Local LLM Benchmarks R9700 episode CSV (Donato Capitella, r9700-episode.csv)

Per-depth episode export behind local-llm-benchmarks.dev/speed, distinct from data/results.json (already filed as local-llm-benchmarks-speed-results-json). Columns: run_id, engine, engine_version, runtime_version, measured_image, model, quant, mode, gpus, concurrency, target_depth, cached_depth_mean, followup_prompt_tokens, computed_prompt_mean, recomputed_history_mean, pp_tps, pp_stddev, tg_tps, tg_stddev. Protocol is the site deterministic-prime-extend-completions-v1: a ~2048-token follow-up at a primed context depth, 256 generated tokens, temperature 0, two repeats. pp_tps is effective follow-up prefill (uncached tokens over client-observed time to first output, so recomputed-history time is included). tg_tps is generation after the first output token, so TTFT is excluded. Every row in this file is gpus=2, i.e. two Radeon AI PRO R9700 cards. The newest run cited from the page writeup (run-4d514b0d7034191a, tcclaviger vLLM) was measured 7 October 2026. Companion sustained-concurrency CSV at https://local-llm-benchmarks.dev/data/r9700-tcclaviger-sustained.csv is the same publisher and measurement date: short 57-68 token prompts, 2048 output tokens, ignore_eos, one batch per row.