# Ornith P3 — Reasoning Budget / Coding

Run: `/home/lab/bench/ornith/p3/p3-20260917T232306Z`

Each profile gets the same 3 deterministic coding-reasoning exams.
Exact PASS = 8/8 answers correct. Closure requires non-empty final content and finish_reason=stop.

| Profile | budget | n | exact pass | closure | score med | reasoning tok med | final tok med | TTFT med ms | total med s | gen t/s med |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| LOW | 256 | 3 | 3/3 (100.00%) | 3/3 (100.00%) | 8.0/8 | 255 | 1638 | 497.51 | 25.58 | 80.78 |
| MEDIUM | 1024 | 3 | 2/3 (66.67%) | 3/3 (100.00%) | 8.0/8 | 1023 | 578 | 474.14 | 19.98 | 81.86 |
| HIGH | 2048 | 3 | 1/3 (33.33%) | 2/3 (66.67%) | 7.0/8 | 2047 | 676 | 480.47 | 33.81 | 81.76 |

## Profiles

- LOW: `thinking_budget_tokens=256`
- MEDIUM: `thinking_budget_tokens=1024`
- HIGH: `thinking_budget_tokens=2048`
- All use `max_tokens=4096`, reasoning_format=auto, temperature=0.
- P3 server is 1 × 32K, Q4/Q4 KV, FlashAttention ON, DFlash OFF.

## What to choose

Use the smallest profile that preserves exact task accuracy and closure.
P3 intentionally does not auto-select a winner; inspect score, closure, latency, and token cost together.

## Baseline restored

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

