interscript · ronaldtse · Sep 2, 2026 · Sep 2, 2026
diff --git a/TODO.publish/01-glm-4-7-row-completion.md b/TODO.publish/01-glm-4-7-row-completion.md
@@ -1,12 +1,18 @@
 # 01 — glm-4.7-flash SadeedDiac-25 row completion
 
-Status: IN FLIGHT (2026-09-02). Fetch state: 992/1,200 distinct rows
-landed, 208 still empty after three passes — the endpoint sits behind
-sustained 429s (code 1305; thinking-disabled accepted, plain
-completion inexpressible). Clean single fetch pass running
-(/tmp/glm47_eval4.log; the #65 guard drops the 208 empties on
-resume). Interim tables from contaminated interleaved passes are
-VOID — only the final full-fetch tables count.
+Status: COMPLETE (2026-09-02). 1,200/1,200, zero empties after 12
+resumable passes past sustained 429s (sentinels 249 -> 208 -> 79 ->
+37 -> 20 -> 13 -> 6 -> 4 -> 2 -> 0). Final: **13.0035 raw /
+13.2256 zero-skip** (w/o-CE 10.0510/10.3206; NFDW 17.60 raw).
+Attribution (validated scorer — reproduces the #69 series exactly:
+5.2 wrong 2.64%, 5.3 missing 3.75/wrong 8.59): **missing 6.67%,
+wrong 9.01%, extra 0.96%** — the only family member regressing on
+both axes; U+0670 convention effect 0.04pp. Bootstrap vs GLM-5.2:
++7.945pp, CI [+7.516, +8.370], p<1e-4. Regression axis complete:
+5.2 2.5060 -> 5.3-Flash 8.5721 -> 5.3 9.9760 -> 4.7-flash 13.0035
+raw. Recorded: results dir + RESULTS.md section/ledger/bootstrap
+rows; paper.adoc row; site note amended (interscript.github.io
+#149).
 
 ## Protocol
 

diff --git a/TODO.publish/04-site-frontier-sentence.md b/TODO.publish/04-site-frontier-sentence.md
@@ -1,9 +1,11 @@
 # 04 — Site frontier sentence (interscript.org ml page)
 
-Status: BLOCKED on TODO.publish/01 (needs the full regression axis).
-The ml.astro page exists on interscript.org (rebrand 2026-08); what
-it lacks is the one-line frontier-context claim that the GLM sweep
-now substantiates.
+Status: COMPLETE (2026-09-02). Frontier-context note merged on the
+site ledger section (interscript.github.io #147), amended with the
+fourth row glm-4.7-flash 13.00 raw DER (#149) — wording "the rest of
+the family" since 4.7 is a predecessor generation. Links
+rababa/docs/RESULTS.md for the disclosed rows. Site guard (34/34),
+astro check, full unit suite (255/255) green on both PRs.
 
 ## The claim (draft, number slots to fill from 01)
 

diff --git a/TODO.training-work/02-glm-4-7-flash-row-completion.md b/TODO.training-work/02-glm-4-7-flash-row-completion.md
@@ -1,10 +1,11 @@
 # 02 — GLM-4.7-Flash row completion
 
-Status: IN FLIGHT (2026-09-01). Old API semantics verified by probe:
-thinking.type=disabled ACCEPTED (plain completion expressible), but
-the endpoint sits behind sustained 429s (code 1305) — running at
-GLM_WORKERS=1, ~2 rows/min aggregate; ETA overnight. Checkpoint
-resumes; empties self-heal (#65).
+Status: COMPLETE (2026-09-02) — 1,200/1,200 after 12 resumable
+passes past sustained 429s (sentinels 249 -> ... -> 0). 13.0035 raw
+/ 13.2256 zero-skip; attribution missing 6.67 / wrong 9.01 / extra
+0.96; bootstrap vs GLM-5.2 +7.945pp CI [+7.516, +8.370]. RESULTS +
+ledger + paper row + site amendment landed. Full record:
+TODO.publish/01.
 
 ## Remaining steps
 

diff --git a/docs/RESULTS.md b/docs/RESULTS.md
@@ -433,6 +433,21 @@ behind 5.2 (CI [-18,028, -14,685] error positions). The dedicated-
 model thesis gains its cleanest form: the newest generalist
 generation lost classical-Arabic mark knowledge its predecessor had.
 
+## glm-4.7-flash on SadeedDiac-25 (2026-09-02)
+
+**13.2256/10.3206 zero-skip, 13.0035/10.0510 raw** (Misraj;
+`results/sadeed-glm-4-7-flash/`; 1,200/1,200, zero empties — the
+endpoint sat behind sustained 429s and needed 12 resumable passes,
+sentinels dropping 249 -> ... -> 0). thinking-disabled accepted:
+the last plain-completion GLM row. Worst frontier DER, and the only
+family member regressing on both axes (missing 6.67% — worst of the
+family, NFDW 17.6 — plus wrong 9.01%); U+0670 convention effect
+0.04pp. Bootstrap vs GLM-5.2: +7.945pp, CI [+7.516, +8.370],
+p<1e-4. The regression axis is complete: GLM-5.2 2.5060 raw ->
+GLM-5.3-Flash 8.5721 -> GLM-5.3 9.9760 -> glm-4.7-flash 13.0035 —
+every successor generation lost classical-Arabic mark knowledge its
+predecessor had.
+
 ## Hebrew s46 on the Dicta test corpora (2026-09-01)
 
