# Sample datasets: the "Avery" persona

These three JSONL files describe four weeks (2024-09-01 through
2024-09-28) in the life of a fictional person, **Avery**, as seen
through three evidence sources:

| File | Kind | Records | Source label |
|---|---|---|---|
| `calendar.sample.jsonl` | `calendar.event` | 30 | `ical:personal` |
| `notes.sample.jsonl` | `note` | 8 | `notes:app` |
| `photos.sample.jsonl` | `photo.meta` | 12 | `photos:library` |

Each line is one evidence record:

```json
{"evidence_id": "...", "kind": "...", "source": "...",
 "observed_at": "ISO-8601 UTC", "attributes": { ... }}
```

## What the derivers should find

The data is constructed so that the built-in derivers produce a small,
inspectable set of claims:

**Routines** (from recurring calendar slots)
- Team standup, Mon/Wed/Fri 09:30 — 12 occurrences, high confidence
- Gym, Tue/Thu 07:00 — 8 occurrences (corroborated by 3 gym photos)
- Spanish class, Thu 18:30 — 4 occurrences
- Dinner with Sam, Fri 19:00 — 4 occurrences

**Places** (from photo geo-clusters and repeated calendar locations)
- A frequent place near (52.5208, 13.4095) labeled "Home - Friedrichshain"
- "Ironworks Gym" as a recurring place
- "Tempelhofer Feld" as a recurring weekend place
- Lisbon appears only twice and should stay low-confidence or absent

**Relationships** (from co-occurring attendees / tagged people / mentions)
- `sam.rivera@example.com` / "Sam Rivera": 4 dinners + 5 tagged photos
  + 2 note mentions → strong social tie
- `priya.nair@example.com` and `marcus.webb@example.com`: 12 shared
  standups + 1 note mention → work ties

**Preferences** (from notes)
- Coffee: oat-milk flat white (ev-note-001)
- Diet: vegetarian (ev-note-002, corroborated by ev-note-007)
- Travel: window seat (ev-note-008)

## Built-in correction story (used by the end-to-end tests)

1. Ingest all three files and run derivation.
2. Refute the "Dinner with Sam, Friday" routine claim — any claim
   derived downstream of it (e.g. a relationship claim citing those
   dinners) is mechanically invalidated, then re-derived from the
   remaining evidence (photos, notes) at lower confidence.
3. Retract `ev-note-002` (the vegetarian note) — the diet preference is
   invalidated and re-derives, if at all, only from ev-note-007 with
   reduced confidence.

## Try it

```sh
python -m mnema.derive.cli init   --node ./avery
python -m mnema.derive.cli ingest --node ./avery data/samples/*.jsonl
python -m mnema.derive.cli run    --node ./avery
python -m mnema.derive.cli claims --node ./avery
python -m mnema.derive.cli explain --node ./avery <claim-id>
python -m mnema.derive.cli refute  --node ./avery <claim-id> --reason "not true"
```

All names, addresses, coordinates, and emails are fictional.