diff --git a/src/mobius/integrations/onnx_genai/__init__.py b/src/mobius/integrations/onnx_genai/__init__.py
index a7f719cec..fc82c6e59 100644
--- a/src/mobius/integrations/onnx_genai/__init__.py
+++ b/src/mobius/integrations/onnx_genai/__init__.py
@@ -71,6 +71,7 @@
     PackageFacts,
     SpecialTokenFact,
     SpecialTokenRole,
+    TokenFacts,
     TokenizerArtifact,
     TokenizerFacts,
     build_tokenizer_facts,
@@ -119,6 +120,7 @@
     "SchedulerConfig",
     "SpecialTokenFact",
     "SpecialTokenRole",
+    "TokenFacts",
     "TokenizerArtifact",
     "TokenizerFacts",
     "add_policy_components_to_workflow",
diff --git a/src/mobius/integrations/onnx_genai/_schema/inference_metadata.schema.json b/src/mobius/integrations/onnx_genai/_schema/inference_metadata.schema.json
index 91861c8f5..48fbb616b 100644
--- a/src/mobius/integrations/onnx_genai/_schema/inference_metadata.schema.json
+++ b/src/mobius/integrations/onnx_genai/_schema/inference_metadata.schema.json
@@ -2234,7 +2234,7 @@
               "type": "null"
             }
           ],
-          "description": "Exact tokenizer, vocabulary, and special-token facts."
+          "description": "Exact tokenizer and vocabulary facts."
         }
       },
       "type": "object"
@@ -2621,7 +2621,12 @@
         },
         {
           "const": "eos_token_ids",
-          "description": "Token ids that end generation.\n\nA set, not a single id: a model may end a turn with one token and a\nmessage with another, and both must stop. Declaring it here is what lets\nthe runtime's stop policy read a package's EOS from the package instead\nof rediscovering it from tokenizer side-files it may not ship.",
+          "description": "Token ids that end generation.\n\nThe request value overrides\n`package.tokenizer.special_tokens.eos_token_id`. It never carries an\nauthored package default.",
+          "type": "string"
+        },
+        {
+          "const": "eos_token_lengths",
+          "description": "Number of valid entries in each row of a padded EOS-id tensor.",
           "type": "string"
         },
         {
@@ -2976,28 +2981,6 @@
       },
       "type": "object"
     },
-    "SpecialTokenFact": {
-      "additionalProperties": false,
-      "description": "One special token, pinned by id and exact surface bytes.",
-      "properties": {
-        "content": {
-          "description": "Exact UTF-8 surface form of the token.",
-          "minLength": 1,
-          "type": "string"
-        },
-        "id": {
-          "description": "Vocabulary id of the token.",
-          "format": "uint32",
-          "minimum": 0,
-          "type": "integer"
-        }
-      },
-      "required": [
-        "id",
-        "content"
-      ],
-      "type": "object"
-    },
     "SpeculationSafety": {
       "description": "Whether an effect may be executed inside a speculative region.",
       "oneOf": [
@@ -3993,6 +3976,94 @@
       ],
       "type": "object"
     },
+    "TokenFacts": {
+      "additionalProperties": false,
+      "description": "Numeric model and control-token facts.\n\nThese ids are model/package facts. Token spellings, added-token maps, and\nchat templates remain in tokenizer assets and are not repeated here.",
+      "properties": {
+        "audio_token_id": {
+          "description": "Prompt placeholder replaced by audio features.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "bos_token_id": {
+          "description": "Beginning-of-sequence token.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "decoder_start_token_id": {
+          "description": "First token fed to an encoder-decoder's autoregressive decoder.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "eos_token_id": {
+          "description": "Every token id that terminates package-default autoregressive generation.",
+          "items": {
+            "format": "uint32",
+            "minimum": 0,
+            "type": "integer"
+          },
+          "type": "array"
+        },
+        "image_token_id": {
+          "description": "Prompt placeholder replaced by image features.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "pad_token_id": {
+          "description": "Padding token used by package-authored tensor contracts.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "sep_token_id": {
+          "description": "Separator token used by sequence-pair models.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "video_token_id": {
+          "description": "Prompt placeholder replaced by video features.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        },
+        "vision_start_token_id": {
+          "description": "Token that opens a vision segment in a multimodal prompt.",
+          "format": "uint32",
+          "minimum": 0,
+          "type": [
+            "integer",
+            "null"
+          ]
+        }
+      },
+      "type": "object"
+    },
     "TokenizerArtifact": {
       "additionalProperties": false,
       "description": "One package-relative tokenizer artifact.",
@@ -4015,7 +4086,10 @@
         "algorithm": {
           "description": "Tokenizer algorithm identifier, e.g. `bpe`, `unigram`, `wordpiece`.",
           "minLength": 1,
-          "type": "string"
+          "type": [
+            "string",
+            "null"
+          ]
         },
         "artifacts": {
           "description": "Package-relative tokenizer artifacts.",
@@ -4031,23 +4105,26 @@
           "type": "boolean"
         },
         "special_tokens": {
-          "additionalProperties": {
-            "$ref": "#/$defs/SpecialTokenFact"
-          },
-          "description": "Special tokens by semantic role, e.g. `bos`, `eos`, `pad`.",
-          "type": "object"
+          "anyOf": [
+            {
+              "$ref": "#/$defs/TokenFacts"
+            },
+            {
+              "type": "null"
+            }
+          ],
+          "description": "Numeric model and control-token facts for this tokenizer vocabulary.\n\nToken strings, added-token mappings, and chat templates remain in the\ntokenizer assets. Request EOS inputs may override these defaults, but\nworkflow literals and termination components do not own another copy."
         },
         "vocab_size": {
           "description": "Number of entries in the vocabulary, including added tokens.",
           "format": "uint",
           "minimum": 1,
-          "type": "integer"
+          "type": [
+            "integer",
+            "null"
+          ]
         }
       },
-      "required": [
-        "algorithm",
-        "vocab_size"
-      ],
       "type": "object"
     },
     "VisionOutputBinding": {
diff --git a/src/mobius/integrations/onnx_genai/auto_export_test.py b/src/mobius/integrations/onnx_genai/auto_export_test.py
index 9a776a896..f54e5c3b0 100644
--- a/src/mobius/integrations/onnx_genai/auto_export_test.py
+++ b/src/mobius/integrations/onnx_genai/auto_export_test.py
@@ -244,11 +244,11 @@ def test_dispatch_decoder(tmp_path):
     }
     assert application_inputs == {
         "request.prompt_lengths",
-        "request.eos_ids",
-        "request.eos_lengths",
         "request.row_max_iterations",
         "request.rng_counter",
     }
+    assert workflow["inputs"]["request.eos_ids"]["role"]["role"] == "eos_token_ids"
+    assert workflow["inputs"]["request.eos_lengths"]["role"]["role"] == "eos_token_lengths"
     assert workflow["inputs"]["request.prompt_lengths"]["default"] == -1
     assert [node["component"] for node in workflow["steps"][0]["setup"]] == [
         "decoder_state_initializer",
diff --git a/src/mobius/integrations/onnx_genai/package_facts.py b/src/mobius/integrations/onnx_genai/package_facts.py
index 74e64542a..53ec00097 100644
--- a/src/mobius/integrations/onnx_genai/package_facts.py
+++ b/src/mobius/integrations/onnx_genai/package_facts.py
@@ -6,28 +6,28 @@
 A published package carries neural graphs and the workflow that drives them, but
 a request arrives as text and media.  Turning that request into the token stream
 the graphs consume requires the vocabulary contract the package was built
-against: which algorithm produced the ids, how many ids exist, and which id
-plays each semantic role.  onnx-genai calls that section ``package.tokenizer``
+against: which algorithm produced the ids, how many ids exist, and which numeric ids
+play each execution-relevant role.  onnx-genai calls that section ``package.tokenizer``
 (``PackageFacts``/``TokenizerFacts`` in ``onnx-genai-metadata``), and it is the
 only place a front end looks for those facts.
 
