From 5670e5162ef24c407faf38a7dac6a9c6c2b941d5 Mon Sep 17 00:00:00 2001
From: echobt <154886644+echobt@users.noreply.github.com>
Date: Tue, 13 Oct 2026 16:33:18 +0000
Subject: [PATCH 1/2] feat(od): activate od-base preset, remove od-tiny; light
 CI excludes od-marked tests

---
 .github/workflows/ci.yml                      |  4 +--
 docs/schemas/RunManifest.json                 |  1 -
 pyproject.toml                                |  1 +
 .../models/opendecision/__init__.py           |  5 ++--
 src/hypertrain/protocol/messages.py           |  3 +--
 src/hypertrain/protocol/messages_v2.py        |  5 +---
 tests/auditor/test_island_bisect_v2.py        |  4 +--
 tests/auditor/test_island_replay_v2.py        |  6 ++++-
 tests/auditor/test_od_bisect.py               |  2 ++
 tests/challenge/test_authority_snapshot_v2.py |  2 ++
 tests/challenge/test_carry_genesis_v2.py      |  2 ++
 tests/e2e/test_service_network_v2_e2e.py      |  1 +
 tests/layout/test_od_island_v2.py             |  2 ++
 tests/miner/test_island_launch_v2.py          |  1 +
 tests/models/test_opendecision.py             |  2 ++
 tests/protocol/test_network_v2.py             |  2 ++
 tests/protocol/test_od_fields.py              | 26 +++++++++----------
 tests/relay/test_transport_v2.py              |  3 +++
 tests/trainer/test_arch_dispatch.py           |  2 +-
 19 files changed, 44 insertions(+), 30 deletions(-)

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 4240f8c..5c065cf 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -81,7 +81,7 @@ jobs:
           done
           printf 'shard %s/%s -> %s files\n' "$SHARD" "$SHARDS" "${#SEL[@]}"
           if [ "${#SEL[@]}" -eq 0 ]; then echo "no files in shard"; exit 0; fi
-          uv run --frozen pytest -q --durations=25 \
+          uv run --frozen pytest -q --durations=25 -m "not od" \
             --deselect=tests/challenge/test_capacity_outer_v2.py::test_actual_bounded_outer_child_matches_original_signed_replay \
             --deselect=tests/challenge/test_service_capacity_v2.py::test_real_capacity_launch_failure_keeps_debit_and_prevents_restart_rerun \
             "--deselect=tests/miner/test_capacity_runtime_v2.py::test_actual_small_kernel_attempt_and_cleanup[success]" \
@@ -113,4 +113,4 @@ jobs:
           enable-cache: true
           cache-dependency-glob: "uv.lock"
       - run: uv sync --locked --all-extras --group dev
-      - run: uv run --frozen pytest -q --durations=25 tests/${{ matrix.suite }}
+      - run: uv run --frozen pytest -q --durations=25 -m "not od" tests/${{ matrix.suite }}
diff --git a/docs/schemas/RunManifest.json b/docs/schemas/RunManifest.json
index 9d14279..529f447 100644
--- a/docs/schemas/RunManifest.json
+++ b/docs/schemas/RunManifest.json
@@ -680,7 +680,6 @@
         },
         "preset": {
           "enum": [
-            "od-tiny",
             "od-base",
             "od-large"
           ],
diff --git a/pyproject.toml b/pyproject.toml
index 00f62cd..db62419 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -65,6 +65,7 @@ testpaths = ["tests"]
 # the `pytest` console script does not put the rootdir on sys.path, `python -m pytest` does.
 pythonpath = ["."]
 addopts = "-ra --strict-markers"
+markers = ["od: builds a real od-base OpenDecision model (excluded from PR CI)"]
 
