From 3e9471c2a7ff6c8482d4970a2d0be165d52212bb Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Tue, 1 Sep 2026 17:17:09 +0200 Subject: [PATCH] docs: paper alignment map + title-level flag for the user --- docs/PAPER-ALIGNMENT.md | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+) create mode 100644 docs/PAPER-ALIGNMENT.md diff --git a/docs/PAPER-ALIGNMENT.md b/docs/PAPER-ALIGNMENT.md new file mode 100644 index 0000000..6266780 --- /dev/null +++ b/docs/PAPER-ALIGNMENT.md @@ -0,0 +1,25 @@ +# Paper alignment (2026-09-01) + +The LaTeX paper (rababa docs/paper-arabic/main.tex) — "Data Quality +over Scale: A 30M-Parameter Encoder for Arabic Diacritization" — +predates the campaign's later results. Slotting map for refresh: + +| Their section | Current evidence to slot in | +|---|---| +| Data | r7 news-domain adaptation (ID 2.2864 / OOD 17.38-11.83 — dual-surface) | +| Model | r8 aux ablation (morph 2.5793 < IPA 2.6588 < none 2.6775) | +| Experiments | CIs via paired bootstrap; GLM-5.3-Flash row (8.59/8.90); 2.0 client tier 4.82; frontier incl. lite 5.78 | +| Reproducibility | labels sha256 provenance; final_preds.jsonl; five-instance subset phenomenon | + +**FLAG for the user (title-level):** the capacity law measured AFTER +the draft — from-scratch 33M collapses at 74.68 full-set even with +clean labels (run-005), and naive stitching fails at both ratios — +sits in direct tension with a "30M-parameter encoder" as a headline +claim, UNLESS the paper's 30M encoder is a differently-trained +data-quality artifact we have not re-measured. The title/thesis needs +the user's decision: refresh the numbers (and possibly the claim), or +reposition as teacher-tier "data quality over scale" (r7 vs frontier +LLMs) with the client tier as the measured boundary. + +My paper-a.adoc scaffold (teacher thesis) remains an alternative +split per the 2026-08-29 two-paper decision.