-Two of those roles are load-bearing rather than informational:
+Two of those facts are load-bearing rather than informational:
 
-``eos``
+``eos_token_id``
     The workflow's termination policy already compares generated ids against a
     stop id.  Publishing the same id under a role means a caller can render and
     trim a transcript without re-deriving it from a side file.
 
-``image_placeholder``
+``image_token_id``
     The prompt token whose position an image's features replace.  A multimodal
     package that omits it declares no place in the token stream for its own
     image features, so an attached image is preprocessed and then dropped.
 
-Every value here is read from the package's own artifacts.  The algorithm,
-vocabulary size and byte-level flag come from the packaged tokenizer definition;
-each special token pairs a config-declared id with the surface form that id has
-in that same vocabulary.  A role the package does not declare is simply absent,
-because a guessed special token is worse than a missing one.
+Every value here is read from the package's own artifacts or resolved config.
+The algorithm, vocabulary size and byte-level flag come from the packaged
+tokenizer definition; numeric token IDs come from the package's runtime config.
+Token spellings and chat templates remain authoritative in tokenizer assets,
+so execution metadata never carries a second, potentially stale vocabulary.
 
 Nothing in this module knows a model name or a literal token id: architecture
 defaults belong to the config adapters under :mod:`mobius._configs`, and reach
@@ -47,9 +47,8 @@
 
 _LOGGER = logging.getLogger(__name__)
 
-#: Semantic role naming the prompt token that stands for one whole image.
-#: Fixed by the onnx-genai runtime, which looks this role up by name.
-IMAGE_PLACEHOLDER_ROLE: Final = "image_placeholder"
+#: Schema field naming the prompt token that stands for one whole image.
+IMAGE_PLACEHOLDER_ROLE: Final = "image_token_id"
 
 
 @dataclasses.dataclass(frozen=True)
@@ -63,21 +62,24 @@ class SpecialTokenRole:
 
     name: str
     fields: tuple[str, ...]
+    multiple: bool = False
 
 
 #: Roles every text-producing package can state about its own vocabulary.
 TEXT_TOKEN_ROLES: Final[tuple[SpecialTokenRole, ...]] = (
-    SpecialTokenRole("bos", ("bos_token_id",)),
-    SpecialTokenRole("eos", ("eos_token_id",)),
-    SpecialTokenRole("pad", ("pad_token_id",)),
-    SpecialTokenRole("unk", ("unk_token_id",)),
+    SpecialTokenRole("pad_token_id", ("pad_token_id",)),
+    SpecialTokenRole("bos_token_id", ("bos_token_id",)),
+    SpecialTokenRole("eos_token_id", ("eos_token_id",), multiple=True),
+    SpecialTokenRole("sep_token_id", ("sep_token_id",)),
+    SpecialTokenRole("decoder_start_token_id", ("decoder_start_token_id",)),
 )
 
 #: Roles a package adds when a modality's features replace a prompt token.
 MEDIA_TOKEN_ROLES: Final[tuple[SpecialTokenRole, ...]] = (
     SpecialTokenRole(IMAGE_PLACEHOLDER_ROLE, ("image_token_id",)),
-    SpecialTokenRole("audio_placeholder", ("audio_token_id",)),
-    SpecialTokenRole("video_placeholder", ("video_token_id",)),
+    SpecialTokenRole("video_token_id", ("video_token_id",)),
+    SpecialTokenRole("audio_token_id", ("audio_token_id",)),
+    SpecialTokenRole("vision_start_token_id", ("vision_start_token_id",)),
 )
 
 #: Nested config objects searched after the root for a role's id.  A composite
@@ -116,7 +118,11 @@ class SpecialTokenRole:
 
 @dataclasses.dataclass(frozen=True)
 class SpecialTokenFact:
-    """One special token, pinned by id and exact surface bytes."""
+    """Compatibility value for callers of the former text-bearing API.
+
+    New metadata uses :class:`TokenFacts`; token content remains in tokenizer
+    assets. Keeping this value exported avoids breaking existing imports.
+    """
 
     id: int
     content: str
@@ -125,6 +131,19 @@ def to_metadata(self) -> dict[str, Any]:
         return {"id": self.id, "content": self.content}
 
 
+@dataclasses.dataclass(frozen=True)
+class TokenFacts:
+    """Numeric model and control-token facts owned by the package."""
+
+    values: Mapping[str, int | tuple[int, ...]] = dataclasses.field(default_factory=dict)
+
+    def to_metadata(self) -> dict[str, Any]:
+        return {
+            name: list(value) if isinstance(value, tuple) else value
+            for name, value in self.values.items()
+        }
+
+
 @dataclasses.dataclass(frozen=True)
 class TokenizerArtifact:
     """One package-relative tokenizer artifact."""
@@ -139,30 +158,27 @@ def to_metadata(self) -> dict[str, Any]:
 class TokenizerFacts:
     """Tokenizer facts and package-relative artifacts.
 
