secryst crystals: native SECRYST_* env vars (rename outright, pre-publication)

docs/RESULTS.md

@@ -0,0 +1,30 @@
+# ml-models results
+
+Evaluated results for models produced in this repository. Each section
+anchor is the provenance target referenced by model metadata
+(`models/*/` metadata.yaml) and `models/metrics-sources.yaml`.
+
+## tha-g2p-base-1.0 — Thai G2P distillation (2026-08-19)
+
+Sequence-level KD: B-K/umt5-thai-g2p-v2-0.5k teacher -> ByT5-base
+student over 48,757 beam-4 teacher-generated labels (Kaikki + epitran
+Wikipedia corpus, deduplicated, degenerate outputs filtered). Harness:
+beam-4, corpus-level PER (total_ed / total_gold over characters of
+joined-piece decode), 1,219 held-out Kaikki Thai test sentences
+(`src/gpu/modal_distill.py::evaluate_per`).
+
+| Model | PER | Exact match |
+|---|---|---|
+| Teacher (B-K/umt5 hub base) | 4.43% | 95.57% |
+| **Student (ByT5-base, gate)** | **9.19%** | 90.81% |
+
+Distillation cost: +4.76pp, inside the +5pp budget
+(docs/DISTILL-SOURCE-PROMPT.md). ByT5-small ablations for reference:
+12.63% on 23K labels, 12.06% on 48.7K labels (capacity-limited, both
+rejected by the gate).
+
+Context: the secryst-published 2.32% umt5 teacher is unrecoverable
+from saved artifacts (transformers 5.15 save drops the untied umt5
+lm_head) and the volume's epitran augmentation corpus is tone-less;
+this release distills the best verified teacher available. A repaired
+2.32%-tier teacher re-enters this pipeline when secryst regenerates it.

models.yaml

@@ -10,8 +10,8 @@
 #      verify sha256, atomically rename into place
 #   4. load the zip (IMF v1: member sha256 verification on load)
 #
-# Overrides: INTERSCRIPT_ML_INDEX (URL or path to an index like this one),
-# INTERSCRIPT_ML_CACHE (cache directory; default ~/.cache/interscript).
+# Overrides: SECRYST_INDEX (URL or path to an index like this one),
+# SECRYST_CACHE (cache directory; default ~/.cache/secryst).
 #
 # Adding a model: it must have passed the WO03 gate (strict validation,
 # parity written into the zip) before an entry ships here.

models/metrics-sources.yaml

@@ -40,3 +40,12 @@ heb-diac-1.0:
     - row: "s43 (production)"
       as: der_beam4
       protocol: "beam=4 standard decode (reference quality; beam search is not in v1 runtimes); Nakdimon test split, 5,095 examples; ByT5-base s43"
+tha-g2p-base-1.0:
+  repo: interscript/ml-models
+  ref: main
+  path: docs/RESULTS.md
+  anchor: tha-g2p-base-1.0
+  protocol: "beam-4, corpus-level PER (total_ed/total_gold over chars of joined-piece decode); 1,219 held-out Kaikki Thai test sentences"
+  tables:
+    - {row: "Teacher (B-K/umt5 hub base)", column: PER, as: per_teacher}
+    - {row: "Student (ByT5-base, gate)", column: PER, as: per_student}

models/tha-g2p-base/tha-g2p-base-1.0.metadata.yaml

@@ -23,10 +23,10 @@ metrics:
       beam-4, corpus-level PER (total_ed/total_gold over chars of
       joined-piece decode); 1,219 Kaikki Thai test sentences;
       B-K/umt5-thai-g2p-v2-0.5k teacher
-    source: interscript/ml-models src/gpu/modal_distill.py::evaluate_per
+    source: interscript/ml-models docs/RESULTS.md#tha-g2p-base-1.0
   - name: per_student
     value: 9.19
     protocol: >-
       beam-4, corpus-level PER, same harness as the teacher; gate
       +4.76pp <= +5pp (docs/DISTILL-SOURCE-PROMPT.md); exact match 90.81%
-    source: interscript/ml-models release tha-g2p-base-1.0
+    source: interscript/ml-models docs/RESULTS.md#tha-g2p-base-1.0

npm/models/package.json

@@ -1,7 +1,7 @@
 {
-  "name": "@interscript/models",
+  "name": "@secryst/models",
   "version": "0.0.1",
-  "description": "Manifest of available Interscript ML models + their current versions, URLs, and checksums.",
+  "description": "Manifest of available secryst (interscript-ml contract) models + their current versions, URLs, and checksums.",
   "license": "MIT",
   "main": "manifest.json",
   "files": [

runtime/README.md

@@ -1,9 +1,14 @@
-# interscript-ml (Python runtime)
+# secryst (Python crystal)
 
