File size: 12,228 Bytes
58c2da3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 | """Single ordered orchestration shell for the completed blind dataset stages."""
from __future__ import annotations
from typing import Any, Protocol
from pydantic import BaseModel
from gcmd_classifier.datasets.errors import DatasetOrchestrationError
from gcmd_classifier.datasets.models import (
DatasetClassificationOutcome,
DatasetEvidencePacket,
ResolvedCMRCollection,
)
from gcmd_classifier.datasets.orchestration import DatasetRunController, DurableInferenceClient
from gcmd_classifier.datasets.runtime_models import (
ArtifactReference,
ArtifactType,
BlindPipelineTerminal,
CheckpointStatus,
DatasetRuntimeFailure,
PersistedExpertSourceBoundary,
PipelineStage,
TerminalOutcome,
)
from gcmd_classifier.datasets.runtime_persistence import (
DatasetArtifactStore,
canonical_json_bytes,
persist_expert_source_boundary,
verify_persisted_blind_result,
)
from gcmd_classifier.llm.base import ModelClient
class BlindStageServices(Protocol):
"""Adapters call the already implemented stage functions; they add no stage logic."""
def execute(
self,
stage: PipelineStage,
context: dict[PipelineStage, Any],
model_client: ModelClient,
) -> Any:
"""Execute one named stage and return its existing typed stage result."""
_ORDER = (
PipelineStage.CMR_SOURCE,
PipelineStage.IDENTITY,
PipelineStage.BLIND_BOUNDARY,
PipelineStage.README_DISCOVERY,
PipelineStage.README_SELECTION,
PipelineStage.README_RETRIEVAL,
PipelineStage.SOURCE_PERSISTENCE,
PipelineStage.EXTRACTION,
PipelineStage.PRODUCT_RESOLUTION,
PipelineStage.EVIDENCE_SELECTION,
PipelineStage.EVIDENCE_PACKET,
PipelineStage.CLASSIFICATION,
)
_ARTIFACT_TYPE = {
PipelineStage.CMR_SOURCE: ArtifactType.CMR_RESPONSE,
PipelineStage.IDENTITY: ArtifactType.COLLECTION_SNAPSHOT,
PipelineStage.BLIND_BOUNDARY: ArtifactType.COLLECTION_SNAPSHOT,
PipelineStage.README_DISCOVERY: ArtifactType.README_METADATA,
PipelineStage.README_SELECTION: ArtifactType.README_METADATA,
PipelineStage.README_RETRIEVAL: ArtifactType.README_PDF,
PipelineStage.SOURCE_PERSISTENCE: ArtifactType.README_METADATA,
PipelineStage.EXTRACTION: ArtifactType.EXTRACTION_MANIFEST,
PipelineStage.PRODUCT_RESOLUTION: ArtifactType.PRODUCT_RESOLUTION,
PipelineStage.EVIDENCE_SELECTION: ArtifactType.EVIDENCE_SELECTION,
PipelineStage.EVIDENCE_PACKET: ArtifactType.EVIDENCE_PACKET,
PipelineStage.CLASSIFICATION: ArtifactType.BLIND_RESULT,
}
class BlindDatasetPipeline:
"""Execute the approved blind order with durable state and artifact boundaries."""
def __init__(
self,
*,
controller: DatasetRunController,
store: DatasetArtifactStore,
services: BlindStageServices,
model_client: ModelClient,
vocabulary_hash: str,
) -> None:
self.controller = controller
self.store = store
self.services = services
self.model_client = DurableInferenceClient(model_client, controller)
self.vocabulary_hash = vocabulary_hash
self._source_boundary: PersistedExpertSourceBoundary | None = None
def run(self) -> BlindPipelineTerminal:
context: dict[PipelineStage, Any] = {}
references: dict[PipelineStage, ArtifactReference] = {}
try:
for stage in _ORDER:
self.controller.begin(stage, f"Begin {stage.value}.")
value = self.services.execute(stage, context, self.model_client)
if stage is PipelineStage.IDENTITY:
if not isinstance(value, ResolvedCMRCollection):
raise DatasetOrchestrationError(
"EXPERT_SOURCE_BINDING_MISSING",
"Identity stage must return the resolved CMR collection.",
)
self._source_boundary = persist_expert_source_boundary(
store=self.store,
run=self.controller.state.run,
session_id=self.controller.state.session.session_id,
collection=value,
cmr_source_artifact=references[PipelineStage.CMR_SOURCE],
)
context.clear()
context[PipelineStage.CMR_SOURCE] = references[PipelineStage.CMR_SOURCE]
value = self._source_boundary.blind_view
if stage is PipelineStage.README_DISCOVERY and value is None:
self.controller.transition(
stage, CheckpointStatus.SKIPPED, "No README candidate was discovered."
