feat(detection-core): validate exact admission context

This commit is contained in:
Marius Ciepluch
2026-10-31 19:38:01 +01:00
parent 361c216402
commit f3adbfe333
9 changed files with 1406 additions and 108 deletions
+731
View File
@@ -0,0 +1,731 @@
"""Exact pre-feasibility admission context validation (Reqs 12.1, 12.2, 13.1)."""
from __future__ import annotations
from dataclasses import dataclass, replace
from typing import TypeVar
import pytest
from detection_core_fixtures import InMemoryContentResolver, stored_artifact
from test_core_admission_review import _attestation, _policy
from test_core_model import ApplicationCase, _application_case
from test_core_readiness import _chain
from test_core_validation_v2 import _catalog, _review_policy, _trust_policy
from openultrasast.core import canonical, custody, validation
from openultrasast.core.admission_context import AdmissionValidationContext
from openultrasast.core.admission_review import (
AdmissionManifestV2,
AdmissionThreshold,
PopulationCensus,
ScopeCensus,
derive_census_digest,
derive_census_payload_digest,
derive_manifest_digest,
derive_manifest_payload_digest,
derive_threshold_set_digest,
)
from openultrasast.core.family_catalog import FamilyCatalogEntry
from openultrasast.core.lineage import LineageManifest, LineageUse, derive_lineage_manifest_digest
from openultrasast.core.measurement import (
BaselineIdentity,
Comparator,
DenominatorKind,
DenominatorPolicy,
IntervalMethod,
MetricId,
MetricRule,
NumeratorKind,
ScopeKey,
Statistic,
derive_denominator_policy_digest,
derive_identity_set_digest,
standard_metric_rules,
)
from openultrasast.core.model import ModelProgramDeclaration
from openultrasast.core.readiness import (
AnalysisIdentity,
FixtureResult,
FixtureRole,
ReadinessEvidenceUse,
ReadinessResult,
ReadinessStage,
StageResult,
)
from openultrasast.core.validation import EvidenceValidationContext
_T = TypeVar("_T")
class MultiPopulationCustody:
def __init__(self, logs: tuple[custody.PopulationCustodyLog, ...], policy: custody.PopulationCustodyPolicy) -> None:
self.logs = {item.population_digest: item for item in logs}
self.policy = policy
def current_log(self, population_digest: str) -> custody.PopulationCustodyLog:
try:
return self.logs[population_digest]
except KeyError as exc:
raise custody.PopulationCustodyError("population custody log not found") from exc
def append(self, request: custody.PopulationUseRequest, custodian: str) -> custody.PopulationUseReceipt:
prior = self.current_log(request.population_digest)
receipt = custody.record_population_use(request, prior, custodian, self.policy)
self.logs[request.population_digest] = custody.append_population_use(prior, receipt, self.policy)
return receipt
class _PolicySubclass(DenominatorPolicy):
pass
class _RuleSubclass(MetricRule):
pass
class _TupleSubclass(tuple[MetricRule, ...]):
pass
class _ScopeSubclass(ScopeKey):
pass
class _ScalarSubclass(str):
pass
class _AdmissionContextSubclass(AdmissionValidationContext):
pass
class _EvidenceContextSubclass(EvidenceValidationContext):
pass
class _CatalogEntrySubclass(FamilyCatalogEntry):
pass
class _ModelProgramSubclass(ModelProgramDeclaration):
pass
class _LineageSubclass(LineageManifest):
pass
class _ReadinessResultSubclass(ReadinessResult):
pass
class _AnalysisIdentitySubclass(AnalysisIdentity):
pass
class _FixtureResultSubclass(FixtureResult):
pass
class _ReadinessTupleSubclass(tuple[ReadinessResult, ...]):
pass
def _as_subclass(value: object, subtype: type[_T]) -> _T:
changed = object.__new__(subtype)
for field, item in vars(value).items():
object.__setattr__(changed, field, item)
return changed
@dataclass(frozen=True)
class AdmissionCase:
manifest: AdmissionManifestV2
denominator: DenominatorPolicy
readiness: tuple[ReadinessResult, ...]