 [tool.ruff]
 target-version = "py312"
diff --git a/src/hypertrain/models/opendecision/__init__.py b/src/hypertrain/models/opendecision/__init__.py
index 10ca07b..f2cb0ae 100644
--- a/src/hypertrain/models/opendecision/__init__.py
+++ b/src/hypertrain/models/opendecision/__init__.py
@@ -88,9 +88,8 @@ def check_manifest(m: RunManifest | RunManifestV2) -> None:
     if ms.param_count != want:
         raise ValueError(f"param_count {ms.param_count} != {want} trainable for {od.objective}")
     n = PRESETS[od.preset].max_len
-    # A6 pins N = max_len (1024) for real presets; od-tiny (tests only) may use any N <= max_len.
-    tiny_ok = od.preset == "od-tiny" and ms.seq_len <= n
-    if od.objective == "mlm" and ms.seq_len != n and not tiny_ok:
+    # A6 pins N = max_len (1024).
+    if od.objective == "mlm" and ms.seq_len != n:
         raise ValueError(f"mlm seq_len must be {n} for {od.preset} (A6)")
 
 
diff --git a/src/hypertrain/protocol/messages.py b/src/hypertrain/protocol/messages.py
index 404ba59..239d7c5 100644
--- a/src/hypertrain/protocol/messages.py
+++ b/src/hypertrain/protocol/messages.py
@@ -50,7 +50,6 @@ def _omit_none(v: object) -> bool:
 
 OD_PRESETS: dict[str, tuple[int, int, int, int, int]] = {
     # preset: (d_model, n_layers, n_heads, head_layers, vocab)
-    "od-tiny": (64, 2, 4, 1, 259),
     "od-base": (768, 12, 12, 2, 50432),
     "od-large": (1024, 24, 16, 4, 50432),
 }
@@ -70,7 +69,7 @@ def record_len(self) -> int:
 
 
 class ODSpec(_M):
-    preset: Literal["od-tiny", "od-base", "od-large"]
+    preset: Literal["od-base", "od-large"]
     head_layers: Pos
     objective: Literal["mlm", "distill", "decision"]
     mask_ratio: F32
diff --git a/src/hypertrain/protocol/messages_v2.py b/src/hypertrain/protocol/messages_v2.py
index eb23577..2511b90 100644
--- a/src/hypertrain/protocol/messages_v2.py
+++ b/src/hypertrain/protocol/messages_v2.py
@@ -151,10 +151,7 @@ def _assign_unit_and_od(self) -> TrainingManifestV2:
                     raise ValueError("OD record required iff objective is not MLM")
                 if od.record is not None and ms.seq_len != od.record.record_len() - 1:
                     raise ValueError("OD record length must equal seq_len+1")
-                if od.objective == "mlm" and (
-                    (od.preset == "od-tiny" and ms.seq_len > 64)
-                    or (od.preset != "od-tiny" and ms.seq_len != 1024)
-                ):
+                if od.objective == "mlm" and ms.seq_len != 1024:
                     raise ValueError("OD MLM sequence length differs from pinned preset")
                 if ds.sample_format != "u16[seq_len+1] token ids":
                     raise ValueError("OD requires u16 records")
diff --git a/tests/auditor/test_island_bisect_v2.py b/tests/auditor/test_island_bisect_v2.py
index b0575f7..d1aba1f 100644
--- a/tests/auditor/test_island_bisect_v2.py
+++ b/tests/auditor/test_island_bisect_v2.py
@@ -61,9 +61,7 @@ def get(i: int):
         (2, 1, None),
         (2, 2, None),
         (4, 2, None),
-        (2, 1, "mlm"),
-        (2, 1, "decision"),
-        (2, 1, "distill"),
+        *(pytest.param(2, 1, o, marks=pytest.mark.od) for o in ("mlm", "decision", "distill")),
     ],
 )
 def test_honest_traces_when_real_island_hooks(n: int, ep: int, objective: str | None) -> None:
diff --git a/tests/auditor/test_island_replay_v2.py b/tests/auditor/test_island_replay_v2.py
index a2b255e..70815d9 100644
--- a/tests/auditor/test_island_replay_v2.py
+++ b/tests/auditor/test_island_replay_v2.py
@@ -146,7 +146,10 @@ def fixture(objective: str | None = None, state_policy: str = "carry"):
     return job, cache, blobs, get, result, key
 
