mirror of
https://github.com/norandom/OpenUltraSAST.git
synced 2026-11-13 01:48:17 +00:00
feat(detection-core): apply bounded model corroboration
This commit is contained in:
+318
-3
@@ -3,12 +3,17 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from dataclasses import dataclass
|
||||
from dataclasses import dataclass, replace
|
||||
|
||||
import pytest
|
||||
from detection_core_fixtures import InMemoryContentResolver, stored_artifact
|
||||
from detection_core_fixtures import (
|
||||
InMemoryContentResolver,
|
||||
InMemoryPopulationCustody,
|
||||
custody_contract,
|
||||
stored_artifact,
|
||||
)
|
||||
|
||||
from openultrasast.core import canonical, evidence, family, model, provenance, semantics
|
||||
from openultrasast.core import canonical, evidence, family, model, model_apply, provenance, semantics
|
||||
|
||||
|
||||
def _digest(value: str) -> str:
|
||||
@@ -326,3 +331,313 @@ def test_core_model_contract_has_no_runtime_dependency_imports() -> None:
|
||||
assert "import dspy" not in source
|
||||
assert "litellm" not in source
|
||||
assert "openai" not in source
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class ApplicationCase:
|
||||
contract: family.FamilyContract
|
||||
record: evidence.EvidenceRecord
|
||||
request: model.ModelRequest
|
||||
result: model.ModelResult
|
||||
resolver: InMemoryContentResolver
|
||||
custody: InMemoryPopulationCustody
|
||||
custody_policy: provenance.PopulationCustodyPolicy
|
||||
custodian: str
|
||||
|
||||
|
||||
def _append_model_use(
|
||||
custody: InMemoryPopulationCustody,
|
||||
policy: provenance.PopulationCustodyPolicy,
|
||||
custodian: str,
|
||||
*,
|
||||
arm_id: str,
|
||||
purpose: provenance.LineagePurpose,
|
||||
artifact_digest: str,
|
||||
labels_accessed: bool = False,
|
||||
) -> provenance.LineageUse:
|
||||
population = custody.current_log(_digest("population"))
|
||||
request = provenance.PopulationUseRequest(
|
||||
population.population_digest,
|
||||
arm_id,
|
||||
purpose,
|
||||
labels_accessed,
|
||||
artifact_digest,
|
||||
population.log_digest,
|
||||
)
|
||||
receipt = custody.append(request, custodian)
|
||||
provenance.validate_population_custody_log(custody.current_log(population.population_digest), policy)
|
||||
return provenance.LineageUse(
|
||||
purpose,
|
||||
population.population_digest,
|
||||
labels_accessed,
|
||||
artifact_digest,
|
||||
receipt.receipt_digest,
|
||||
)
|
||||
|
||||
|
||||
def _application_case(
|
||||
*,
|
||||
contract: family.FamilyContract | None = None,
|
||||
record: evidence.EvidenceRecord | None = None,
|
||||
failed: bool = False,
|
||||
) -> ApplicationCase:
|
||||
contract = contract or _contract()
|
||||
record = record or _record(contract)
|
||||
question = contract.residual_questions[0]
|
||||
completeness = model.completeness_snapshot(record, contract)
|
||||
bundle = model.derive_model_input_bundle_digest(record, contract, question, completeness)
|
||||
initial, policy, custodian = custody_contract()
|
||||
custody = InMemoryPopulationCustody(initial, policy)
|
||||
artifacts = (
|
||||
(provenance.LineagePurpose.PROMPT, _digest("prompt")),
|
||||
(provenance.LineagePurpose.EXAMPLE, _digest("examples")),
|
||||
(provenance.LineagePurpose.FITTED_MODEL, _digest("fitted")),
|
||||
(provenance.LineagePurpose.DECISION_MAPPING, _digest("mapping")),
|
||||
)
|
||||
uses = tuple(
|
||||
_append_model_use(
|
||||
custody,
|
||||
policy,
|
||||
custodian,
|
||||
arm_id="model-arm",
|
||||
purpose=purpose,
|
||||
artifact_digest=artifact,
|
||||
)
|
||||
for purpose, artifact in artifacts
|
||||
)
|
||||
