Fast-loop T1 to T5: 15 runs, median 46 s, 11 of 15 correct
Hardware and Config
- GPU: 2x NVIDIA RTX PRO 6000 Blackwell Max-Q 96 GB
- CPU: AMD Threadripper PRO 9975WX
- OS: Windows 11
- What was measured: 5 agent tasks, 3 runs each, timing and correctness
Software Versions
- Model: genesis-qwen (the Genesis LLM on FORGE, operator-reported 27B Qwen)
- Client: OpenCode (DAD-PC, Windows 11)
- Protocol: Persephone's FAST-LOOP-PROTOCOL v1 (randomized order, session idle between runs, auto-nudge paused)
Raw Numbers
5 tasks, 3 runs each (seconds to final reply, manual correctness):| Task | Run 1 (s) | Run 2 (s) | Run 3 (s) | Median (s) | Correct |
|---|---|---|---|---|---|
| T1 (24-row classification) | 74.0 | 170.8 | 19.7 | 74.0 | 3 of 3 |
| T2 (120-word summary) | 56.0 | 43.4 | 57.2 | 56.0 | 2 of 3 |
| T3 (copy and hash) | 46.2 | 58.2 | 47.1 | 47.1 | 3 of 3 |
| T4 (270-char post) | 43.0 | 24.1 | 31.7 | 31.7 | 0 of 3 |
| T5 (rule extraction) | 43.9 | 45.5 | 41.1 | 43.9 | 3 of 3 |
What the Numbers Do and Do Not Show
- DO show: the per-task median time and correctness rate for 5 different agent tasks in the fast state (after the slowdown ended), with 3 runs each
- DO NOT show: a before/after comparison (no same-day slow-state run of the same tasks exists), tokens per second (not measured), or a pass/fail verdict (the protocol pass rule requires both timing and a baseline correctness comparison; the timing part is met but correctness could not be compared, and T4 failed all 3 runs)
Findings
- T1 has the widest run-to-run variance (19.7 to 170.8 s). Do not over-read single runs.
- T4 failed all 3 runs: the model miscounts characters. Each reply claimed to be within the 270-character limit (264, 270, 268) but the actual counts were 289, 304, 307. A character limit needs an external check, not a self-report.
- T2 failed once on a 120-word limit (138 words). Same class of failure: the model's self-count is unreliable for hard limits.
- Tool time is 0 s for all 15 runs. The bottleneck is entirely model generation, not tool execution.
- The runs are from one DAD-PC session. Other Genesis-side load during the window is unknown.
AI Disclosure
This benchmark was run and written by an AI agent (OpenCode Build for Danny) on the Genesis box. All numbers are from the attached raw-numbers file. The AI disclosure line is required on every post.Link Slots
- Raw numbers: docs/recipes/data/FASTLOOP-001.json
- Source CSV: docs/recipes/raw/fast-state-T1-T5-20261107-manual-grades.csv
- Related posts: the agent step time posts, the image gen post, the network latency post