 
-@pytest.mark.parametrize("objective", [None, "mlm", "decision", "distill"])
+@pytest.mark.parametrize(
+    "objective",
+    [None, *(pytest.param(o, marks=pytest.mark.od) for o in ("mlm", "decision", "distill"))],
+)
 def test_full_carry_and_window_equality(objective: str | None) -> None:
     job, cache, blobs, get, expected, _ = fixture(objective)
     outputs = emulate(2, lambda c: audit_island(job, c, get, *blobs, cache, now_round=2))
@@ -373,6 +376,7 @@ def test_missing_predecessor_reconstructs_once_under_two_h_budget() -> None:
         )
 
 
+@pytest.mark.od
 def test_verified_anchor_restart_corruption_and_od_warm_start(tmp_path: Path) -> None:
     from hypertrain.auditor.replay import tensor_root
     from hypertrain.models.opendecision import extend_params
diff --git a/tests/auditor/test_od_bisect.py b/tests/auditor/test_od_bisect.py
index 2347b51..130713d 100644
--- a/tests/auditor/test_od_bisect.py
+++ b/tests/auditor/test_od_bisect.py
@@ -15,6 +15,8 @@
 from hypertrain.trainer.loop import train_round
 from hypertrain.trainer.model import init_params
 
+pytestmark = pytest.mark.od
+
 N = 2
 HON = Keypair(b"\x07" * 32).ss58
 CHEAT = Keypair(b"\x08" * 32).ss58
diff --git a/tests/challenge/test_authority_snapshot_v2.py b/tests/challenge/test_authority_snapshot_v2.py
index 65a8c5b..79ae943 100644
--- a/tests/challenge/test_authority_snapshot_v2.py
+++ b/tests/challenge/test_authority_snapshot_v2.py
@@ -18,6 +18,8 @@
 from hypertrain.protocol.hashing import sha256_hex
 from hypertrain.protocol.jcs import canonicalize
 
+pytestmark = pytest.mark.od
+
 ROOT = Path(__file__).resolve().parents[2]
 spec = importlib.util.spec_from_file_location(
     "portable_authority_adapter", ROOT / "scripts/network_authority_snapshot.py"
diff --git a/tests/challenge/test_carry_genesis_v2.py b/tests/challenge/test_carry_genesis_v2.py
index f388cc6..5af8c15 100644
--- a/tests/challenge/test_carry_genesis_v2.py
+++ b/tests/challenge/test_carry_genesis_v2.py
@@ -18,6 +18,8 @@
 from hypertrain.trainer.config import TrainConfig
 from hypertrain.trainer.model import init_params
 
+pytestmark = pytest.mark.od
+
 spec = importlib.util.spec_from_file_location(
     "carry_genesis_fixture", Path(__file__).parents[1] / "e2e/test_service_network_v2_e2e.py"
 )
diff --git a/tests/e2e/test_service_network_v2_e2e.py b/tests/e2e/test_service_network_v2_e2e.py
index dde9da3..8b4980a 100644
--- a/tests/e2e/test_service_network_v2_e2e.py
+++ b/tests/e2e/test_service_network_v2_e2e.py
@@ -39,6 +39,7 @@ def carry_setup(manifest, policy, admission_policy, rows, tree):
     yield from fixtures.network.__wrapped__(tmp_path, request)
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("od_carry_network", ["mlm", "decision", "distill"], indirect=True)
 def test_od_carry_admission_reference_requires_real_full_state(od_carry_network):
     """First production prerequisite: genuine corresponding-manifest graduation work."""
diff --git a/tests/layout/test_od_island_v2.py b/tests/layout/test_od_island_v2.py
index daa5189..1797cb5 100644
--- a/tests/layout/test_od_island_v2.py
+++ b/tests/layout/test_od_island_v2.py
@@ -20,6 +20,8 @@
 from hypertrain.trainer.loop import Assignment
 from hypertrain.trainer.model import init_params
 
