# Results: how often a method finds a pattern that is not there

**Run 2026-09-18.** Method committed before the run: `METHOD.md`, in this
folder, in the git history. Generator and runner:
`scripts/study/false-alarms.cjs`. Raw output: `results.json`.

## The headline

100 synthetic diaries per row, built with **no real pattern in them at all**.
Every pattern reported is therefore a false alarm.

| Diary length | A plain search for correlations | The same, corrected | Arkhelion |
|---|---|---|---|
| 60 days | 90% | 6% | 3% |
| 90 days | 88% | 11% | 4% |
| 180 days | 95% | 12% | 2% |
| 292 days | 97% | 10% | 0% |

Read the top right corner first: on nine months of a diary with nothing in it,
**a plain hunt for correlations reported a pattern 97 times out of 100.**

Chart: `docs/distribution/assets/study-false-alarms.png`.

## The half that keeps this honest

A method that never reports anything would score perfectly above and be
worthless. So the same three ran on diaries with **one real pattern planted**
(a measure raising the outcome two days later):

| Diary length | Plain search | Corrected | Arkhelion |
|---|---|---|---|
| 60 days | 72% | 29% | 16% |
| 90 days | 82% | 38% | 29% |
| 180 days | 96% | 80% | 71% |
| 292 days | 100% | 97% | 96% |

The plain search "wins" this table only because it says yes to almost
everything, which is what the first table measures.

**Our own limitation, stated plainly: at 60 days we miss a real pattern five
times out of six.** Short records are honestly weak. It takes about six months
before this finds a real pattern most of the time. That belongs in the product,
not just in a paper.

## What the study does not show

- These are synthetic diaries, not people. They carry a weekly rhythm,
  day-to-day drift and 15% missing entries, which is closer to real logs than
  clean random numbers, but it is not the same thing.
- Arm A is a plain correlation search of the kind a spreadsheet performs. **No
  AI assistant was tested.** Saying otherwise would be exactly the sort of
  claim this study exists to check.
- One effect size was planted. A weaker one would be missed more often by
  everybody.
- "Pattern" never means cause here, in any arm.

## Reproduce it

    node scripts/study/false-alarms.cjs --diaries 100

Seeds are fixed (`20260918 + index + days*1000` for the noise diaries), so the
same diaries come back every time.