_claude_memory/feedback_experiments_md_live_doc.md

metadata

name: EXPERIMENTS.md is the live paper-grade master doc — update on every new result
description: >-
  Standing instruction — every time a new experiment result lands, append/update
  the corresponding row in regureasoner_loop/docs/EXPERIMENTS.md, not just in
  results/*.md
type: feedback
originSessionId: 4037f43b-2133-46c6-84bd-02f7d454ec8b

EXPERIMENTS.md (regureasoner_loop/docs/EXPERIMENTS.md on the mllm-integrate-server2 branch) is the canonical paper-section draft target. Every time a new bench/training/oracle result lands, update the matching section in EXPERIMENTS.md AND push to the remote branch.

Why: the user explicitly asked: "update this EXPERIMENTS.md every time you have new experiment results" (2026-04-27). The doc is intended as the experiment section for a top AI conference paper; results/*.md files are raw data tables that get referenced from EXPERIMENTS.md, but EXPERIMENTS.md is the narrative + headline numbers source of truth.

How to apply:

• When a metrics.json / genqual.json / training final.json lands, open EXPERIMENTS.md, find the matching §3.x or §4-§7 section, replace any "queued" / "running" placeholder with the real number, and append a paired comparison block when both raw and enriched rows are available.
• Bump §8 "Pipeline state (live)" to reflect the new state and what triggered next.
• Add a row to §10 "Code commits" with the commit SHA + 1-line note.
• Always commit + push to mllm-integrate-server2 immediately (companion results/<task>_*.md if the per-result table is long).
• The HF mirror of EXPERIMENTS.md is blocked by sandbox policy — do not retry; the canonical copy is the GitHub branch.