-    Mirrors onnx-genai's ``TokenizerFacts``.  ``algorithm`` and ``vocab_size``
-    are required by that contract; the rest are omitted when empty rather than
-    published as defaults.
+    Mirrors onnx-genai's ``TokenizerFacts``. ``algorithm`` and ``vocab_size``
+    are omitted when no tokenizer definition is available, while numeric token
+    facts can still describe the package's execution contract.
     """
 
-    algorithm: str
-    vocab_size: int
+    algorithm: str | None = None
+    vocab_size: int | None = None
     byte_level: bool = False
     artifacts: tuple[TokenizerArtifact, ...] = ()
-    special_tokens: Mapping[str, SpecialTokenFact] = dataclasses.field(default_factory=dict)
+    special_tokens: TokenFacts = dataclasses.field(default_factory=TokenFacts)
 
     def to_metadata(self) -> dict[str, Any]:
-        facts: dict[str, Any] = {
-            "algorithm": self.algorithm,
-            "vocab_size": self.vocab_size,
-            "byte_level": self.byte_level,
-        }
+        facts: dict[str, Any] = {"byte_level": self.byte_level}
+        if self.algorithm is not None:
+            facts["algorithm"] = self.algorithm
+        if self.vocab_size is not None:
+            facts["vocab_size"] = self.vocab_size
         if self.artifacts:
             facts["artifacts"] = [artifact.to_metadata() for artifact in self.artifacts]
-        if self.special_tokens:
-            facts["special_tokens"] = {
-                role: token.to_metadata()
-                for role, token in sorted(self.special_tokens.items())
-            }
+        if self.special_tokens.values:
+            facts["special_tokens"] = self.special_tokens.to_metadata()
         return facts
 
 
@@ -433,14 +449,23 @@ def _config_value(config: Any, field: str) -> Any:
 
 
 def _token_id(value: Any) -> int | None:
-    """Coerce a declared token id, taking the first of a stop-id list."""
-    if isinstance(value, (list, tuple)):
-        value = value[0] if value else None
+    """Coerce one declared token id."""
     if isinstance(value, bool) or not isinstance(value, int):
         return None
     return value if value >= 0 else None
 
 
+def _token_ids(value: Any) -> tuple[int, ...]:
+    """Coerce an ordered token-id set without silently dropping multi-EOS."""
+    values = value if isinstance(value, (list, tuple)) else (value,)
+    result: list[int] = []
+    for candidate in values:
+        token_id = _token_id(candidate)
+        if token_id is not None and token_id not in result:
+            result.append(token_id)
+    return tuple(result)
+
+
 _MISSING: Final = object()
 
 
@@ -491,41 +516,39 @@ def source_declared_roles(
     return declared
 
 
-def build_special_tokens(
-    definition: TokenizerDefinition,
+def build_token_facts(
     config: Any,
     roles: Iterable[SpecialTokenRole],
     *,
     declared: Mapping[str, Any] | None = None,
-) -> dict[str, SpecialTokenFact]:
-    """Resolve declared roles against the packaged vocabulary.
+) -> TokenFacts:
+    """Resolve numeric package token facts.
 
     ``declared`` overrides the config for roles the emitting workflow has
     already resolved for itself, so a document never states one id for its
     termination policy and a different one for the same role here.
 
-    A role survives only when both halves of the fact are known: an id the
-    package declares *and* the exact surface bytes that id has in this
-    vocabulary.  An id with no surface form is a stale config field pointing
-    outside the shipped vocabulary, and publishing it would let a front end
-    splice a token the tokenizer cannot render.
+    Multi-valued roles preserve their declared order. Scalar roles take exactly
+    one non-negative integer. Token spellings remain in tokenizer assets.
     """
     declared = declared or {}
-    special_tokens: dict[str, SpecialTokenFact] = {}
+    values: dict[str, int | tuple[int, ...]] = {}
     for role in roles:
-        token_id = _token_id(declared.get(role.name))
-        if token_id is None:
+        value = declared.get(role.name)
+        if value is None:
             for field in role.fields:
-                token_id = _token_id(_config_value(config, field))
-                if token_id is not None:
+                value = _config_value(config, field)
+                if value is not None:
                     break
-        if token_id is None:
+        if role.multiple:
+            token_ids = _token_ids(value)
+            if token_ids:
+                values[role.name] = token_ids
             continue
-        content = definition.surface_forms.get(token_id)
-        if not content:
-            continue
-        special_tokens[role.name] = SpecialTokenFact(id=token_id, content=content)
-    return special_tokens
+        token_id = _token_id(value)
+        if token_id is not None:
+            values[role.name] = token_id
+    return TokenFacts(values)
 
 
 def build_tokenizer_facts(
@@ -560,8 +583,11 @@ def build_tokenizer_facts(
     address nothing, so the vocabulary is the width a caller can actually render.
     """
     definition = read_tokenizer_definition(package_dir) or read_tokenizer_definition(source)
-    if definition is None or definition.vocab_size <= 0:
-        return None
+    special_tokens = build_token_facts(
+        config,
+        roles,
+        declared=source_declared_roles(source, roles),
+    )
     artifacts = ()
     if package_dir is not None:
         artifacts = tuple(
@@ -569,17 +595,14 @@ def build_tokenizer_facts(
             for name in TOKENIZER_ARTIFACT_NAMES
             if os.path.isfile(os.path.join(package_dir, name))
         )
+    if definition is None and not artifacts and not special_tokens.values:
+        return None
     return TokenizerFacts(
-        algorithm=definition.algorithm,
-        vocab_size=definition.vocab_size,
-        byte_level=definition.byte_level,
+        algorithm=definition.algorithm if definition is not None else None,
+        vocab_size=definition.vocab_size if definition is not None else None,
+        byte_level=definition.byte_level if definition is not None else False,
         artifacts=artifacts,
-        special_tokens=build_special_tokens(
-            definition,
-            config,
-            roles,
-            declared=source_declared_roles(source, roles),
-        ),
+        special_tokens=special_tokens,
     )
 