context: AdmissionValidationContext
port: MultiPopulationCustody
resolver: InMemoryContentResolver
@dataclass(frozen=True)
class ContextSetup:
model_case: ApplicationCase
port: MultiPopulationCustody
context: AdmissionValidationContext
threshold: AdmissionThreshold
threshold_lineage: LineageManifest
admission_population: str
def _append_lineage(
port: MultiPopulationCustody,
population: str,
arm: str,
purpose: custody.LineagePurpose,
artifact: str,
custodian: str,
*,
labels: bool = False,
) -> LineageManifest:
current = port.current_log(population)
request = custody.PopulationUseRequest(population, arm, purpose, labels, artifact, current.log_digest)
receipt = port.append(request, custodian)
use = LineageUse(purpose, population, labels, artifact, receipt.receipt_digest)
return LineageManifest(arm, artifact, (use,), derive_lineage_manifest_digest(arm, artifact, (use,)))
def _reviewed_census(population: str, scope: ScopeKey, disclosure: canonical.ContentRef) -> PopulationCensus:
producer = canonical.content_digest("benchmark-producer")
reviewer = canonical.content_digest("independent-reviewer")
row = ScopeCensus(scope, 8, 2, 2, 2, 4, 12, 3)
provisional_review = _attestation("0" * 64, producer, reviewer, disclosure)
provisional = PopulationCensus(population, (row,), False, "0" * 64, provisional_review, "0" * 64)
payload = derive_census_payload_digest(provisional)
census = replace(provisional, census_payload_digest=payload, review=_attestation(payload, producer, reviewer, disclosure))
return replace(census, census_digest=derive_census_digest(census))
def _resign_manifest(
manifest: AdmissionManifestV2,
disclosure: canonical.ContentRef,
) -> AdmissionManifestV2:
producer = canonical.content_digest("benchmark-producer")
reviewer = canonical.content_digest("independent-reviewer")
provisional = replace(
manifest,
manifest_payload_digest="0" * 64,
review=_attestation("0" * 64, producer, reviewer, disclosure),
manifest_digest="0" * 64,
)
payload = derive_manifest_payload_digest(provisional)
reviewed = replace(provisional, manifest_payload_digest=payload, review=_attestation(payload, producer, reviewer, disclosure))
return replace(reviewed, manifest_digest=derive_manifest_digest(reviewed))
def _support_lineages(
port: MultiPopulationCustody,
source_population: str,
custodian: str,
) -> tuple[LineageManifest, LineageManifest]:
matcher = _append_lineage(
port,
source_population,
"matcher-arm",
custody.LineagePurpose.MATCHING_BEHAVIOR,
canonical.content_digest("matcher"),
custodian,
)
promotion = _append_lineage(
port,
source_population,
"promotion-arm",
custody.LineagePurpose.PROMOTION,
canonical.content_digest("promotion"),
custodian,
)
return matcher, promotion
def _context_setup() -> ContextSetup:
from openultrasast.core import admission_context
model_case = _application_case()
contract = model_case.contract
catalog = _catalog(contract)
admission_population = canonical.content_digest("admission-population")
custodian = model_case.custodian
admission_log = custody.empty_population_custody_log(admission_population, custodian)
source_log = model_case.custody.current_log(canonical.content_digest("population"))
port = MultiPopulationCustody((source_log, admission_log), model_case.custody_policy)
matcher, promotion = _support_lineages(port, source_log.population_digest, custodian)
threshold = AdmissionThreshold(MetricId.VULNERABLE_RECALL, Statistic.WILSON_LOWER, Comparator.AT_LEAST, 0.5)
threshold_set = derive_threshold_set_digest((threshold,))
threshold_lineage = _append_lineage(
port,
source_log.population_digest,
"reference-arm",
custody.LineagePurpose.THRESHOLD,
threshold_set,
custodian,
)
evidence_context = validation.make_evidence_validation_context(
family_catalog=catalog,
fact_packs=(),
model_programs=(model_case.request.program,),
lineages=(model_case.request.lineage, matcher, promotion, threshold_lineage),
fact_review_policy=_review_policy(),
provenance_trust_policy=_trust_policy(),
custody_policy=model_case.custody_policy,
)
review_policy = _policy(canonical.content_digest("independent-reviewer"))