 The disclosed modern-text gap, measured (eval_hebrew_dicta.py, TODO
@@ -496,6 +511,7 @@ row here does not belong in the paper.
 | GLM-5.2 (2.5060 raw / 2.6911 zero-skip) | same evaluator | temp 0, `thinking.type=disabled` (plain completion) | raw: word-structure skips; zero-skip: projected | yes — script + CSVs (results/sadeed-glm-5-2/) |
 | GLM-5.3-Flash (8.5721 raw / 8.7978 zero-skip) | same evaluator | temp 0, `reasoning_effort=low` (thinking CANNOT be disabled — API 400 code 1210) | same; dagger-alif convention skips; 0 empty responses after sentinel purge | yes — script + CSVs (results/sadeed-glm-5-3-flash/) |
 | GLM-5.3 (9.9760 raw / 9.8971 zero-skip) | same evaluator | temp 0, `reasoning_effort=low` (same 400/1210 rejection) | same; 41 sentinels caught + refetched, 0 empties final | yes — script + CSVs (results/sadeed-glm-5-3/) |
+| glm-4.7-flash (13.0035 raw / 13.2256 zero-skip) | same evaluator | temp 0, `thinking.type=disabled` (still accepted on 4.x) | same; 12 passes past sustained 429s, sentinels 249 -> ... -> 0, 0 empties final | yes — script + CSVs (results/sadeed-glm-4-7-flash/) |
 | Claude-3.7-Sonnet (1.3941) | published number | their protocol, undisclosed to us | unknown | no — vendor-published |
 | Gemini-Flash-2.0 (3.1926) / GPT-4 (3.8645) | published numbers | theirs | unknown | no — vendor-published |
 | Sadeed-1.5B (7.2915) | published, same benchmark | theirs (their repo reports 1.2 under its own split — not comparable) | unknown | partially — paper + code, split differs |
@@ -519,6 +535,7 @@ identical.
 | r7 vs r6 | −0.367 | [−0.467, −0.278] | <1e-4 — improvement is real |
 | r7 vs GLM-5.2 (projected) | +0.539 (GLM worse) | [+0.202, +0.939] | 0.0003 — our teacher leads |
 | GLM-5.3-Flash vs GLM-5.2 (projected) | +7.966 | [+7.520, +8.368] | <1e-4 — the frontier regression is overwhelming |
+| glm-4.7-flash vs GLM-5.2 (projected) | +7.945 | [+7.516, +8.370] | <1e-4 — the axis endpoint; 4.7 regresses on both missing and wrong |
 
 Policy (B1 of the improvement plan): every delta quoted in docs or
 the paper carries its CI or an explicit note that the run predates

diff --git a/results/sadeed-glm-4-7-flash/README.md b/results/sadeed-glm-4-7-flash/README.md
@@ -0,0 +1,50 @@
+# glm-4.7-flash on SadeedDiac-25 (2026-09-02)
+
+Same harness as every LLM row: neutral prompt, `temperature 0`,
+structure-preserving cleanup, Misraj evaluator, 1,200/1,200 responses,
+zero empty responses. Decode protocol: `thinking: {"type": "disabled"}`
+— glm-4.7-flash still accepts disabled thinking (unlike the 5.x
+generation, which rejects it with HTTP 400 code 1210), so this row is
+the last plain-completion GLM row.
+
+Provider note: the endpoint sat behind sustained 429s (code 1305) —
+the full set took 12 resumable passes (the #65 guard dropped
+exhausted-retry empty sentinels between passes: 249 -> 208 -> 79 ->
+37 -> 20 -> 13 -> 6 -> 4 -> 2 -> 0). All 1,200 final rows are real
+responses.
+
+## Results (Misraj evaluator, percentages)
+
+| Protocol | DER (CE) | DER (w/o CE) | WER (CE) | WER (w/o CE) | NFDW |
+|---|---|---|---|---|---|
+| raw | **13.0035** | 10.0510 | 32.4140 | 26.7750 | 17.60 |
+| projected zero-skip | **13.2256** | 10.3206 | 36.0985 | 30.1002 | 18.56 |
+
+## Attribution (same 1,200 paragraphs, convention-normalized, rates over GT-marked positions)
+
+| model | missing | **wrong haraqat** | extra | protocol |
+|---|---|---|---|---|
+| our r7 teacher | 0.11% | **2.62%** | 0.15% | greedy argmax |
+| GLM-5.2 | 0.61% | **2.64%** | 0.17% | thinking disabled |
+| GLM-5.3-Flash | 1.01% | **10.05%** | 0.20% | reasoning_effort=low |
+| GLM-5.3 (full) | 3.75% | **8.59%** | 0.11% | reasoning_effort=low |
+| glm-4.7-flash | **6.67%** | **9.01%** | 0.96% | thinking disabled |
+
+glm-4.7-flash is the worst frontier row on DER and the only one that
+regresses on BOTH axes simultaneously: the highest missing rate of
+the family (6.67% — it under-diacritizes, NFDW 17.6%) alongside a
+5.3-class wrong rate (9.01%). The U+0670 convention effect is
+0.04pp — negligible, consistent with the family. Paired bootstrap vs
+GLM-5.2 (paragraph-level simplified scorer, 10,000 resamples, seed
+42): **+7.945pp, 95% CI [+7.516, +8.370], one-sided p<1e-4** —
+decisively behind 5.2, whose plain-completion DER was 2.5060 raw /
+2.6911 zero-skip.
+
+The regression axis is now complete: GLM-5.2 2.51 -> GLM-5.3-Flash
+8.57 -> GLM-5.3 9.98 -> glm-4.7-flash 13.00 raw DER. Every successor
+generation lost classical-Arabic mark knowledge its predecessor had;
+the dedicated-model thesis holds across the whole measurable family.
+
+## Files
+
+- `sadeed_preds_raw.csv`, `sadeed_preds_projected.csv`