# Ornith P1 — Cache / Prefix Reuse

Run: `/home/lab/bench/ornith/p1/p1-cache-20260917T221619Z`

Steady-state statistics use turns 2–5. Turn 1 is the cold priming turn.

| Scenario | n | TTFT med ms | p10 | p90 | reuse med % | processed med | cached med | prompt t/s | gen t/s | LCP | f_keep |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| P1-A-cache-on-same | 12 | 161.93 | 158.71 | 168.26 | 99.48 | 31 | 17902 | 231.18 | 75.16 | 0.998 | 1.000 |
| P1-B-cache-on-rotate | 12 | 3421.65 | 488.29 | 6595.30 | 49.22 | 9006 | 8880 | 2053.53 | 76.78 | - | - |
| P1-C-cache-off-same | 12 | 6114.08 | 6028.08 | 6156.58 | 0.00 | 17933 | 0 | 2960.65 | 75.68 | - | - |
| P1-D-cache-off-rotate | 12 | 6436.19 | 6252.08 | 6525.65 | 0.00 | 17933 | 0 | 2937.71 | 76.07 | - | - |
| P1-E-cold | 12 | 5871.94 | 5599.84 | 6026.31 | 0.00 | 17933 | 0 | 3076.68 | 75.38 | - | - |

## Derived comparisons

- Cold/hot median TTFT ratio: **36.26×**.
- Median reuse same-slot: **99.48%**; rotated: **49.22%**.
- Cache-OFF reuse fell to **0.00%** median.

## Baseline restored

- 3 × 256K
- KV Q4/Q4
- FlashAttention ON (all-quants build)
- DFlash OFF
- vision OFF
- continuous batching ON
- cache-prompt ON
- hard output 4096