context = admission_context.make_admission_validation_context(
evidence=evidence_context,
admission_review_policy=review_policy,
proof_trust_policy=None,
)
return ContextSetup(model_case, port, context, threshold, threshold_lineage, admission_population)
def _baseline(setup: ContextSetup, denominator: DenominatorPolicy, scope: ScopeKey) -> BaselineIdentity:
from openultrasast.core import admission_context
catalog = setup.context.evidence.family_catalog
program_digest = setup.model_case.request.program.declaration_digest
threshold_set = derive_threshold_set_digest((setup.threshold,))
return BaselineIdentity(
"reference-arm",
setup.admission_population,
canonical.content_digest("admission-freeze"),
(scope,),
(catalog.entries[0].contract_digest,),
catalog.catalog_digest,
admission_context.derive_fact_schema_digest(),
validation.derive_fact_set_digest(()),
canonical.content_digest("acquisition"),
"analyzer@abc123",
canonical.content_digest("configuration"),
(program_digest,),
derive_identity_set_digest("model_program", (program_digest,)),
denominator.policy_digest,
threshold_set,
admission_context.derive_outcome_mapping_digest(catalog),
)
def _manifest(
setup: ContextSetup,
baseline: BaselineIdentity,
census: PopulationCensus,
readiness: tuple[ReadinessResult, ...],
disclosure: canonical.ContentRef,
) -> AdmissionManifestV2:
provisional_review = _attestation(
"0" * 64,
canonical.content_digest("benchmark-producer"),
canonical.content_digest("independent-reviewer"),
disclosure,
)
provisional = AdmissionManifestV2(
"0" * 64,
provisional_review,
True,
False,
baseline,
census,
tuple(sorted(item.result_digest for item in readiness)),
setup.port.current_log(setup.admission_population).log_digest,
baseline.denominator_policy_digest,
(setup.threshold,),
baseline.threshold_set_digest,
setup.threshold_lineage.manifest_digest,
setup.context.context_digest,
"0" * 64,
)
return _resign_manifest(provisional, disclosure)
def _case() -> AdmissionCase:
from openultrasast.core import admission_context
setup = _context_setup()
contract = setup.model_case.contract
scope = ScopeKey(contract.family_id, "python")
denominator = DenominatorPolicy(standard_metric_rules((scope,)), IntervalMethod.WILSON_95, "")
baseline = _baseline(setup, denominator, scope)
program_digest = setup.model_case.request.program.declaration_digest
catalog = setup.context.evidence.family_catalog
current_identity = admission_context.derive_analysis_identity(baseline, catalog, scope, program_digest)
readiness = _chain(identity=current_identity)[:6]
disclosure, resolver = stored_artifact(b'{"conflicts":[]}', artifact_id="admission-context-conflicts")
census = _reviewed_census(setup.admission_population, scope, disclosure)
manifest = _manifest(setup, baseline, census, readiness, disclosure)
return AdmissionCase(manifest, denominator, readiness, setup.context, setup.port, resolver)
def _validate(case: AdmissionCase) -> None:
from openultrasast.core.admission_context import validate_admission_context
validate_admission_context(
case.manifest,
case.denominator,
case.readiness,
case.context,
case.port,
case.resolver,
)
def _rebind_denominator(case: AdmissionCase, denominator: DenominatorPolicy) -> AdmissionCase:
object.__setattr__(denominator, "policy_digest", derive_denominator_policy_digest(denominator))
identity = replace(case.manifest.identity, denominator_policy_digest=denominator.policy_digest)
manifest = _resign_manifest(
replace(case.manifest, identity=identity, denominator_policy_digest=denominator.policy_digest),
case.manifest.review.conflict_disclosure,
)
return replace(case, manifest=manifest, denominator=denominator)
def _forged_policy(template: DenominatorPolicy, rules: tuple[MetricRule, ...]) -> DenominatorPolicy:
changed = object.__new__(DenominatorPolicy)
object.__setattr__(changed, "rules", rules)
object.__setattr__(changed, "interval_method", template.interval_method)
object.__setattr__(changed, "policy_digest", template.policy_digest)
return changed
def _smuggled_denominator(case: AdmissionCase, kind: str) -> DenominatorPolicy:
policy = case.denominator
first = policy.rules[0]
