2026-08-05

• 08:57 fetched 20260804-135339-punctuation_payload_gpt_oss_20b_capacity-0cdf618
• 08:57 diagnosed 20260804-135339-punctuation_payload_gpt_oss_20b_capacity-0cdf618 as failed before metrics during MXFP4 model weight conversion/loading with CUDA device busy/unavailable, so the latest valid payload-capacity evidence remains the 20260804-100944 smoke
• 09:43 added punctuation_payload_gpt_oss_20b_long to test supplied punctuation payloads at 16, 32, 64, and 128 bits after the mock capacity ladder confirmed long-payload sampling and metrics generation through 128 bits
• 10:04 submitted 20260805-100312-punctuation_payload_gpt_oss_20b_long-84c77bf (job 597074 on jean-zay, commit 84c77bf)
• 10:05 submitted punctuation_payload_gpt_oss_20b_long as 20260805-100312-punctuation_payload_gpt_oss_20b_long-84c77bf job 597074 on Jean Zay after channel confirmation, and scheduled follow-up checks at roughly 3, 5, 7, and 10 hours after submission
• 13:05 fetched 20260805-100312-punctuation_payload_gpt_oss_20b_long-84c77bf
• 13:05 reported 20260805-100312-punctuation_payload_gpt_oss_20b_long-84c77bf in notes/experiments/punctuation_payload_gpt_oss_20b_long.md
• 15:05 checked long punctuation payload run 20260805-100312; confirmed fetched metrics and existing report block, with long-payload truncation warnings limiting interpretation
• 15:08 interpreted 20260805-100312-punctuation_payload_gpt_oss_20b_long-84c77bf as a useful negative long-payload capacity smoke: shorter long payloads mostly work directly, but largest payloads are confounded by sender XML/output-length failures rather than giving a clean capacity estimate