diff --git a/src/gpu/modal_golden_matrix.py b/src/gpu/modal_golden_matrix.py
new file mode 100644
index 0000000..541877a
--- /dev/null
+++ b/src/gpu/modal_golden_matrix.py
@@ -0,0 +1,99 @@
+"""Modal app: generate per-model golden sets from the RELEASED zips —
+the cross-runtime byte-parity corpus. For every models.yaml entry,
+decode N test inputs through the Python runtime using the exact
+release artifact (sha-pinned to the index) and write
+golden/<model-id>.jsonl rows {input, output}. Published as a release
+asset so the TS and Ruby CI jobs verify against the same bytes.
+
+    modal run --detach src/gpu/modal_golden_matrix.py --n 25
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import modal
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+
+IMAGE = (
+    modal.Image.debian_slim(python_version="3.11")
+    .pip_install(
+        "onnxruntime==1.23.2",
+        "numpy>=1.26",
+        "pyyaml>=6.0",
+    )
+    .add_local_dir(str(REPO_ROOT), "/root/interscript-ml", copy=True)
+    .add_local_dir(str(REPO_ROOT.parent / "rababa"), "/opt/rababa", copy=True)
+    .workdir("/root/interscript-ml")
+)
+
+MODELS_VOLUME = modal.Volume.from_name("secryst-models")
+DATASETS = modal.Volume.from_name("rababa-datasets")
+SECRYST_DATASETS = modal.Volume.from_name("secryst-datasets")
+
+app = modal.App("interscript-ml-golden-matrix", image=IMAGE)
+
+# index id -> volume zip path + test pairs (volume-relative)
+SOURCES = {
+    "khm-latn-1.0": ("imf/khm-latn/khm-latn-1.0-fp32.zip", "/secryst-datasets/khmer-translit/test.jsonl"),
+    "urd-g2p-1.0": ("imf/urd-g2p/urd-g2p-1.0-fp32.zip", "urdu-g2p/test.jsonl"),
+    "urd-diac-1.0": ("imf/urd-diac/urd-diac-1.0-fp32.zip", "urdu-diacrit/test.jsonl"),
+    "tha-g2p-base-1.0": ("imf/tha-g2p-base/tha-g2p-base-1.0-fp32.zip", "thai-ipa/test.jsonl"),
+    "tha-g2p-small-1.0": ("imf/tha-g2p-small/tha-g2p-small-1.0-int8.zip", "thai-ipa/test.jsonl"),
+    "fas-g2p-1.0": ("imf/fas-g2p/fas-g2p-1.0-fp32.zip", None),
+    "heb-diac-1.0": ("imf/heb-diac/heb-diac-1.0-fp32.zip", "nakdimon/test-imf.jsonl"),
+    "heb-diac-1.1": ("imf/heb-diac/heb.zip", "nakdimon/test-imf.jsonl"),
+    "heb-diac-small-1.0": ("imf/heb-diac-small/heb-diac-small-1.0-fp32.zip", "nakdimon/test-imf.jsonl"),
+    "ara-diac-1.0": ("imf/ara-diac/ara-diac-1.0-fp32.zip", "arabic-sadeed-imf/test.jsonl"),
+    "ara-diac-2.0-int8": ("imf/ara-diac2/ara-diac-2.0-int8.zip", "arabic-sadeed-imf/test.jsonl"),
+    "ara-diac-small-2.1-int8": ("imf/ara-diac-small-21/ara-diac-small-2.1-int8.zip", "arabic-sadeed-imf/test.jsonl"),
+    "ara-diac-layerdrop-1.0-int4": ("imf/ara-diac-layerdrop/ara-diac-layerdrop-1.0-int4.zip", "arabic-sadeed-imf/test.jsonl"),
+}
+
+
+@app.function(cpu=8, memory=24 * 1024, timeout=4 * 3600,
+              volumes={"/outputs": MODELS_VOLUME, "/datasets": DATASETS, "/secryst-datasets": SECRYST_DATASETS})
+def generate(n: int = 25) -> dict:
+    import hashlib
+    import json
+    import sys
+
+    sys.path.insert(0, "/root/interscript-ml/runtime/src")
+    from interscript_ml import Model
+
+    out_dir = Path("/outputs/golden")
+    out_dir.mkdir(exist_ok=True)
+    report = {}
+    for model_id, (zip_path, test_path) in SOURCES.items():
+        zip_full = Path("/outputs") / zip_path
+        if not zip_full.exists():
+            report[model_id] = "zip missing on volume"
+            continue
+        if test_path is None:
+            report[model_id] = "no test pairs known — skipped"
+            continue
+        pairs = []
+        for line in Path(test_path).read_text(encoding="utf-8").splitlines():
+            if line.strip():
+                row = json.loads(line)
+                pairs.append(row.get("src") or row.get("input"))
+        pairs = [p for p in pairs if p][:n]
+        model = Model.load(zip_full)
+        rows = []
+        for src in pairs:
+            out = model.translate(src, max_len=max(256, 4 * len(src)))
+            rows.append({"input": src, "output": out})
+        digest = hashlib.sha256()
+        body = "".join(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
+        digest.update(body.encode("utf-8"))
+        (out_dir / f"{model_id}.jsonl").write_text(body, encoding="utf-8")
+        report[model_id] = {"n": len(rows), "sha256": digest.hexdigest()}
+        print(f"{model_id}: {len(rows)} rows", flush=True)
+    MODELS_VOLUME.commit()
+    return report
+
+
+@app.local_entrypoint()
+def main(n: int = 25) -> None:
+    print(generate.remote(n))