if kind == "policy_subclass":
changed_policy = object.__new__(_PolicySubclass)
object.__setattr__(changed_policy, "rules", policy.rules)
object.__setattr__(changed_policy, "interval_method", policy.interval_method)
object.__setattr__(changed_policy, "policy_digest", policy.policy_digest)
return changed_policy
if kind == "policy_state":
object.__setattr__(policy, "hidden", "state")
return policy
changed: MetricRule
if kind == "rule_subclass":
changed = _RuleSubclass(first.metric_id, first.scopes, first.numerator, first.denominator)
elif kind == "rule_state":
changed = replace(first)
object.__setattr__(changed, "hidden", "state")
elif kind == "scope_subclass":
scope = first.scopes[0]
changed = replace(first, scopes=(_ScopeSubclass(scope.family_id, scope.ecosystem),))
elif kind == "scope_state":
scope = replace(first.scopes[0])
object.__setattr__(scope, "hidden", "state")
changed = replace(first, scopes=(scope,))
elif kind == "scalar_subclass":
scope = first.scopes[0]
changed = replace(first, scopes=(ScopeKey(_ScalarSubclass(scope.family_id), scope.ecosystem),))
else:
return _forged_policy(policy, _TupleSubclass(policy.rules))
return _forged_policy(policy, (changed, *policy.rules[1:]))
def test_exact_reviewed_context_is_accepted_before_feasibility() -> None:
case = _case()
_validate(case)
assert case.port.current_log(case.manifest.identity.population_digest).status is custody.PopulationStatus.AVAILABLE
@pytest.mark.parametrize(
"field",
("family_contract_set_digest", "fact_set_digest", "model_program_digests", "outcome_mapping_digest"),
)
def test_claimed_digest_sets_are_rederived_from_canonical_objects(field: str) -> None:
case = _case()
if field == "model_program_digests":
models = (canonical.content_digest("other-model"),)
identity = replace(
case.manifest.identity,
model_program_digests=models,
model_program_set_digest=derive_identity_set_digest("model_program", models),
)
elif field == "family_contract_set_digest":
identity = replace(case.manifest.identity, family_contract_set_digest=canonical.content_digest("other-catalog"))
elif field == "fact_set_digest":
identity = replace(case.manifest.identity, fact_set_digest=canonical.content_digest("other-facts"))
else:
identity = replace(case.manifest.identity, outcome_mapping_digest=canonical.content_digest("other-mapping"))
changed = _resign_manifest(
replace(case.manifest, identity=identity),
case.manifest.review.conflict_disclosure,
)
with pytest.raises(ValueError, match="catalog|fact set|model program|outcome mapping"):
_validate(replace(case, manifest=changed))
def test_scope_and_denominator_are_exact_not_digest_only() -> None:
case = _case()
other_scope = ScopeKey("other-family", "python")
denominator = DenominatorPolicy(standard_metric_rules((other_scope,)), IntervalMethod.WILSON_95, "")
identity = replace(case.manifest.identity, denominator_policy_digest=denominator.policy_digest)
changed = _resign_manifest(
replace(
case.manifest,
identity=identity,
denominator_policy_digest=denominator.policy_digest,
),
case.manifest.review.conflict_disclosure,
)
with pytest.raises(ValueError, match="denominator scope"):
_validate(replace(case, manifest=changed, denominator=denominator))
def test_rederived_denominator_digest_cannot_hide_fixed_algebra_mutation() -> None:
case = _case()
first = case.denominator.rules[0]
changed_rule = replace(first, numerator=NumeratorKind.FIXED_DEFECT_CASES)
object.__setattr__(case.denominator, "rules", (changed_rule, *case.denominator.rules[1:]))
object.__setattr__(case.denominator, "policy_digest", derive_denominator_policy_digest(case.denominator))
identity = replace(case.manifest.identity, denominator_policy_digest=case.denominator.policy_digest)
changed = _resign_manifest(
replace(
case.manifest,
identity=identity,
denominator_policy_digest=case.denominator.policy_digest,
),
case.manifest.review.conflict_disclosure,
)
with pytest.raises(ValueError, match="fixed algebra"):
_validate(replace(case, manifest=changed))
@pytest.mark.parametrize(
"kind",
(
"policy_subclass",
"policy_state",
"rule_subclass",
"rule_state",
"tuple_subclass",
"scope_subclass",
