# Ornith P2 — Long-context Retrieval

Run: `/home/lab/bench/ornith/p2/p2-20260917T225043Z`

PASS requires all three exact needles in FINAL `content`, not merely in `reasoning_content`.

| Scenario | n | passes | pass rate | raw prompt med | server prompt med | TTFT med ms | total med s | prompt t/s med | gen t/s med |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| P2-NATIVE-262K | 2 | 2 | 100.00% | 244980 | 245166 | 159535.53 | 164.94 | 1542.74 | 32.80 |
| P2-YARN-512K | 2 | 2 | 100.00% | 499892 | 500156 | 477570.38 | 490.91 | 1047.69 | 19.22 |

## Profiles

- `P2-NATIVE-262K`: 262144 context, native model RoPE.
- `P2-YARN-512K`: 524288 context, YaRN, `--rope-scale 4`, `--yarn-orig-ctx 262144`.
- Q4/Q4 KV, FlashAttention ON, DFlash OFF, vision OFF.
- Output hard cap 4096.
- Every round starts from a fresh llama-server (cold cache).

## Baseline restored

- 3 × 256K
- KV Q4/Q4
- FlashAttention ON
- DFlash OFF
- cache-prompt ON
- hard output 4096