-The reference Python runtime for **IMF v1** model zips — the phonological
-layer of Interscript. The Ruby (secryst gem) and TypeScript
-(@interscript/ml) runtimes are diffed against this one on shared golden
-sets.
+The Python crystal — reference implementation of **IMF v1** model zips
+and the `models.yaml` index (the **interscript-ml** contract), the
+phonological layer of Interscript. The Ruby (`secryst` gem) and
+TypeScript (`@secryst/ml`) crystals are diffed against this one on
+shared golden sets. This crystal owns golden generation and numerical
+adjudication for the family.
+
+Home repo: https://github.com/secryst/secryst-py (this copy in
+ml-models/runtime is the frozen origin; the package now lives there).
 
 ```python
 from interscript_ml import Model
@@ -25,14 +30,14 @@ model = Model.load("khm-latn-1.0.zip")    # or: a local zip path directly
 - Dynamic fetch per the `models.yaml` contract (shared with the Ruby and
   TypeScript runtimes): resolve id -> channel URL, download to temp,
   verify whole-file sha256 against the index, atomically install into
-  `~/.cache/interscript/models/<id>/`. Overrides:
-  `INTERSCRIPT_ML_INDEX` (URL or path), `INTERSCRIPT_ML_CACHE`.
+  `~/.cache/secryst/models/<id>/`. Overrides:
+  `SECRYST_INDEX` (URL or path), `SECRYST_CACHE`.
 
 Install: `pip install ./runtime` (from the ml-models checkout) or
 `pip install -e "./runtime[dev]"` for development.
 
 Tests: `python -m pytest runtime/tests` — tiny-graph zips, no torch
-needed. The end-to-end golden test runs when `INTERSCRIPT_ML_E2E_ZIP`
+needed. The end-to-end golden test runs when `SECRYST_E2E_ZIP`
 points at a real zip (e.g. `models/khm-latn/khm-latn-1.0-fp32.zip`)
 and asserts byte-identical outputs against `golden/khm-latn-100.jsonl`.
 

runtime/src/interscript_ml/registry.py

@@ -21,8 +21,8 @@
 DEFAULT_INDEX_URL = (
     "https://raw.githubusercontent.com/interscript/ml-models/main/models.yaml"
 )
-ENV_INDEX = "INTERSCRIPT_ML_INDEX"
-ENV_CACHE = "INTERSCRIPT_ML_CACHE"
+ENV_INDEX = "SECRYST_INDEX"
+ENV_CACHE = "SECRYST_CACHE"
 
 
 class RegistryError(ValueError):
@@ -51,7 +51,7 @@ class IndexEntry:
 def cache_dir() -> Path:
     if os.environ.get(ENV_CACHE):
         return Path(os.environ[ENV_CACHE])
-    return Path.home() / ".cache" / "interscript"
+    return Path.home() / ".cache" / "secryst"
 
 
 def load_index(index_url: str | None = None) -> dict[str, IndexEntry]:

runtime/tests/test_model.py

@@ -2,7 +2,7 @@
 
 Tiny-graph zips built with the onnx package (no torch, no training
 repo). The end-to-end golden test runs only when a real zip is provided
-via INTERSCRIPT_ML_E2E_ZIP.
+via SECRYST_E2E_ZIP.
 """
 
 from __future__ import annotations
@@ -147,9 +147,9 @@ def test_missing_graph_rejected(tmp_path: Path) -> None:
 
 def test_golden_set_e2e() -> None:
     """Run against a real zip: byte-identical outputs on the golden set."""
-    zip_path = os.environ.get("INTERSCRIPT_ML_E2E_ZIP")
+    zip_path = os.environ.get("SECRYST_E2E_ZIP")
     if not zip_path:
-        pytest.skip("set INTERSCRIPT_ML_E2E_ZIP to a real IMF zip")
+        pytest.skip("set SECRYST_E2E_ZIP to a real IMF zip")
     golden = Path(__file__).resolve().parent.parent.parent / "golden" / "khm-latn-100.jsonl"
     if "khm" not in Path(zip_path).name:
         pytest.skip("golden file is khm-latn specific")

runtime/tests/test_registry.py

@@ -38,7 +38,7 @@ def test_resolve_downloads_verifies_and_caches(tmp_path: Path) -> None:
     zip_path = build_tiny_zip(tmp_path / "channel" / "tiny.zip")
     index = _index_file(tmp_path, zip_path)
     cache = tmp_path / "cache"
-    os.environ["INTERSCRIPT_ML_CACHE"] = str(cache)
+    os.environ["SECRYST_CACHE"] = str(cache)