lineage = provenance.LineageManifest(
|
||||
"model-arm",
|
||||
bundle,
|
||||
uses,
|
||||
provenance.derive_lineage_manifest_digest("model-arm", bundle, uses),
|
||||
)
|
||||
program = model.make_model_program_declaration(
|
||||
family_contract_digest=family.family_contract_digest(contract),
|
||||
provider="provider",
|
||||
model_name="model-v1",
|
||||
program_digest=_digest("program"),
|
||||
prompt_digest=_digest("prompt"),
|
||||
example_set_digest=_digest("examples"),
|
||||
fitted_component_digest=_digest("fitted"),
|
||||
configuration_digest=_digest("model-configuration"),
|
||||
adapter_revision="adapter-v1",
|
||||
decision_mapping_digest=_digest("mapping"),
|
||||
input_bundle_digest=bundle,
|
||||
lineage_manifest_digest=lineage.manifest_digest,
|
||||
)
|
||||
request = model.make_model_request(
|
||||
record=record,
|
||||
contract=contract,
|
||||
residual_question_id=question.question_id,
|
||||
program=program,
|
||||
lineage=lineage,
|
||||
custody_policy_digest=policy.policy_digest,
|
||||
proposal_program_digests=(_digest("proposer"),),
|
||||
trusted_response_store_identities=("trusted-store",),
|
||||
maximum_response_bytes=256,
|
||||
)
|
||||
response = model.make_preserved_model_response(
|
||||
normalized_fields=(semantics.FieldValue("decision", semantics.FieldKind.TEXT, "unsafe"),),
|
||||
inline_redacted_text="[redacted] unsafe",
|
||||
artifact=None,
|
||||
redaction_policy_digest=_digest("redaction-policy"),
|
||||
)
|
||||
result = model.make_model_result(
|
||||
request=request,
|
||||
status=model.ModelStatus.FAILED if failed else model.ModelStatus.SUCCEEDED,
|
||||
response=response,
|
||||
family_decision=None if failed else "unsafe",
|
||||
failure_reason="provider_timeout" if failed else None,
|
||||
)
|
||||
return ApplicationCase(contract, record, request, result, InMemoryContentResolver({}), custody, policy, custodian)
|
||||
|
||||
|
||||
def test_apply_model_result_preserves_packet_and_adds_only_model_corroboration() -> None:
|
||||
case = _application_case()
|
||||
packet_digest = evidence.derive_packet_digest(case.record.packet)
|
||||
|
||||
derived = model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
assert isinstance(derived, evidence.EvidenceRecord)
|
||||
assert derived.packet is case.record.packet
|
||||
assert evidence.derive_packet_digest(derived.packet) == packet_digest
|
||||
assert derived.decision.basis is evidence.EvidenceBasis.MODEL_CORROBORATED
|
||||
assert derived.decision.family_code == "unsafe"
|
||||
assert derived.decision.provenance.model_result_digest == case.result.result_digest
|
||||
assert derived.decision.provenance.proof_record_digest is None
|
||||
|
||||
|
||||
def test_failed_model_result_preserves_the_deterministic_record() -> None:
|
||||
case = _application_case(failed=True)
|
||||
|
||||
derived = model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
assert derived is case.record
|
||||
|
||||
|
||||
def _incomplete_application_case() -> ApplicationCase:
|
||||
base_contract = _contract()
|
||||
unknown = family.DecisionRule(
|
||||
"unknown",
|
||||
evidence.NeutralOutcome.INDETERMINATE,
|
||||
("flow",),
|
||||
(semantics.RelationStatus.INCOMPLETE,),
|
||||
)
|
||||
contract = replace(base_contract, decisions=(*base_contract.decisions, unknown))
|
||||
complete = _record(contract)
|
||||
relation = complete.packet.transitions[0]
|
||||
incomplete_relation = semantics.make_relation(
|
||||
kind=relation.kind,
|
||||
role=relation.role,
|
||||
source=relation.source,
|
||||
target=relation.target,
|
||||
location=relation.location,
|
||||