 
@@ -602,4 +625,5 @@ def attach_package_facts(
     tokenizer = build_tokenizer_facts(source, config, roles=roles, package_dir=package_dir)
     if tokenizer is None:
         return
+    metadata["schema_version"] = "v1.2"
     metadata.setdefault("package", {})["tokenizer"] = tokenizer.to_metadata()
diff --git a/src/mobius/integrations/onnx_genai/package_facts_test.py b/src/mobius/integrations/onnx_genai/package_facts_test.py
index d327b7df9..876879806 100644
--- a/src/mobius/integrations/onnx_genai/package_facts_test.py
+++ b/src/mobius/integrations/onnx_genai/package_facts_test.py
@@ -23,6 +23,7 @@
 
 from mobius._configs import MMSConfig
 from mobius._model_package import ModelPackage
+from mobius.integrations.onnx_genai import SpecialTokenFact
 from mobius.integrations.onnx_genai.auto_export import write_onnx_genai_config
 from mobius.integrations.onnx_genai.auto_export_test import _vlm_package, _VlmCfg
 from mobius.integrations.onnx_genai.inference_metadata_test import (
@@ -81,6 +82,7 @@ class _MediaCfg(_VlmCfg):
     """A vision config that can also declare a video placeholder."""
 
     video_token_id: int | None = None
+    audio_token_id: int | None = None
 
 
 def _write_tokenizer(
@@ -165,6 +167,12 @@ def test_an_unreadable_tokenizer_states_nothing(self, tmp_path):
 class TestDecoderPackageFacts:
     """A text package states the vocabulary the ids it emits belong to."""
 
+    def test_legacy_special_token_fact_remains_importable(self):
+        assert SpecialTokenFact(7, "<legacy>").to_metadata() == {
+            "id": 7,
+            "content": "<legacy>",
+        }
+
     def test_facts_survive_into_package_metadata(self, tmp_path):
         _write_tokenizer(tmp_path)
         metadata = build_decoder_workflow_metadata(
@@ -175,10 +183,11 @@ def test_facts_survive_into_package_metadata(self, tmp_path):
         assert tokenizer["vocab_size"] == _TEXT_VOCAB_SIZE
         assert tokenizer["byte_level"] is True
         assert tokenizer["special_tokens"] == {
-            "bos": {"id": _BOS_ID, "content": "<s>"},
-            "eos": {"id": _EOS_ID, "content": "</s>"},
-            "pad": {"id": _PAD_ID, "content": "<pad>"},
+            "pad_token_id": _PAD_ID,
+            "bos_token_id": _BOS_ID,
+            "eos_token_id": [_EOS_ID],
         }
+        assert metadata["schema_version"] == "v1.2"
         _validate(metadata)
 
     def test_the_stated_width_is_the_vocabulary_not_the_logits_row(self, tmp_path):
@@ -205,8 +214,8 @@ def test_pipeline_workflow_remains_the_sole_representation(self, tmp_path):
     def test_the_stated_stop_token_is_the_one_the_loop_terminates_on(self, tmp_path):
         # A repackaged checkpoint may be retuned without its config being
         # rewritten, so genai_config.json outranks the config. Both the
-        # termination policy and the published role must follow it, or one
-        # document names two different stop tokens.
+        # package facts must follow it, while the workflow only exposes the
+        # request override consumed by termination.
         _write_tokenizer(tmp_path, added={_IMAGE_ID: "<|endoftext|>"})
         (tmp_path / "genai_config.json").write_text(
             json.dumps({"model": {"eos_token_id": _IMAGE_ID}}), encoding="utf-8"
@@ -215,29 +224,49 @@ def test_the_stated_stop_token_is_the_one_the_loop_terminates_on(self, tmp_path)
             _decoder_package(_TextCfg()), _TextCfg(), source=str(tmp_path)
         )
         workflow = metadata["pipeline"]["workflow"]
-        assert metadata["package"]["tokenizer"]["special_tokens"]["eos"] == {
-            "id": _IMAGE_ID,
-            "content": "<|endoftext|>",
-        }
-        assert workflow["inputs"]["package.eos_ids"]["default"] == _IMAGE_ID
+        assert metadata["package"]["tokenizer"]["special_tokens"]["eos_token_id"] == [
+            _IMAGE_ID
+        ]
+        assert "package.eos_ids" not in workflow["inputs"]
+        assert workflow["inputs"]["request.eos_ids"]["role"]["role"] == "eos_token_ids"
+        assert "default" not in workflow["inputs"]["request.eos_ids"]
 