"scope_state",
"scalar_subclass",
),
)
def test_denominator_graph_rejects_subclasses_and_injected_state(kind: str) -> None:
case = _case()
changed = _rebind_denominator(case, _smuggled_denominator(case, kind))
with pytest.raises(ValueError, match="closed"):
_validate(changed)
@pytest.mark.parametrize(
"member",
(
MetricId.PAIRED_CORRECTNESS,
NumeratorKind.CORRECT_PAIRS,
DenominatorKind.PAIRS,
IntervalMethod.WILSON_95,
Statistic.WILSON_LOWER,
Comparator.AT_LEAST,
),
)
def test_admission_graph_rejects_injected_enum_singleton_state(member: object) -> None:
case = _case()
object.__setattr__(member, "hidden", "state")
try:
with pytest.raises(ValueError, match="closed"):
_validate(case)
finally:
object.__delattr__(member, "hidden")
@pytest.mark.parametrize("kind", ("subclass", "state"))
def test_admission_context_root_is_exact(kind: str) -> None:
case = _case()
context: AdmissionValidationContext
if kind == "subclass":
context = _as_subclass(case.context, _AdmissionContextSubclass)
else:
context = case.context
object.__setattr__(context, "hidden", "state")
with pytest.raises(ValueError, match="closed|exact"):
_validate(replace(case, context=context))
@pytest.mark.parametrize(
"kind",
(
"root_subclass",
"root_state",
"entry_subclass",
"entry_state",
"program_subclass",
"program_state",
"lineage_subclass",
"lineage_state",
),
)
def test_evidence_context_graph_is_closed_before_digest_validation(kind: str) -> None:
case = _case()
evidence = case.context.evidence
changed: EvidenceValidationContext
if kind == "root_subclass":
changed = _as_subclass(evidence, _EvidenceContextSubclass)
elif kind == "root_state":
changed = evidence
object.__setattr__(changed, "hidden", "state")
elif kind.startswith("entry"):
entry = evidence.family_catalog.entries[0]
changed_entry = _as_subclass(entry, _CatalogEntrySubclass) if kind.endswith("subclass") else entry
if kind.endswith("state"):
object.__setattr__(changed_entry, "hidden", "state")
catalog = replace(evidence.family_catalog, entries=(changed_entry,))
changed = replace(evidence, family_catalog=catalog)
elif kind.startswith("program"):
program = evidence.model_programs[0]
changed_program = _as_subclass(program, _ModelProgramSubclass) if kind.endswith("subclass") else program
if kind.endswith("state"):
object.__setattr__(changed_program, "hidden", "state")
changed = replace(evidence, model_programs=(changed_program,))
else:
lineage = evidence.lineages[0]
changed_lineage = _as_subclass(lineage, _LineageSubclass) if kind.endswith("subclass") else lineage
if kind.endswith("state"):
object.__setattr__(changed_lineage, "hidden", "state")
changed = replace(evidence, lineages=(changed_lineage, *evidence.lineages[1:]))
context = replace(case.context, evidence=changed)
with pytest.raises(ValueError, match="closed"):
_validate(replace(case, context=context))
@pytest.mark.parametrize(
"kind",
(
"root_subclass",
"root_state",
"identity_subclass",
"identity_state",
"fixture_subclass",
"fixture_state",
"tuple_subclass",
),
)
def test_readiness_graph_and_container_are_closed(kind: str) -> None:
case = _case()
first = case.readiness[0]
changed_first: ReadinessResult
if kind == "root_subclass":
changed_first = _as_subclass(first, _ReadinessResultSubclass)
elif kind == "root_state":
changed_first = first
object.__setattr__(changed_first, "hidden", "state")
elif kind.startswith("identity"):
identity = _as_subclass(first.identity, _AnalysisIdentitySubclass) if kind.endswith("subclass") else first.identity
if kind.endswith("state"):
object.__setattr__(identity, "hidden", "state")
changed_first = _as_subclass(first, ReadinessResult)
object.__setattr__(changed_first, "identity", identity)
elif kind.startswith("fixture"):
fixture = first.fixture_results[0]
changed_fixture = _as_subclass(fixture, _FixtureResultSubclass) if kind.endswith("subclass") else fixture
if kind.endswith("state"):
object.__setattr__(changed_fixture, "hidden", "state")
changed_first = _as_subclass(first, ReadinessResult)
object.__setattr__(changed_first, "fixture_results", (changed_fixture, *first.fixture_results[1:]))
else:
readiness: tuple[ReadinessResult, ...] = _ReadinessTupleSubclass(case.readiness)
with pytest.raises(ValueError, match="tuple"):
_validate(replace(case, readiness=readiness))
return
readiness = (changed_first, *case.readiness[1:])
with pytest.raises(ValueError, match="closed"):
_validate(replace(case, readiness=readiness))
@pytest.mark.parametrize(
"member",
(
ReadinessStage.INSTRUMENT,
ReadinessEvidenceUse.REGRESSION_FIXTURE,
StageResult.PASSED,
FixtureRole.POSITIVE,
),
)
def test_readiness_graph_rejects_injected_enum_singleton_state(member: object) -> None:
case = _case()
object.__setattr__(member, "hidden", "state")
try:
with pytest.raises(ValueError, match="closed readiness graph"):
_validate(case)
finally:
object.__delattr__(member, "hidden")
def test_missing_or_stale_readiness_is_refused() -> None:
case = _case()
shortened = case.readiness[:-1]
missing = _resign_manifest(
replace(case.manifest, readiness_digests=tuple(sorted(item.result_digest for item in shortened))),
case.manifest.review.conflict_disclosure,
)
with pytest.raises(ValueError, match="development readiness"):
_validate(replace(case, manifest=missing, readiness=shortened))
stale_identity = replace(case.readiness[0].identity, acquisition_record_digest=canonical.content_digest("drift"))
stale = _chain(identity=stale_identity)[:6]
stale_manifest = _resign_manifest(
replace(case.manifest, readiness_digests=tuple(sorted(item.result_digest for item in stale))),
case.manifest.review.conflict_disclosure,
)
with pytest.raises(ValueError, match="stale|analysis identity|acquisition"):
_validate(replace(case, manifest=stale_manifest, readiness=stale))
def test_context_substitution_and_lineage_object_tampering_are_refused() -> None:
case = _case()
evidence_context = case.context.evidence
substituted_evidence = validation.make_evidence_validation_context(
family_catalog=evidence_context.family_catalog,
fact_packs=evidence_context.fact_packs,
model_programs=evidence_context.model_programs,
lineages=evidence_context.lineages[:-1],
fact_review_policy=evidence_context.fact_review_policy,
provenance_trust_policy=evidence_context.provenance_trust_policy,
custody_policy=evidence_context.custody_policy,
)
from openultrasast.core import admission_context
substituted = admission_context.make_admission_validation_context(
evidence=substituted_evidence,
admission_review_policy=case.context.admission_review_policy,
proof_trust_policy=None,
)
with pytest.raises(ValueError, match="validation context"):
_validate(replace(case, context=substituted))
object.__setattr__(evidence_context.lineages[0].uses[0], "labels_accessed", True)
with pytest.raises(ValueError, match="lineage|label"):
_validate(case)
def test_label_influenced_promotion_is_refused_while_admission_population_is_available() -> None:
case = _case()
promotion = next(item for item in case.context.evidence.lineages if item.uses[0].purpose is custody.LineagePurpose.PROMOTION)
source = promotion.uses[0].source_population_digest
current = case.port.current_log(source)
request = custody.PopulationUseRequest(
source,
promotion.arm_id,
custody.LineagePurpose.PROMOTION,
True,
promotion.subject_digest,
current.log_digest,
)
case.port.append(request, current.custodian_identity)
assert case.port.current_log(case.manifest.identity.population_digest).status is custody.PopulationStatus.AVAILABLE
with pytest.raises(ValueError, match="omitted current label-influenced use"):
_validate(case)
def test_current_custody_head_must_remain_available() -> None:
case = _case()
population = case.manifest.identity.population_digest
current = case.port.current_log(population)
request = custody.PopulationUseRequest(
population,
case.manifest.identity.arm_id,
custody.LineagePurpose.ADMISSION_LABEL_OPEN,
True,
case.manifest.manifest_digest,
current.log_digest,
)
case.port.append(request, current.custodian_identity)
spent = _resign_manifest(
replace(case.manifest, custody_log_digest=case.port.current_log(population).log_digest),
case.manifest.review.conflict_disclosure,
)
with pytest.raises(ValueError, match="spent|available"):
_validate(replace(case, manifest=spent))