+pytestmark = pytest.mark.od
+
 
 def od_manifest(objective: str, n: int, zero1: bool, dtype: str = "fp32") -> RunManifestV2:
     spec = importlib.util.spec_from_file_location(
diff --git a/tests/miner/test_island_launch_v2.py b/tests/miner/test_island_launch_v2.py
index 94a84cd..b39486a 100644
--- a/tests/miner/test_island_launch_v2.py
+++ b/tests/miner/test_island_launch_v2.py
@@ -1026,6 +1026,7 @@ def test_d2_signed_reservation_rejects_before_dispatch(tmp_path, fault):
     assert len(lifecycle.j.all("network_plan_intake")) == (1 if fault == "replay" else 0)
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("objective", ["mlm", "decision", "distill"])
 def test_real_od_two_rank_worker_paths(tmp_path: Path, objective: str) -> None:
     import importlib.util
diff --git a/tests/models/test_opendecision.py b/tests/models/test_opendecision.py
index 622688f..06627e2 100644
--- a/tests/models/test_opendecision.py
+++ b/tests/models/test_opendecision.py
@@ -36,6 +36,8 @@
 from hypertrain.trainer.loop import replay, train_round
 from hypertrain.trainer.model import forward as dispatched_forward
 
+pytestmark = pytest.mark.od
+
 HEAD_PREFIXES = ("head.", "scorer.", "type_emb.", "unknown", "log_temp.")
 
 
diff --git a/tests/protocol/test_network_v2.py b/tests/protocol/test_network_v2.py
index d71b0ae..40eaeb4 100644
--- a/tests/protocol/test_network_v2.py
+++ b/tests/protocol/test_network_v2.py
@@ -124,6 +124,7 @@ def commit(tokens: int = 7680) -> CommitV2:
     )
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("n", [1, 2, 4, 8])
 @pytest.mark.parametrize("objective", ["mlm", "decision", "distill"])
 @pytest.mark.parametrize("zero1", [False, True])
@@ -147,6 +148,7 @@ def test_od_general_n_when_v2(n: int, objective: str, zero1: bool) -> None:
             RunManifest.model_validate(body)
 