-    def test_a_role_the_vocabulary_cannot_render_is_omitted(self, tmp_path):
-        # A config field pointing outside the shipped vocabulary is stale. A
-        # front end that spliced that id would emit a token the tokenizer
-        # cannot render, so the role is dropped rather than half-stated.
+    def test_multi_eos_order_is_preserved_without_a_workflow_literal(self, tmp_path):
+        _write_tokenizer(tmp_path, added={_IMAGE_ID: "<|im_end|>"})
+        (tmp_path / "genai_config.json").write_text(
+            json.dumps({"model": {"eos_token_id": [_EOS_ID, _IMAGE_ID]}}),
+            encoding="utf-8",
+        )
+
+        metadata = build_decoder_workflow_metadata(
+            _decoder_package(_TextCfg()), _TextCfg(), source=str(tmp_path)
+        )
+
+        assert metadata["package"]["tokenizer"]["special_tokens"]["eos_token_id"] == [
+            _EOS_ID,
+            _IMAGE_ID,
+        ]
+        assert "package.eos_ids" not in metadata["pipeline"]["workflow"]["inputs"]
+
+    def test_numeric_facts_do_not_duplicate_tokenizer_surface_forms(self, tmp_path):
+        # Token spellings belong to tokenizer assets. The package still states
+        # its execution-relevant numeric fact without copying vocabulary text.
         _write_tokenizer(tmp_path)
         metadata = build_decoder_workflow_metadata(
             _decoder_package(_TextCfg()),
             _TextCfg(bos_token_id=9_999),
             source=str(tmp_path),
         )
-        assert "bos" not in metadata["package"]["tokenizer"]["special_tokens"]
+        assert metadata["package"]["tokenizer"]["special_tokens"]["bos_token_id"] == 9_999
 
-    def test_a_package_without_a_readable_tokenizer_states_no_facts(self, tmp_path):
+    def test_a_package_without_a_readable_tokenizer_still_states_numeric_facts(self, tmp_path):
         metadata = build_decoder_workflow_metadata(
             _decoder_package(_TextCfg()), _TextCfg(), source=str(tmp_path)
         )
-        assert "package" not in metadata
+        tokenizer = metadata["package"]["tokenizer"]
+        assert "algorithm" not in tokenizer
+        assert "vocab_size" not in tokenizer
+        assert tokenizer["special_tokens"]["eos_token_id"] == [_EOS_ID]
 
     def test_artifacts_name_only_files_the_package_contains(self, tmp_path):
         source = tmp_path / "source"
@@ -311,13 +340,10 @@ def _materialize_tokenizer(output_dir, source, *, revision=None):
         metadata = yaml.safe_load(Path(artifacts["inference_metadata"]).read_text())
         tokenizer = metadata["package"]["tokenizer"]
         assert tokenizer["vocab_size"] == _IMAGE_ID + 1
-        assert tokenizer["special_tokens"]["eos"] == {
-            "id": _IMAGE_ID,
-            "content": "<|endoftext|>",
-        }
+        assert tokenizer["special_tokens"]["eos_token_id"] == [_IMAGE_ID]
         assert tokenizer["artifacts"] == [{"location": "tokenizer.json"}]
         workflow = metadata["pipeline"]["workflow"]
-        assert workflow["inputs"]["package.eos_ids"]["default"] == _IMAGE_ID
+        assert "package.eos_ids" not in workflow["inputs"]
 
 
 class TestMultimodalPackageFacts:
@@ -338,11 +364,8 @@ def _metadata(self, source, **overrides):
 
     def test_image_placeholder_role_is_published(self, source):
         tokenizer = self._metadata(source)["package"]["tokenizer"]
-        assert tokenizer["special_tokens"][IMAGE_PLACEHOLDER_ROLE] == {
-            "id": _IMAGE_ID,
-            "content": "<image>",
-        }
-        assert tokenizer["special_tokens"]["eos"] == {"id": _EOS_ID, "content": "</s>"}
+        assert tokenizer["special_tokens"][IMAGE_PLACEHOLDER_ROLE] == _IMAGE_ID
+        assert tokenizer["special_tokens"]["eos_token_id"] == [_EOS_ID]
 
     def test_image_input_routing_survives_alongside_the_facts(self, source):
         metadata = self._metadata(source)
@@ -374,15 +397,15 @@ def test_image_input_routing_survives_alongside_the_facts(self, source):
 
     def test_a_declared_video_placeholder_is_published_too(self, source):
         tokenizer = self._metadata(source, video_token_id=_VIDEO_ID)["package"]["tokenizer"]
-        assert tokenizer["special_tokens"]["video_placeholder"] == {
-            "id": _VIDEO_ID,
-            "content": "<video>",
-        }
+        assert tokenizer["special_tokens"]["video_token_id"] == _VIDEO_ID
+
+    def test_a_declared_audio_placeholder_is_published_too(self, source):
+        tokenizer = self._metadata(source, audio_token_id=_VIDEO_ID)["package"]["tokenizer"]
+        assert tokenizer["special_tokens"]["audio_token_id"] == _VIDEO_ID
 
     def test_an_undeclared_modality_role_is_absent(self, source):
         tokenizer = self._metadata(source)["package"]["tokenizer"]
-        assert "video_placeholder" not in tokenizer["special_tokens"]
-        assert "audio_placeholder" not in tokenizer["special_tokens"]
+        assert "video_token_id" not in tokenizer["special_tokens"]
 
 
 class TestScatteredAddedTokens:
@@ -446,10 +469,10 @@ def test_the_multimodal_package_publishes_the_scattered_placeholder(self, tmp_pa
         )
         config = _MediaCfg(image_token_id=_IMAGE_ID, eos_token_id=_EOS_ID)
         metadata = build_vlm_workflow_metadata(_vlm_package(), config, source=str(tmp_path))
-        assert metadata["package"]["tokenizer"]["special_tokens"][IMAGE_PLACEHOLDER_ROLE] == {
-            "id": _IMAGE_ID,
-            "content": "<|image_pad|>",
-        }
+        assert (
+            metadata["package"]["tokenizer"]["special_tokens"][IMAGE_PLACEHOLDER_ROLE]
+            == _IMAGE_ID
+        )
 
 
 class TestAlgorithmIsReadNotDefaulted:
@@ -625,7 +648,8 @@ def test_blank_id_and_the_published_pad_role_name_one_token(self, tmp_path):
         assert tokenizer["algorithm"] == "word_level"
         assert tokenizer["vocab_size"] == config.vocab_size
         blank_id = metadata["profiles"]["transcription"]["decoding"]["blank_id"]
-        assert tokenizer["special_tokens"]["pad"] == {"id": blank_id, "content": "<pad>"}
+        assert tokenizer["special_tokens"]["pad_token_id"] == blank_id
+        assert set(tokenizer["special_tokens"]) == {"pad_token_id"}
         _validate(metadata)
 
