#2617 bench/results: RTX 5070 Ti 16 GB on Linux, A/B of five opt-in settings (Swift 1.5 IQ3_S, 0.1.44.1)

Description

Results only: one new folder, `bench/results/2026-10-15-community-rtx-5070-ti-opt-in-ab/`, its line in `docs/COMMUNITY_BENCHMARKS.md` and its row in `bench/results/COMMUNITY.md`. No engine, tool or setup change.

**Machine:** RTX 5070 Ti 16 GB (gen 4 x16), Ryzen 7 5800X3D (AVX2, no AVX-512), 80 GiB DDR4-3200, Ubuntu 26.04, driver 595.99.02, CUDA 13.1; Strata 0.1.44.1 source build (`296b59d2`, sm_120), Swift 1.5 IQ3_S with setup's settings.

**What was measured:** five A/Bs of 4 alternating rounds each, a fresh server per arm and round, eight short prompts and one long prompt per round, 256 tokens, greedy; engine timings, every request and engine log attached.

| Change (against the default) | Pooled short decode, B vs A per round | Medians (tok/s) |
| --- | --- | --- |
| `STRATA_ROUTE_TAIL_SKIP=0` | -23.7, -24.7, -21.7, -23.4% | 60.11 vs 78.73 |
| `--spec 5` (vs 4) | -4.7, -5.2, -2.8, -3.3% | 76.27 vs 79.79 |
| `STRATA_MTP_ADAPTIVE=1` (PR #2529's build) | -3.5, -1.0, -1.2, -2.5% | 78.22 vs 80.00 |
| `--adapt-async 1` (at 1M context) | +1.7, -0.6, -1.0, +2.3% | 73.26 vs 73.52 |
| `--mtp-window 8192` (63K-token prompt) | +0.1, -1.2, -0.4, +0.4% | 79.81 vs 79.94 |

The `--adapt-async`, tail-skip and PR #2529 results were also posted on #1903, #2482 and #2529.

**What it leaves out:** one machine and one model size, short generations only, answers differ between arms (speeds are paired by round, not by text), no quality check; the first A/B ran at 1,048,576 tokens and the other four at 262,144. Listed in the README.

**How it was checked:** a second reader recomputed every number from the attached `rows.jsonl`, engine logs, configs and telemetry summaries, and searched all files for user and host names, paths, addresses and keys. The same machine's speed report is PR #2580; both PRs add a line at the end of the same list in `docs/COMMUNITY_BENCHMARKS.md` and a row at the end of `bench/results/COMMUNITY.md`, so the second to merge has a one-line conflict there.

🤖 Generated with [Claude Code](https://claude.com/claude-code)