     try:
         local = resolve("tiny-1.0", index_url=str(index))
         assert local == cache / "models" / "tiny-1.0" / "tiny.zip"
@@ -47,18 +47,18 @@ def test_resolve_downloads_verifies_and_caches(tmp_path: Path) -> None:
         zip_path.unlink()
         assert resolve("tiny-1.0", index_url=str(index)) == local
     finally:
-        os.environ.pop("INTERSCRIPT_ML_CACHE", None)
+        os.environ.pop("SECRYST_CACHE", None)
 
 
 def test_resolve_rejects_bad_download(tmp_path: Path) -> None:
     zip_path = build_tiny_zip(tmp_path / "channel" / "tiny.zip")
     index = _index_file(tmp_path, zip_path, sha256="0" * 64)
-    os.environ["INTERSCRIPT_ML_CACHE"] = str(tmp_path / "cache")
+    os.environ["SECRYST_CACHE"] = str(tmp_path / "cache")
     try:
         with pytest.raises(RegistryError, match="sha256 mismatch"):
             resolve("tiny-1.0", index_url=str(index))
     finally:
-        os.environ.pop("INTERSCRIPT_ML_CACHE", None)
+        os.environ.pop("SECRYST_CACHE", None)
 
 
 def test_resolve_unknown_id(tmp_path: Path) -> None:
@@ -71,15 +71,15 @@ def test_resolve_unknown_id(tmp_path: Path) -> None:
 def test_model_load_by_id(tmp_path: Path) -> None:
     zip_path = build_tiny_zip(tmp_path / "channel" / "tiny.zip")
     index = _index_file(tmp_path, zip_path)
-    os.environ["INTERSCRIPT_ML_CACHE"] = str(tmp_path / "cache")
+    os.environ["SECRYST_CACHE"] = str(tmp_path / "cache")
     try:
         from interscript_ml import Model
 
         model = Model.load("tiny-1.0", index_url=str(index))
         assert model.id == "tiny-1.0"
         assert isinstance(model.translate("he", max_len=4), str)
     finally:
-        os.environ.pop("INTERSCRIPT_ML_CACHE", None)
+        os.environ.pop("SECRYST_CACHE", None)
 
 
 def test_resolve_parts_assembles_and_verifies(tmp_path: Path) -> None:
@@ -118,7 +118,7 @@ def test_resolve_parts_assembles_and_verifies(tmp_path: Path) -> None:
     index_path = tmp_path / "models.yaml"
     index_path.write_text(yaml.safe_dump(index), encoding="utf-8")
     cache = tmp_path / "cache"
-    os.environ["INTERSCRIPT_ML_CACHE"] = str(cache)
+    os.environ["SECRYST_CACHE"] = str(cache)
     try:
         local = resolve("tiny-1.0", index_url=str(index_path))
         assert local == cache / "models" / "tiny-1.0" / "tiny.zip"
@@ -127,7 +127,7 @@ def test_resolve_parts_assembles_and_verifies(tmp_path: Path) -> None:
         (channel / "tiny.zip.part-00").unlink()
         assert resolve("tiny-1.0", index_url=str(index_path)) == local
     finally:
-        os.environ.pop("INTERSCRIPT_ML_CACHE", None)
+        os.environ.pop("SECRYST_CACHE", None)
 
 
 def test_resolve_parts_rejects_corrupt_part(tmp_path: Path) -> None:
@@ -162,9 +162,9 @@ def test_resolve_parts_rejects_corrupt_part(tmp_path: Path) -> None:
     }
     index_path = tmp_path / "models.yaml"
     index_path.write_text(yaml.safe_dump(index), encoding="utf-8")
-    os.environ["INTERSCRIPT_ML_CACHE"] = str(tmp_path / "cache")
+    os.environ["SECRYST_CACHE"] = str(tmp_path / "cache")
     try:
         with pytest.raises(RegistryError, match="part 0"):
             resolve("tiny-1.0", index_url=str(index_path))
     finally:
-        os.environ.pop("INTERSCRIPT_ML_CACHE", None)
+        os.environ.pop("SECRYST_CACHE", None)

scripts/publish_model.py

@@ -131,7 +131,7 @@ def release_notes(model_id: str, meta: dict, filename: str, size: int,
         "Runtimes reassemble split parts transparently and verify every sha256:",
         "",
         "```python",
-        "from interscript_ml import Model",
+        "from secryst import Model",
         f'model = Model.load("{model_id}")',
         "```",
     ]