 
+@pytest.mark.od
 def test_assign_unit_when_only_island_batch_is_divisible() -> None:
     # Given
     body = od_body()
diff --git a/tests/protocol/test_od_fields.py b/tests/protocol/test_od_fields.py
index e21fa36..d19cd17 100644
--- a/tests/protocol/test_od_fields.py
+++ b/tests/protocol/test_od_fields.py
@@ -27,20 +27,20 @@ def od_body() -> dict[str, Any]:
     b = copy.deepcopy(example_manifest().body())
     b["model"] |= {
         "arch": "od-encoder",
-        "n_layers": 2,
-        "d_model": 64,
-        "n_heads": 4,
-        "n_kv_heads": 4,
-        "d_ff": 256,
-        "vocab": 259,
-        "seq_len": 64,
-        "param_count": 187_140,
+        "n_layers": 12,
+        "d_model": 768,
+        "n_heads": 12,
+        "n_kv_heads": 12,
+        "d_ff": 3072,
+        "vocab": 50432,
+        "seq_len": 1024,
+        "param_count": 123_753_984,
         "compute_dtype": "bf16",
         "capacity_factor": f32hex(1.0),
         "aux_loss_coef": f32hex(0.0),
         "od": {
-            "preset": "od-tiny",
-            "head_layers": 1,
+            "preset": "od-base",
+            "head_layers": 2,
             "objective": "mlm",
             "mask_ratio": f32hex(0.3),
             "mask_seed": 7,
@@ -121,8 +121,8 @@ def test_unit_root_without_unit_rejected() -> None:
         lambda b: b["reference_spec"].update(profile="od-bf16-det-v1"),
         lambda b: b["reference_spec"].pop("profile"),
         lambda b: b["reference_spec"]["layout"].update(zero1=True),
-        lambda b: b["model"].update(vocab=260),
-        lambda b: b["model"]["od"].update(head_layers=2),
+        lambda b: b["model"].update(vocab=50433),
+        lambda b: b["model"]["od"].update(head_layers=1),
         lambda b: b["model"]["od"].update(objective="decision"),
         lambda b: b["model"]["od"].update(tokenizer_offset=4),
     ],
@@ -143,7 +143,7 @@ def test_decision_record_len_pins_seq_len() -> None:
     b = od_body()
     b["model"]["od"] |= {"objective": "decision", "record": rec}
     b["model"]["seq_len"] = 1835
-    # d_model etc. still od-tiny; only seq_len/record coupling is under test
+    # d_model etc. still od-base; only seq_len/record coupling is under test
     RunManifest.model_validate(b)
     b["model"]["seq_len"] = 1834
     with pytest.raises(ValidationError):
diff --git a/tests/relay/test_transport_v2.py b/tests/relay/test_transport_v2.py
index 2ad3c3d..79f363a 100644
--- a/tests/relay/test_transport_v2.py
+++ b/tests/relay/test_transport_v2.py
@@ -1039,6 +1039,7 @@ def command(*args, **kwargs):
         assert outcome["provider_after"] is None and outcome["proof_status"] == "CENSORED"
 
 
+@pytest.mark.od
 def test_provider_fix_pinned_kind_version_and_fail_closed(
     tmp_path: Path, monkeypatch: pytest.MonkeyPatch
 ) -> None:
@@ -1140,6 +1141,7 @@ def test_provider_fix_pinned_kind_version_and_fail_closed(
         module.require_capped_node(journal, "missing-marker")
 
 
+@pytest.mark.od
 def test_shim_raw_bytes_exit_and_pinned_kind_empty_preflight(tmp_path: Path) -> None:
     import json
     import os
@@ -1239,6 +1241,7 @@ def test_shim_raw_bytes_exit_and_pinned_kind_empty_preflight(tmp_path: Path) ->
     assert not any(c[0] in {"run", "create", "start"} for c in calls)
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("kernel_mode", ["valid", "wrong", "missing", "foreign", "forged"])
 def test_pinned_kind_contract_through_capped_start(tmp_path: Path, kernel_mode: str) -> None:
     import json
diff --git a/tests/trainer/test_arch_dispatch.py b/tests/trainer/test_arch_dispatch.py
index 3a150e8..fc406fe 100644
--- a/tests/trainer/test_arch_dispatch.py
+++ b/tests/trainer/test_arch_dispatch.py
@@ -66,7 +66,7 @@ def od_cfg() -> TrainConfig:
 def test_config_fills_od_fields(od_cfg: TrainConfig) -> None:
     m = od_cfg.model
     assert (m.arch, m.profile) == ("od-encoder", "od-bf16-det-eager-v1")
-    assert m.od is not None and m.od.preset == "od-tiny"
+    assert m.od is not None and m.od.preset == "od-base"
     hash(m)  # frozen ODSpec keeps ModelConfig hashable for the adapter cache
 
 

From a1158c250d6d26b606ebb809bfa39e55153e070d Mon Sep 17 00:00:00 2001
From: echobt <154886644+echobt@users.noreply.github.com>
Date: Tue, 13 Oct 2026 17:21:35 +0000
Subject: [PATCH 2/2] test(od): mark e2e OD-network tests od (excluded from
 light CI)

