Calibration
Calibrate Auwai against the model and endpoint you actually use. The repository does not publish a reference machine, a hosted endpoint, or a hardware-specific throughput claim. Those values change with the machine, the model build, the quantization, the context window, and the serving stack.
Turn formula
turns = ceil(total indexed chunk tokens / per-turn budget) + 1 reserve turn
The per-turn budget comes from the context window, the fixed prompt, the output
reserve, and the safety margin. Use auwai budget or the planner output to
read the resolved values.
At a large context window, the policy table is:
| Policy | Per-turn budget | Auwai turns |
|---|---|---|
| Default: reserve 20K, ratio 25% | 61,000 | 6 (5 + 1) |
| Tuned: reserve 14K, ratio 22% | 55,000 | 7 (6 + 1) |
The tuned policy leaves more context for reasoning and output and costs about 10% more turns.
Time per turn
Measure prefill and decode on your own endpoint with one fixed prompt. Then:
prefill seconds = (read tokens + fixed prompt) / measured prefill tok/s
output seconds = output tokens / measured decode tok/s
An Auwai turn is several model calls. Prefix caching changes the result substantially, so measure a full plan/read/record/report turn rather than a single completion. Record the endpoint shape, model id, context window, and hardware class in the run notes, not in this repository.
Corpora
Turn counts are exact from the Auwai index. Wall time depends on your measured throughput.
| Repository | Chunk tokens | Turns at 61K | Turns at 55K |
|---|---|---|---|
| Auwai | 282,315 | 6 | 7 |
| Linux kernel | 364,909,391 | 5,984 | 6,636 |
The kernel full-read estimate is a mathematical ceiling over all architectures, drivers, and documentation. It is not a practical target. Task-scoped plans are far smaller.
Recommendation
Use the reserve and ratio that leave enough room for the answer and reasoning on your endpoint. Use a larger per-turn read budget only when minimizing total turns matters more than headroom.
auwai workspace brief --workspace my-workspace --context-window 262144 \
--reserve 14000 --ratio 22%