Land the post-remake refresh pieces held back by the quota

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

scripts/publish_atomic_results.py

@@ -295,6 +295,29 @@ def method_specs(root: Path, family: str, tasks: tuple[str, ...]) -> dict[str, t
     return out
 
 
+def alm_specs(root: Path, family: str, tasks: tuple[str, ...]):
+    """Specs for the remade authorial ensemble, or None until it exists.
+
+    ``scripts/alm_attribution_rows.py`` writes one predictions/metrics pair per
+    task from the remade models' scores. The directory is absent until the
+    remake has landed, and a leaderboard build before then simply carries no
+    ensemble row rather than a stale one.
+    """
+    family_root = root.parent / "alm-remake-20260902" / "leaderboard" / family
+    if not family_root.is_dir():
+        return None
+    specs, models = [], set()
+    for task in tasks:
+        metrics_path = family_root / task / "metrics.json"
+        if not metrics_path.is_file():
+            return None
+        metrics = read_json(metrics_path)
+        models.update(metrics["models"])
+        specs.append((task, family_root / task / "predictions.jsonl", metrics_path,
+                      float(metrics["test_metrics"]["macro_f1"])))
+    return specs, len(models)
+
+
 def dimensions_for_specs(specs: list[tuple[str, Path, Path, float]]) -> list[int]:
     return [feature_count(read_json(metrics_path)) for _, _, metrics_path, _ in specs]
 
@@ -385,6 +408,31 @@ def make_rows(results_root: Path, sphragis_data: Path, metre_data: Path):
         "kainobert-syntax", "kainobert-char-tfidf", "kainobert-tfidf-alm",
         "kainobert-burrows-alm", "kainobert-burrows-char", "kainobert-char-alm",
     )
+    alm_main = alm_specs(results_root, "sphragis-main", MAIN_TASKS)
+    alm_source = alm_specs(results_root, "sphragis-source-held-out", SOURCE_TASKS)
+    alm_metre = alm_specs(results_root, "sphragis-metre", VERSE_TASKS)
+    alm_status = (
+        "One OLMo-1B model further pretrained per author on the current release with "
+        "validation-loss early stopping; attribution is the lowest mean per-token surprisal; "
+        "no base-model or epoch search; test evaluated once"
+    )
+    alm_rows = []
+    if alm_main and alm_source:
+        alm_rows.append(build_row(
+            model="OLMo-1B authorial language models",
+            url="https://github.com/Urdatorn/sphragis_models",
+            dimensionality=f"1.18B parameters per author; {alm_main[1]} and {alm_source[1]} models",
+            classifier="Lowest mean per-token surprisal", status=alm_status,
+            task_specs=alm_main[0] + alm_source[0], dataset_roots=roots))
+    alm_metre_rows = []
+    if alm_metre:
+        alm_metre_rows.append(build_row(
+            model="OLMo-1B authorial language models",
+            url="https://github.com/Urdatorn/sphragis_models",
+            dimensionality=f"1.18B parameters per author; {alm_metre[1]} models",
+            classifier="Lowest mean per-token surprisal", status=alm_status,
+            task_specs=alm_metre[0], dataset_roots=roots))
+
     method_main = method_specs(results_root, "sphragis-main", MAIN_TASKS)
     method_source = method_specs(results_root, "sphragis-source-held-out", SOURCE_TASKS)
     method_sphragis: dict[str, tuple] = {}
@@ -644,6 +692,7 @@ def make_rows(results_root: Path, sphragis_data: Path, metre_data: Path):
             )
             for pair in pairs
         ),
+        *alm_rows,
         *(
             build_row(
                 model=pair_model_name(combo),
@@ -759,6 +808,7 @@ def make_rows(results_root: Path, sphragis_data: Path, metre_data: Path):
             )
             for pair in metre_pairs if pair in metre_pair_data
         ),
+        *alm_metre_rows,
         *(
             build_row(
                 model=pair_model_name(combo),