     def test_a_package_that_ships_no_vocabulary_still_states_its_facts(self, tmp_path):
@@ -677,7 +701,7 @@ def test_facts_are_stated_once_for_both_components(self, tmp_path):
         )
         tokenizer = metadata["package"]["tokenizer"]
         assert tokenizer["vocab_size"] == _TEXT_VOCAB_SIZE
-        assert tokenizer["special_tokens"]["eos"] == {"id": _EOS_ID, "content": "</s>"}
+        assert tokenizer["special_tokens"]["eos_token_id"] == [_EOS_ID]
         # Nothing was shipped next to the document, so it claims no artifacts.
         assert "artifacts" not in tokenizer
         assert set(metadata["pipeline"]) == {"workflow"}
diff --git a/src/mobius/integrations/onnx_genai/workflow_metadata.py b/src/mobius/integrations/onnx_genai/workflow_metadata.py
index db6cd3b37..ef3c370ed 100644
--- a/src/mobius/integrations/onnx_genai/workflow_metadata.py
+++ b/src/mobius/integrations/onnx_genai/workflow_metadata.py
@@ -134,6 +134,9 @@ def ignore_aliases(self, data: Any) -> bool:
         return True
 
 
+_CTC_TOKEN_ROLES = tuple(role for role in TEXT_TOKEN_ROLES if role.name == "pad_token_id")
+
+
 def _dump_yaml(metadata: dict[str, Any], handle: Any) -> None:
     yaml.dump(metadata, handle, Dumper=_NoAliasSafeDumper, sort_keys=False)
 
@@ -143,14 +146,6 @@ def _source_model_value(source: str | None, name: str, fallback: Any) -> Any:
     return source_declared_value(source, name, fallback)
 
 
-def _source_token_id(source: str | None, name: str, fallback: Any) -> int:
-    """Resolve a generation token ID from packaged runtime metadata when available."""
-    fallback = _source_model_value(source, name, fallback)
-    if isinstance(fallback, list):
-        fallback = fallback[0] if fallback else 0
-    return int(fallback or 0)
-
-
 def _contract(value: ir.Value) -> dict[str, Any]:
     port = _port(value)
     dtype = {"fp16": "float16", "bf16": "bfloat16", "fp32": "float32"}.get(
@@ -278,7 +273,7 @@ def _declare_batch_capacities(
             capacity["uniform_dimensions"] = uniform_dimensions
         component["batch_capacity"] = capacity
         declared = True
-    if declared:
+    if declared and metadata.get("schema_version") == "v1":
         metadata["schema_version"] = "v1.1"
 
