docs: GLM-5.3-Flash attribution — wrong haraqat (10.05%), not convention (0.125pp)

docs/RESULTS.md

@@ -66,12 +66,19 @@ protocol, zero skipped paragraphs.
   REJECTS disabled thinking (HTTP 400 code 1210) — reasoning cannot be
   turned off, only dialed to low/high/max — so this ran at
   `reasoning_effort=low`, the nearest analog to plain completion.
-  ~3.4x worse DER than GLM-5.2 and behind Sadeed-1.5B, driven by
-  Quranic-convention orthography (dagger-alif forms the GT does not
-  use, 9.8% not-fully-diacritized words) — a generalist regression on
-  classical-knowledge output conventions, recorded as a dedicated-model
-  data point. Zero empty responses after purging 140 checkpoint rows
-  the pre-fix payload had retried into empty strings.
+  ~3.4x worse DER than GLM-5.2 and behind Sadeed-1.5B. Attribution
+  (2026-09-01, per-position decomposition over the same 1,200
+  paragraphs, convention-normalized): **wrong-haraqat rate 10.05% vs
+  GLM-5.2's 2.64%** — which matches our r7 teacher's 2.62% almost
+  exactly — while missing is 1.01%, extra 0.20%, and the entire
+  dagger-alif (U+0670) convention effect is 0.125pp (310 marks, zero
+  in GT; rules derived from the aligned positions: drop after ى,
+  fatha elsewhere; controls r7/G-5.2 move <=0.009pp). The regression
+  is genuine mark errors, not orthographic convention — the frontier
+  generalist lost core classical-Arabic knowledge its predecessor
+  had. The dagger-alif observation explains the raw-mode skip flood,
+  not the DER gap. Zero empty responses after purging 140 checkpoint
+  rows the pre-fix payload had retried into empty strings.
 
 - **r3 = r2 + 1 epoch on the decontaminated Misraj corpus (1M lines) +
   150k MSA replay**: best non-frontier-LLM result on the benchmark;

results/sadeed-glm-5-3-flash/README.md

@@ -23,17 +23,27 @@ For reference, GLM-5.2 (thinking disabled): raw 2.5060/1.5537/7.9929,
 zero-skip 2.6911/1.7179/8.3037. GLM-5.3-Flash is ~3.4x worse on DER
 than its predecessor under the nearest equivalent protocol.
 
-## Why the delta is real but protocol-shaped
-
-- **Orthography**: the model emits Quranic-convention marks (dagger
-  alif: عَلَىٰ, هٰذِهِ, ذَٰلِكَ; also بِهِۦ) where the benchmark's GT uses
-  plain MSA forms. The evaluator skips whole sentences on word
-  mismatches (survivorship in raw mode); surviving dagger-alif words
-  count as diacritic errors. Not-fully-diacritized words run 9.8%.
-- **Thinking floor**: reasoning cannot be turned off, so the
-  plain-completion protocol the published LLM rows used is not
-  expressible for this model; `low` still engaged reasoning on long
-  paragraphs (reasoning_content observed in-flight).
+## Why the delta is real (attribution, 2026-09-01)
+
+Per-position decomposition over the same 1,200 paragraphs,
+convention-normalized (U+0670 rules derived from aligned positions:
+drop after ى, fatha on other letters; controls move <=0.009pp):
+
+| model | missing | **wrong haraqat** | extra | U+0670 convention |
+|---|---|---|---|---|
+| our r7 teacher | 0.11% | **2.62%** | 0.15% | — |
+| GLM-5.2 | 0.61% | **2.64%** | 0.17% | ~0.009pp |
+| GLM-5.3-Flash | 1.01% | **10.05%** | 0.20% | 0.125pp |
+
+The regression is overwhelmingly WRONG haraqat at ~4x its
+predecessor's rate (which matches our dedicated teacher's to within
+0.02pp) — not the dagger-alif orthography (0.125pp total), not
+under-diacritization (1.01%), not thinking overhead. The Quranic
+marks (عَلَىٰ, هٰذِهِ, ذَٰلِكَ; 310 in the outputs, zero in GT) explain the
+raw-protocol evaluator skips, not the DER gap.
+- **Thinking floor** (protocol caveat, unchanged): reasoning cannot be
+  turned off; `low` still engaged reasoning on long paragraphs
+  (reasoning_content observed in-flight).
 - **Measurement hygiene**: an initial run resumed from a 140-row
   checkpoint whose rows were empty responses produced by the
   pre-fix both-knobs payload (HTTP 400s retried into empty strings) —