---
 tests/e2e/test_network_service_proof.py  | 5 +++++
 tests/e2e/test_service_network_v2_e2e.py | 3 +++
 2 files changed, 8 insertions(+)

diff --git a/tests/e2e/test_network_service_proof.py b/tests/e2e/test_network_service_proof.py
index 9447a99..14c09b9 100644
--- a/tests/e2e/test_network_service_proof.py
+++ b/tests/e2e/test_network_service_proof.py
@@ -499,6 +499,7 @@ def respond(request):
     )
 
 
+@pytest.mark.od
 def test_decoder_probe_consumer_real_routes_receipt_and_debit_metadata(tmp_path, monkeypatch):
     """One real reference/proof transition; metadata oracle, not independent gradients."""
     from types import SimpleNamespace
@@ -1328,6 +1329,7 @@ def test_endpoint_requires_explicit_https_without_ambient_credentials(tmp_path,
         driver.Settings.model_validate(body)
 
 
+@pytest.mark.od
 def test_digest_authority_rejects_before_service_write(tmp_path, signed_package):
     directory, _, _, _, _, _ = signed_package
     cfg = settings(tmp_path).model_copy(update={"snapshot": directory})
@@ -1336,6 +1338,7 @@ def test_digest_authority_rejects_before_service_write(tmp_path, signed_package)
     assert not cfg.evidence.exists()
 
 
+@pytest.mark.od
 def test_wrong_role_rejects_before_live_write(tmp_path, monkeypatch, signed_package):
     directory, digest, roles, template, _, _ = signed_package
     cfg = settings(tmp_path).model_copy(update={"snapshot": directory, "snapshot_digest": digest})
@@ -1409,6 +1412,7 @@ def forbidden(*args, **kwargs):
     assert not (tmp_path / "none.log").exists()
 
 
+@pytest.mark.od
 def test_missing_service_adapter_rejects_before_evidence(tmp_path, monkeypatch, signed_package):
     directory, digest, _, _, _, _ = signed_package
     monkeypatch.setattr(driver, "durable_destination", lambda path: True)
@@ -1583,6 +1587,7 @@ def fail_copy(*args, **kwargs):
     assert result["errors"] and not index["export_complete"]
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("fault", ["offset", "beacon", "valid"])
 def test_prepare_clock_probe_rejects_before_any_output(
     tmp_path, monkeypatch, fault, signed_package
diff --git a/tests/e2e/test_service_network_v2_e2e.py b/tests/e2e/test_service_network_v2_e2e.py
index 8b4980a..32f0421 100644
--- a/tests/e2e/test_service_network_v2_e2e.py
+++ b/tests/e2e/test_service_network_v2_e2e.py
@@ -146,6 +146,7 @@ def test_actual_shadow_trial_does_not_mutate_production(network):
     assert network.store._services(network.manifest.run_id())[0].balances().issued == 40000
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("network", ["mlm", "decision", "distill"], indirect=True)
 def test_od_two_rank_signed_service_reference_and_proof(network):
     # All OD objectives use the real signed admission/reference/proof/finality routes,
@@ -1379,6 +1380,7 @@ def _graduate_four(network):
     assert network.store._services(network.manifest.run_id())[0].balances().issued == 40000
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("od_carry_network", ["mlm", "decision", "distill"], indirect=True)
 def test_od_four_active_actual_client_two_round_carry_checkpoint(od_carry_network, monkeypatch):
     """Production service path with same-manifest admission and actual HTTP/TLS miners."""
@@ -1866,6 +1868,7 @@ async def now():
     assert verify_network_checkpoint(checkpoint, fixtures.COORD.ss58) == []
 
 
+@pytest.mark.od
 @pytest.mark.parametrize("od_carry_network", ["mlm"], indirect=True)
 def test_public_watch_layer_over_actual_fastapi_tls(od_carry_network, monkeypatch):
     """Funded original audits; real public CLI/TLS routes and committed long-poll wake."""