 
@@ -6437,7 +6432,6 @@ def build_vlm_workflow_metadata(
     batch_int = _request_aligned({"dtype": "int64", "rank": 1, "shape": [batch]})
     batch_bool = _request_aligned({"dtype": "bool", "rank": 1, "shape": [batch]})
     control_int = {"dtype": "int64", "rank": 1, "shape": [1]}
-    eos = _source_token_id(source, "eos_token_id", getattr(config, "eos_token_id", 0))
     inputs: dict[str, Any] = {
         "request.prompt_tokens": {
             "contract": _contract(token_input),
@@ -6472,18 +6466,14 @@ def build_vlm_workflow_metadata(
             "default": -1,
         },
         "request.eos_ids": {
-            "contract": {"dtype": "int64", "rank": 2, "shape": [batch, "num_eos"]},
-            "role": {"kind": "opaque"},
-            "source": {"kind": "application", "name": "eos_ids"},
-            "required": False,
-            "default": eos,
-        },
-        "request.eos_lengths": {
-            "contract": batch_int,
-            "role": {"kind": "opaque"},
-            "source": {"kind": "application", "name": "eos_lengths"},
-            "required": False,
-            "default": 1,
+            "contract": (
+                {"dtype": "int64", "rank": 2, "shape": [batch, "num_eos"]}
+                if cache_pairs
+                else {"dtype": "int64", "rank": 1, "shape": ["num_eos"]}
+            ),
+            "role": {"kind": "runtime", "version": "1.0", "role": "eos_token_ids"},
+            "source": {"kind": "request"},
+            "required": True,
         },
         "request.row_max_iterations": {
             "contract": batch_int,
@@ -6492,13 +6482,22 @@ def build_vlm_workflow_metadata(
             "required": False,
             "default": -1,
         },
-        "package.eos_ids": {
-            "contract": {"dtype": "int64", "rank": 1, "shape": [1]},
-            "role": {"kind": "opaque"},
-            "source": {"kind": "literal"},
-            "required": False,
-            "default": eos,
-        },
+        **(
+            {
+                "request.eos_lengths": {
+                    "contract": batch_int,
+                    "role": {
+                        "kind": "runtime",
+                        "version": "1.0",
+                        "role": "eos_token_lengths",
+                    },
+                    "source": {"kind": "request"},
+                    "required": True,
+                },
+            }
+            if cache_pairs
+            else {}
+        ),
         "package.max_context": {
             "contract": control_int,
             "role": {"kind": "opaque"},
@@ -7315,7 +7314,7 @@ def build_vlm_workflow_metadata(
         },
     }
     metadata = {
-        "schema_version": "v1",
+        "schema_version": "v1.2",
         "preprocessing": preprocessing,
         "pipeline": {"workflow": _publish_workflow_v1(workflow)},
     }
@@ -8533,10 +8532,6 @@ def _build_autoregressive_workflow_metadata(
     batch_int = _request_aligned({"dtype": "int64", "rank": 1, "shape": [batch_dimension]})
     batch_bool = _request_aligned({"dtype": "bool", "rank": 1, "shape": [batch_dimension]})
     control_int = {"dtype": "int64", "rank": 1, "shape": [1]}
-    # The published `special_tokens.eos` fact is resolved the same way, so a
-    # repackaged checkpoint cannot state one stop id for its termination policy
-    # and a different one for the same role.
-    eos_token_id = _source_token_id(source, "eos_token_id", getattr(config, "eos_token_id", 0))
     workflow_inputs.update(
         {
             "request.max_iterations": {
@@ -8549,13 +8544,6 @@ def _build_autoregressive_workflow_metadata(
                 "source": {"kind": "request", "field": "max_output_tokens"},
                 "required": True,
             },
-            "package.eos_ids": {
-                "contract": {"dtype": "int64", "rank": 1, "shape": ["E"]},
-                "role": {"kind": "opaque"},
-                "source": {"kind": "literal"},
-                "required": False,
-                "default": eos_token_id,
-            },
             "package.one_token": {
                 "contract": batch_int,
                 "role": {"kind": "opaque"},
@@ -8608,17 +8596,23 @@ def _build_autoregressive_workflow_metadata(
                         "rank": 2,
                         "shape": [batch_dimension, "num_eos"],
                     },
-                    "role": {"kind": "opaque"},
-                    "source": {"kind": "application", "name": "eos_ids"},
-                    "required": False,
-                    "default": eos_token_id,
+                    "role": {
+                        "kind": "runtime",
+                        "version": "1.0",
+                        "role": "eos_token_ids",
+                    },
+                    "source": {"kind": "request"},
+                    "required": True,
                 },
                 "request.eos_lengths": {
                     "contract": batch_int,
-                    "role": {"kind": "opaque"},
-                    "source": {"kind": "application", "name": "eos_lengths"},
-                    "required": False,
-                    "default": 1,
+                    "role": {
+                        "kind": "runtime",
+                        "version": "1.0",
+                        "role": "eos_token_lengths",
+                    },
+                    "source": {"kind": "request"},
+                    "required": True,
                 },
                 "request.row_max_iterations": {
                     "contract": batch_int,
@@ -8632,6 +8626,17 @@ def _build_autoregressive_workflow_metadata(
                 },
             }
         )
+    else:
+        workflow_inputs["request.eos_ids"] = {
+            "contract": {"dtype": "int64", "rank": 1, "shape": ["num_eos"]},
+            "role": {
+                "kind": "runtime",
+                "version": "1.0",
+                "role": "eos_token_ids",
+            },
+            "source": {"kind": "request"},
+            "required": True,
+        }
     stochastic_sampler = sampler != "greedy"
     sampler_with_rng = stochastic_sampler or bool(cache_pairs)
     if sampler_with_rng:
@@ -9377,7 +9382,7 @@ def _build_autoregressive_workflow_metadata(
                 "termination",
                 {
                     "tokens": "sample.body",
-                    "eos_ids": ("termination.eos_ids" if cache_pairs else "package.eos_ids"),
+                    "eos_ids": ("termination.eos_ids" if cache_pairs else "request.eos_ids"),
                     **({"eos_lengths": "termination.eos_lengths"} if cache_pairs else {}),
                     "iteration": "loop.iteration",
                     "max_iterations": (
@@ -9551,7 +9556,7 @@ def _build_autoregressive_workflow_metadata(
         },
     }
     metadata = {
-        "schema_version": "1.0",
+        "schema_version": "v1.2",
         **({"preprocessing": {"audio": audio_program}} if audio_program is not None else {}),
         "pipeline": {"workflow": _publish_workflow_v1(workflow)},
     }
@@ -10178,10 +10183,10 @@ def build_ctc_asr_workflow_metadata(
         },
         "pipeline": {"workflow": _publish_workflow_v1(workflow)},
     }
-    # The transcript is rendered from class ids, so the class table is the
-    # package's tokenizer. Publishing it under the same roles a decoder uses
-    # means ``blank_id`` above and the ``pad`` role name one id, not two.
-    attach_package_facts(metadata, source, config)
+    # CTC uses the padding class as its blank id. BOS/EOS are unrelated
+    # HuggingFace config defaults here and must not imply autoregressive
+    # generation semantics for this logits-only workflow.
+    attach_package_facts(metadata, source, config, roles=_CTC_TOKEN_ROLES)
     return metadata
 
 
@@ -10195,7 +10200,13 @@ def write_ctc_asr_workflow_metadata(
     """Write one-file CTC ASR metadata into *output_dir*."""
     os.makedirs(output_dir, exist_ok=True)
     metadata = build_ctc_asr_workflow_metadata(pkg, config, source=source)
-    attach_package_facts(metadata, source, config, package_dir=output_dir)
+    attach_package_facts(
+        metadata,
+        source,
+        config,
+        roles=_CTC_TOKEN_ROLES,
+        package_dir=output_dir,
+    )
     path = os.path.join(output_dir, "inference_metadata.yaml")
     with open(path, "w", encoding="utf-8") as handle:
         _dump_yaml(metadata, handle)
diff --git a/src/mobius/integrations/onnx_genai/workflow_metadata_test.py b/src/mobius/integrations/onnx_genai/workflow_metadata_test.py
index 5828ed7c6..08fe71233 100644
--- a/src/mobius/integrations/onnx_genai/workflow_metadata_test.py
+++ b/src/mobius/integrations/onnx_genai/workflow_metadata_test.py
@@ -255,7 +255,8 @@ def test_vlm_writer_derives_real_decoder_contract_from_artifact(tmp_path):
     assert "&id" not in serialized
     assert "*id" not in serialized
     with open(path, encoding="utf-8") as handle:
-        workflow = yaml.safe_load(handle)["pipeline"]["workflow"]
+        metadata = yaml.safe_load(handle)
+    workflow = metadata["pipeline"]["workflow"]
     decoder_invokes = []
     policy_invokes = {}
 