)
return self._terminal(TerminalOutcome.COMPLETED_SKIPPED)
if stage is PipelineStage.SOURCE_PERSISTENCE:
# The stage exists as an explicit service checkpoint; exact source bytes
# were already published at their acquisition boundaries.
reference = references[PipelineStage.README_RETRIEVAL]
else:
reference = self._persist(stage, value, context)
references[stage] = reference
context[stage] = value
checkpoint_artifacts = (reference,)
if stage is PipelineStage.IDENTITY and self._source_boundary is not None:
checkpoint_artifacts = (
reference,
self._source_boundary.cmr_source_artifact,
self._source_boundary.sealed_expert_artifact,
self._source_boundary.expert_source_binding_artifact,
)
self.controller.complete(
stage, f"Completed {stage.value}.", artifacts=checkpoint_artifacts
)
if stage is PipelineStage.CLASSIFICATION:
self.controller.mark_blind_result(produced=True)
self.controller.begin(
PipelineStage.BLIND_RESULT_PERSISTENCE,
"Persist blind result.",
)
self.controller.complete(
PipelineStage.BLIND_RESULT_PERSISTENCE,
"Blind result atomically persisted.",
artifacts=(reference,),
)
self.controller.mark_blind_result(persisted=True)
result = context[PipelineStage.CLASSIFICATION]
packet = self._packet(context[PipelineStage.EVIDENCE_PACKET])
verified = verify_persisted_blind_result(
store=self.store,
run=self.controller.state.run,
session_id=self.controller.state.session.session_id,
reference=references[PipelineStage.CLASSIFICATION],
expected_packet_sha256=packet.packet_sha256,
expected_native_id=result.identity.native_id,
vocabulary_hash=self.vocabulary_hash,
source_boundary=self._required_source_boundary(),
)
self.controller.begin(
PipelineStage.BLIND_RESULT_VERIFICATION, "Verify persisted blind result."
)
self.controller.complete(
PipelineStage.BLIND_RESULT_VERIFICATION,
"Persisted blind-result hash verified.",
artifacts=(verified.artifact,),
)
self.controller.mark_blind_result(verified=True)
terminal = (
TerminalOutcome.COMPLETED_NOT_CLASSIFIED
if result.classification_outcome is DatasetClassificationOutcome.NOT_CLASSIFIED
else TerminalOutcome.COMPLETED_CLASSIFIED
)
return self._terminal(terminal, verified=verified)
except Exception as exc:
attempted = self.controller.state.classification_attempted
failure = DatasetRuntimeFailure(
code=getattr(exc, "code", "BLIND_PIPELINE_FAILED"),
stage=self.controller.state.checkpoints[-1].stage,
message="Blind dataset pipeline stage failed.",
retryable=False,
classification_attempted=attempted,
)
return self._terminal(
TerminalOutcome.POST_INFERENCE_FAILURE
if attempted
else TerminalOutcome.PRE_INFERENCE_FAILURE,
failure=failure,
)
def _required_source_boundary(self) -> PersistedExpertSourceBoundary:
if self._source_boundary is None:
raise DatasetOrchestrationError(
"COMPARISON_PROVENANCE_INCOMPLETE",
"Verified blind results require persisted expert source provenance.",
)
return self._source_boundary
def _persist(
self, stage: PipelineStage, value: Any, context: dict[PipelineStage, Any]
) -> ArtifactReference:
if stage is PipelineStage.CMR_SOURCE and hasattr(value, "response_bytes"):
data = value.response_bytes
elif stage is PipelineStage.README_RETRIEVAL and hasattr(value, "content"):
data = value.content
elif isinstance(value, bytes):
data = value
elif isinstance(value, BaseModel):
data = canonical_json_bytes(value)
elif isinstance(value, dict):
data = canonical_json_bytes(value)
else:
raise DatasetOrchestrationError(
"UNPERSISTABLE_STAGE_RESULT", "Stage did not return a typed persistable value."
)
if stage not in {PipelineStage.CMR_SOURCE} and b'"ScienceKeywords"' in data:
raise DatasetOrchestrationError(
"BLIND_BOUNDARY_VIOLATION", "Expert keywords entered a blind artifact."
)
upstream = tuple(reference.sha256 for reference in self._upstream_references(context))
name = (
"blind_result.json"
if stage is PipelineStage.CLASSIFICATION
else f"artifacts/{stage.value}.json"
)
if stage is PipelineStage.README_RETRIEVAL:
name = "selected_readme.pdf"
return self.store.persist_bytes(
run=self.controller.state.run,
session_id=self.controller.state.session.session_id,
relative_name=name,
data=data,
artifact_type=_ARTIFACT_TYPE[stage],
upstream_sha256=upstream,
)
def _upstream_references(
self, context: dict[PipelineStage, Any]
) -> tuple[ArtifactReference, ...]:
refs: list[ArtifactReference] = []
for checkpoint in self.controller.state.checkpoints:
if checkpoint.status is CheckpointStatus.COMPLETED:
refs.extend(checkpoint.artifact_references)
return tuple(refs)
@staticmethod
def _packet(value: Any) -> DatasetEvidencePacket:
if isinstance(value, DatasetEvidencePacket):
return value
packet = getattr(value, "packet", None)
if isinstance(packet, DatasetEvidencePacket):
return packet
raise DatasetOrchestrationError(
"PACKET_RESULT_INVALID", "Evidence packet stage returned no typed packet."
)
def _terminal(
self,
outcome: TerminalOutcome,
*,
verified: Any = None,
failure: DatasetRuntimeFailure | None = None,
) -> BlindPipelineTerminal:
status = (
CheckpointStatus.SKIPPED
if outcome is TerminalOutcome.COMPLETED_SKIPPED
else CheckpointStatus.FAILED
if failure is not None
else CheckpointStatus.COMPLETED
)
self.controller.transition(
PipelineStage.TERMINAL,
status,
f"Terminal outcome: {outcome.value}.",
terminal_outcome=outcome,
)
return BlindPipelineTerminal(
state=self.controller.state,
terminal_outcome=outcome,
verified_blind_result=verified,
failure=failure,
)
|