# The two-week test

A design you have not tested is a belief.

About an hour. It is the only part of this session that needs a desk and a quiet head, and the obvious
version cheats.

## The method

1. **Take two past weeks.** Your own, or the two in `05_sample_mailbox.md`.
2. **Mark what actually mattered** in each week. You know, because the weeks already happened. This is
   your ground truth.
3. **Tune against week one only.** Adjust your searches, your rules and your ten lines until the system
   catches what it should.
4. **Run it untouched against week two.** Change nothing first.
5. **Count both directions.** How much of what mattered did it surface, and how much did it surface
   that did not matter?

**If you tune and measure on the same week you will prove only that you can describe the past.** That
is the mistake almost everybody makes with this.

A triage that surfaces forty things to catch the two real ones has not helped you. Both counts matter,
and only one of them is flattering.

---

## Answer key for the sample mailbox

Do not read this until you have run yours.

### Week one

**Genuinely needs action from you — 5**

| # | Why |
|---|---|
| 1 | Priya asks for the revised sequence before Friday. Softened by 8, but still owed. |
| 12 | Sana asks a direct question and is blocked until you answer. |
| 13 | An invoice with payment terms. |
| 14 | Building management needs a named contact and a date range before they will issue the letter. |
| 10 | Marcus asks whether they need to provide anything. Low cost, genuine ask. |

**Contradicts your plan — 1, and it is the one that matters**

| # | Why |
|---|---|
| **3** | **"The cabinets will now arrive on the fifteenth rather than the eighth."** Nothing is asked of you. No rule catches it. Your install starts on the eighth and, per message 16, the crew is booked and non-refundable. This is the single most expensive item in the week and it is in paragraph three of a friendly update. |

**Nothing asked, no contradiction — 10**: 2, 4, 5, 6, 7, 8, 9, 11, 15, 16.

**The trap in week one is message 7.** "The quarterly review has been scheduled" is an announcement.
Nobody asked you to confirm anything. Most assistants return "confirm attendance at the quarterly
review" — and the quoted sentence underneath will not say that. Two seconds to catch, and only because
of the quote.

### Week two

**Genuinely needs action from you — 6**

| # | Why |
|---|---|
| 3 | Dan needs a delivery day confirmed. |
| 6 | Marcus wants to move the walkthrough to the eleventh. Needs a decision. |
| 8 | Priya still needs the sequence, now with a Monday deadline. |
| 10 | The one-pager is due on the fifth. |
| 12 | The invoice is past terms and needs chasing. |
| 2 | Sign-off received, so the renewal now moves to you. |

**Contradicts your plan — 3**

| # | Why |
|---|---|
| 6 | The walkthrough date in your block is the fourteenth. |
| 13 | **Sixty-day terms, not thirty.** You believed thirty. Nothing is asked of you. |
| 15 | **Priya is leaving the programme on 1 December.** Explicitly "no action for you" — and she is the person you owe the sequence to and who briefs the crew. |

**Nothing asked, no contradiction — 7**: 1, 4, 5, 7, 9, 11, 14, 16.

### How to read your score

- **Did you catch week two's items 13 and 15?** Both say nothing is being asked of you, and both change
  your plan. That is rule three working. If you missed them, your ten lines are too thin — they are the
  only reason a model can see the contradiction.
- **Did you get a false positive on week two's 4 or 7?** Both are announcements. Rule two exists for
  exactly these.
- **How many of the seven no-action messages came back as action items?** That is your noise rate, and
  it is the number that decides whether you keep using this.

A reasonable result on week two is five or six of the six asks, two or three of the three
contradictions, and no more than one or two false positives. If you are at forty items, the problem is
not the model — it is that rule two is not in your prompt, or your ten lines are missing.