@@ -285,7 +286,8 @@ def collect_decoder_invokes(node):
         == 104
     )
     assert workflow["inputs"]["package.max_context"]["default"] == 131072
-    assert workflow["inputs"]["package.eos_ids"]["default"] == 200001
+    assert "package.eos_ids" not in workflow["inputs"]
+    assert metadata["package"]["tokenizer"]["special_tokens"]["eos_token_id"] == [200001]
     assert workflow["state"]["cache_103"]["contract"] == {
         "dtype": "bfloat16",
         "rank": 4,
diff --git a/tests/fixtures/onnx_genai_workflows/decoder/inference_metadata.yaml b/tests/fixtures/onnx_genai_workflows/decoder/inference_metadata.yaml
index f17cd0911..fa2c9bd07 100644
--- a/tests/fixtures/onnx_genai_workflows/decoder/inference_metadata.yaml
+++ b/tests/fixtures/onnx_genai_workflows/decoder/inference_metadata.yaml
@@ -1,4 +1,4 @@
-schema_version: v1.1
+schema_version: v1.2
 pipeline:
   workflow:
     manifest:
@@ -42,18 +42,6 @@ pipeline:
         source:
           kind: request
         required: true
-      package.eos_ids:
-        contract:
-          dtype: int64
-          rank: 1
-          shape:
-          - E
-        role:
-          kind: opaque
-        source:
-          kind: literal
-        required: false
-        default: 127
       package.one_token:
         contract:
           dtype: int64
@@ -120,12 +108,12 @@ pipeline:
             kind: request_aligned
             axis: 0
         role:
-          kind: opaque
+          kind: runtime
+          version: '1.0'
+          role: eos_token_ids
         source:
-          kind: application
-          name: eos_ids
-        required: false
-        default: 127
+          kind: request
+        required: true
       request.eos_lengths:
         contract:
           dtype: int64
@@ -136,12 +124,12 @@ pipeline:
             kind: request_aligned
             axis: 0
         role:
-          kind: opaque
+          kind: runtime
+          version: '1.0'
+          role: eos_token_lengths
         source:
-          kind: application
-          name: eos_lengths
-        required: false
-        default: 1
+          kind: request
+        required: true
       request.row_max_iterations:
         contract:
           dtype: int64
@@ -1397,3 +1385,9 @@ pipeline:
           rank: 1
           shape:
           - 1
+package:
+  tokenizer:
+    byte_level: false
+    special_tokens:
+      eos_token_id:
+      - 127
diff --git a/tests/fixtures/onnx_genai_workflows/static_cache/inference_metadata.yaml b/tests/fixtures/onnx_genai_workflows/static_cache/inference_metadata.yaml
index 3d101cf37..2f4bd5598 100644
--- a/tests/fixtures/onnx_genai_workflows/static_cache/inference_metadata.yaml
+++ b/tests/fixtures/onnx_genai_workflows/static_cache/inference_metadata.yaml
@@ -1,4 +1,4 @@
-schema_version: v1.1
+schema_version: v1.2
 pipeline:
   workflow:
     manifest:
@@ -42,18 +42,6 @@ pipeline:
         source:
           kind: request
         required: true
-      package.eos_ids:
-        contract:
-          dtype: int64
-          rank: 1
-          shape:
-          - E
-        role:
-          kind: opaque
-        source:
-          kind: literal
-        required: false
-        default: 127
       package.one_token:
         contract:
           dtype: int64
@@ -132,12 +120,12 @@ pipeline:
             kind: request_aligned
             axis: 0
         role:
-          kind: opaque
+          kind: runtime
+          version: '1.0'
+          role: eos_token_ids
         source:
-          kind: application
-          name: eos_ids
-        required: false
-        default: 127
+          kind: request
+        required: true
       request.eos_lengths:
         contract:
           dtype: int64
@@ -148,12 +136,12 @@ pipeline:
             kind: request_aligned
             axis: 0
         role:
-          kind: opaque
+          kind: runtime
+          version: '1.0'
+          role: eos_token_lengths
         source:
-          kind: application
-          name: eos_lengths
-        required: false
-        default: 1
+          kind: request
+        required: true
       request.row_max_iterations:
         contract:
           dtype: int64
@@ -1368,3 +1356,9 @@ pipeline:
           rank: 1
           shape:
           - 1
+package:
+  tokenizer:
+    byte_level: false
+    special_tokens:
+      eos_token_id:
+      - 127
diff --git a/tests/fixtures/onnx_genai_workflows/vlm/inference_metadata.yaml b/tests/fixtures/onnx_genai_workflows/vlm/inference_metadata.yaml
index f2dce1a06..8c503adb1 100644
--- a/tests/fixtures/onnx_genai_workflows/vlm/inference_metadata.yaml
+++ b/tests/fixtures/onnx_genai_workflows/vlm/inference_metadata.yaml
@@ -1,4 +1,4 @@
-schema_version: v1
+schema_version: v1.2
 preprocessing:
   image:
     transforms:
@@ -155,13 +155,13 @@ pipeline:
             kind: request_aligned
             axis: 0
         role:
-          kind: opaque
+          kind: runtime
+          version: '1.0'
+          role: eos_token_ids
         source:
-          kind: application
-          name: eos_ids
-        required: false
-        default: 2
-      request.eos_lengths:
+          kind: request
+        required: true
+      request.row_max_iterations:
         contract:
           dtype: int64
           rank: 1
@@ -174,10 +174,10 @@ pipeline:
           kind: opaque
         source:
           kind: application
-          name: eos_lengths
+          name: row_max_iterations
         required: false
-        default: 1
-      request.row_max_iterations:
+        default: -1
+      request.eos_lengths:
         contract:
           dtype: int64
           rank: 1
@@ -187,24 +187,12 @@ pipeline:
             kind: request_aligned
             axis: 0
         role:
-          kind: opaque
-        source:
-          kind: application
-          name: row_max_iterations
-        required: false
-        default: -1
-      package.eos_ids:
-        contract:
-          dtype: int64
-          rank: 1
-          shape:
-          - 1
-        role:
-          kind: opaque
+          kind: runtime
+          version: '1.0'
+          role: eos_token_lengths
         source:
-          kind: literal
-        required: false
-        default: 2
+          kind: request
+        required: true
       package.max_context:
         contract:
           dtype: int64
@@ -1534,3 +1522,10 @@ pipeline:
           rank: 1
           shape:
           - 1
+package:
+  tokenizer:
+    byte_level: false
+    special_tokens:
+      eos_token_id:
+      - 2
+      image_token_id: 32000