Changed the outcome statement (frame v0.2) and created the referral state |
| A2 | Customers trust an automatic decision enough not to contact support about it | value | High — the contact-rate half of the outcome fails even if the latency half succeeds | Weak: inferred from the 41% contact rate, never tested directly | Five-participant moderated comprehension test on a clickable prototype of accept, refuse and refer. 3 designer-days | Marco | **Inconclusive.** 2 of 5 read "your label is being prepared" as a refusal. n = 5 cannot separate a wording problem from two participants. The wording changed anyway — cheap and reversible — and a retest at n = 12 was booked for 15 September |
| A3 | The carrier creates a label fast enough to sit inside the request that accepts the return | feasibility | High — it decides whether label creation is synchronous, which decides the failure model of the whole flow | Weak: a latency figure in the vendor's own documentation, never measured by us | Four-hour timeboxed spike against the carrier sandbox: 200 sequential creates measuring p50/p95/p99, then 20 induced timeouts retried, ten with a client reference and ten without. Throw the harness away | Dana | **Falsified.** p50 640 ms, p95 1.9 s, p99 11.4 s. With a client reference, 0 duplicates in 10; without, 2 in 10. Superseded ADR 0002 with ADR 0003 |
| A4 | Warehouse receipt events are reliable enough to trigger a refund automatically | feasibility | Medium — a missed event strands money; a duplicate double-refunds | Moderate: six months of history and an existing consumer | Replay 30 days of receipt events against order ids; count missing, duplicate and late. 1 engineer-day | Lea | **Confirmed with a caveat.** No missing or duplicate events; 0.4% arrive more than 24 hours after the physical scan. The refund SLA therefore starts at the event, not the scan |
| A5 | Support agents will accept the system's refusal reasons rather than override them | viability | Medium — silent overriding means the policy model is wrong and nobody finds out | Weak: agent opinion in the discovery workshop | Shadow ten refusals with two agents; record every override and its stated reason. 1 day | Nils | **Partial.** 3 of 10 overridden, all three for gift orders, which no rule covers. Produced one new rule, deferred to slice 3 rather than smuggled into the pilot |
| A6 | Fraud loss does not rise when the decision becomes automatic | viability | High — a guardrail breach stops the initiative regardless of the outcome measures | Weak, and untestable before exposure: fraud responds to the presence of the new path, not to a prototype | **No pre-test bought.** Converted into a rollout guardrail: exposed cohort versus matched control, with a stop threshold and a named authority | Risk | Deferred to rollout by design. Held at every stage — and the closeout names what that sample size could and could not detect |
| A7 | Policy rules change often enough to justify a separate, independently releasable policy module | feasibility | Medium — if false, the modifiability driver is imaginary and the module seam is unearned cost | Moderate: an existing policy change log nobody had read | Read the change log. 2 hours | Dana | **Confirmed.** Nine policy changes in twelve months, four of them mid-quarter. Justifies driver D4 and the eligibility seam |
| A8 | Customers would rather self-serve than call, for this class of request | value | High | **Strong:** 71% of baseline contacts are status chasing on an already-submitted request, not requests for help deciding | None bought, deliberately. Buying evidence for a well-evidenced belief is a way of looking busy | — | Held as a stated premise and re-checked at closeout |

**Two habits worth stealing from this table.** First, the test is sized to the *decision*,
not to the assumption: A3 got a four-hour spike because a number would change a structural
choice, while A7 got two hours of reading because the log already existed. Second, "no test"
is a legitimate recorded choice — A6 could not be answered before exposure, so it became a
rollout guardrail with a stop threshold, and A8 was already evidenced.

**Traces to:** A1 and A2 carry the outcome; A5 and A6 carry guardrails; A3 is the force
behind driver D1 and ADR 0003; A7 is the evidence for driver D4 and the eligibility module
seam in `c4-review-checklist.md`; A6 becomes a stop threshold in `rollout-contract.md`.