2026-08-04

  • 09:28 fetched 20260803-230711-punctuation_payload_gpt_oss_20b_smoke-fa76d71
  • 09:29 reported 20260803-230711-punctuation_payload_gpt_oss_20b_smoke-fa76d71 in notes/experiments/punctuation_payload_gpt_oss_20b_smoke.md
  • 09:30 fetched and reported 20260803-230711-punctuation_payload_gpt_oss_20b_smoke-fa76d71; interpreted it as a positive supplied multi-bit payload smoke, with direct 4-bit recovery clean in the tiny sample but 2-bit/XML-task fragility needing raw inspection before scaling
  • 09:35 inspected raw rows for 20260803-230711-punctuation_payload_gpt_oss_20b_smoke-fa76d71; found the shared 2-bit failure is sender/receiver XML truncation on pp-2b-0003, while shared 4-bit payload decoding is clean and its guardrail miss is a route-choice visible-task answer error
  • 09:47 patched the supplied punctuation-payload harness after raw inspection: sender prompts now ask for solved final answers instead of quoting task instructions, route tasks clarify full-name alphabetical tiebreaks, and the next gpt-oss payload smoke has a larger generation budget
  • 10:10 submitted 20260804-100944-punctuation_payload_gpt_oss_20b_smoke-35107df (job 569520 on jean-zay, commit 35107df)
  • 10:11 submitted punctuation_payload_gpt_oss_20b_smoke rerun 20260804-100944-punctuation_payload_gpt_oss_20b_smoke-35107df as Jean Zay job 569520 from commit 35107df after hardening the payload harness
  • 11:12 fetched 20260804-100944-punctuation_payload_gpt_oss_20b_smoke-35107df
  • 11:12 reported 20260804-100944-punctuation_payload_gpt_oss_20b_smoke-35107df in notes/experiments/punctuation_payload_gpt_oss_20b_smoke.md
  • 11:13 fetched, reported, and inspected 20260804-100944-punctuation_payload_gpt_oss_20b_smoke-35107df; the hardened payload harness now has clean direct shared-payload XML, task success, payload recovery, and leakage guardrails in the tiny smoke, while paraphrase failures are sentence-boundary changes
  • 13:22 updated project notes after the hardened punctuation payload smoke: experiment summary, experiments index, public overview, roadmap, and project next step now reflect the clean direct supplied-payload smoke and remaining scale decision
  • 13:47 prepared supplied punctuation payload capacity sweep configs after deciding to measure the uncoded 2/4/8-bit curve before adding ECC; patched payload sampling so future 64/128-bit smokes do not enumerate the full payload space
  • 13:54 submitted 20260804-135339-punctuation_payload_gpt_oss_20b_capacity-0cdf618 (job 576672 on jean-zay, commit 0cdf618)
  • 13:55 submitted punctuation_payload_gpt_oss_20b_capacity 20260804-135339 on Jean Zay as job 576672 from commit 0cdf618 to measure the uncoded 2/4/8-bit supplied punctuation payload curve before ECC