status=semantics.RelationStatus.INCOMPLETE,
|
||||
conditions=relation.conditions,
|
||||
reasons=("missing",),
|
||||
provenance=relation.provenance,
|
||||
)
|
||||
packet = replace(complete.packet, transitions=(incomplete_relation,), completeness_reasons=("missing",))
|
||||
decision = evidence.Decision(
|
||||
"unknown",
|
||||
evidence.NeutralOutcome.INDETERMINATE,
|
||||
evidence.EvidenceBasis.DETERMINISTIC,
|
||||
packet.scope,
|
||||
("missing",),
|
||||
evidence.derive_packet_digest(packet),
|
||||
provenance.DecisionProvenance(None, None, None, None, None, None),
|
||||
)
|
||||
return _application_case(contract=contract, record=evidence.EvidenceRecord(packet, decision, ""))
|
||||
|
||||
|
||||
def test_incomplete_residual_prerequisite_remains_indeterminate() -> None:
|
||||
case = _incomplete_application_case()
|
||||
|
||||
derived = model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
assert derived is case.record
|
||||
assert derived.decision.outcome is evidence.NeutralOutcome.INDETERMINATE
|
||||
assert derived.decision.basis is evidence.EvidenceBasis.DETERMINISTIC
|
||||
|
||||
|
||||
def test_model_application_refuses_proposer_adjudicator_collision_and_family_substitution() -> None:
|
||||
case = _application_case()
|
||||
object.__setattr__(case.request, "proposal_program_digests", (case.request.program.program_digest,))
|
||||
with pytest.raises(model.ModelContractError, match="propose|adjudicat"):
|
||||
model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
case = _application_case()
|
||||
other_contract = replace(case.contract, family_id="other-family")
|
||||
with pytest.raises(model.ModelContractError, match="family|contract"):
|
||||
model_apply.apply_model_result(
|
||||
case.record,
|
||||
other_contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
|
||||
def test_model_application_refuses_stale_or_label_contaminated_lineage() -> None:
|
||||
case = _application_case()
|
||||
_append_model_use(
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
case.custodian,
|
||||
arm_id="model-arm",
|
||||
purpose=provenance.LineagePurpose.PROMPT,
|
||||
artifact_digest=case.request.program.prompt_digest,
|
||||
)
|
||||
with pytest.raises(model.ModelContractError, match="omitted|current"):
|
||||
model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
case = _application_case()
|
||||
_append_model_use(
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
case.custodian,
|
||||
arm_id="other-arm",
|
||||
purpose=provenance.LineagePurpose.DECISION_MAPPING,
|
||||
artifact_digest=case.request.program.decision_mapping_digest,
|
||||
labels_accessed=True,
|
||||
)
|
||||
with pytest.raises(model.ModelContractError, match="label|contaminat|spent"):
|
||||
model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
|
||||
def test_model_application_refuses_custody_policy_and_result_substitution() -> None:
|
||||
case = _application_case()
|
||||
other_policy = replace(case.custody_policy, policy_digest=_digest("other-policy"))
|
||||
with pytest.raises(model.ModelContractError, match="custody policy"):
|
||||
model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
other_policy,
|
||||
)
|
||||
|
||||
case = _application_case()
|
||||
object.__setattr__(case.result, "packet_digest", _digest("other-packet"))
|
||||
with pytest.raises(model.ModelContractError, match="packet"):
|
||||
model_apply.apply_model_result(
|
||||
case.record,
|
||||
case.contract,
|
||||
case.request,
|
||||
case.result,
|
||||
case.resolver,
|
||||
case.custody,
|
||||
case.custody_policy,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user