G0~G8 성과·동맹 측정 OS 작업 일괄 고정
8월 7일까지 워킹트리에만 남아 있던 미커밋 작업을 커밋한다. 여러 사본 폴더(worktree·clone)에 흩어져 있던 중간 스냅샷을 정리하기 전에 원본을 git 이력으로 고정하는 것이 목적이다. - contracts/routes/services: measurement, outcome_trajectory, rupture_repair, deliberate_practice, calibration_transfer, supervision_research, multimodal_alliance, continuous_improvement 계열 신규 모듈과 테스트 - infra/db/init: 07~16 마이그레이션(측정 기반~calibration transfer 실행) - apps/web: 세션 리뷰 카드·관리 화면·E2E 스펙 추가 - docs/ops: G0~G8 라이브 통합·배포·롤백 증거 문서와 evidence JSON/PNG - scripts: smoke·ledger·릴리스 에이전트·NAS 프리뷰 운영 스크립트 engine.public 로그 .bak과 apps/web/test-results 산출물은 커밋에서 제외했다.
This commit is contained in:
Yun Chan
parent
93dd8f82d7
commit
16e791e044
390 changed files with 243188 additions and 499 deletions
255
apps/api/app/contracts/continuous_improvement.py
Normal file
255
apps/api/app/contracts/continuous_improvement.py
Normal file
|
|
@ -0,0 +1,255 @@
|
|||
"""G8 자율 콘텐츠 생성·적대 검토·승격·운영 환류 계약."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Literal
|
||||
|
||||
from pydantic import BaseModel, ConfigDict, Field, model_validator
|
||||
|
||||
|
||||
ReviewDimension = Literal[
|
||||
"safety",
|
||||
"identity",
|
||||
"answer_leakage",
|
||||
"cultural_bias",
|
||||
"difficulty",
|
||||
"pii",
|
||||
"grounding",
|
||||
]
|
||||
FindingSeverity = Literal["blocker", "high", "moderate", "low"]
|
||||
FindingState = Literal["open", "resolved", "accepted_risk"]
|
||||
ModelChangeDecision = Literal["promote", "rollback", "quarantine"]
|
||||
|
||||
|
||||
class ContentSourceArtifact(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
source_id: str = Field(pattern=r"^oas-g8-source-[a-z0-9-]+$")
|
||||
version: str = Field(min_length=1, max_length=80)
|
||||
content_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
provenance_uri: str = Field(pattern=r"^(repo|db|audit)://[a-zA-Z0-9_./:-]+$")
|
||||
usage_status: Literal["approved", "restricted", "rejected"]
|
||||
citation_label: str = Field(min_length=1, max_length=300)
|
||||
|
||||
|
||||
class GeneratedContentDraft(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
draft_id: str = Field(pattern=r"^oas-g8-draft-[a-z0-9-]+$")
|
||||
content_kind: Literal["case", "rupture", "practice", "benchmark"]
|
||||
source_refs: tuple[str, ...] = Field(min_length=1)
|
||||
generation_model: str = Field(min_length=1, max_length=180)
|
||||
prompt_version: str = Field(min_length=1, max_length=80)
|
||||
prompt_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
payload_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
synthetic_identity_id: str = Field(pattern=r"^synthetic-identity-[a-z0-9-]+$")
|
||||
difficulty_level: int = Field(ge=1, le=5)
|
||||
hidden_answer_fingerprint: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
visible_answer_overlap_tokens: int = Field(ge=0)
|
||||
pii_findings: int = Field(ge=0)
|
||||
unsupported_clinical_claims: int = Field(ge=0)
|
||||
|
||||
|
||||
class RedTeamFinding(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
finding_id: str = Field(pattern=r"^oas-g8-finding-[a-z0-9-]+$")
|
||||
dimension: ReviewDimension
|
||||
severity: FindingSeverity
|
||||
state: FindingState
|
||||
evidence_ref: str = Field(min_length=1, max_length=220)
|
||||
remediation_ref: str | None = Field(default=None, max_length=220)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def require_resolution_evidence(self) -> "RedTeamFinding":
|
||||
if self.state == "resolved" and not self.remediation_ref:
|
||||
raise ValueError("resolved red-team finding requires remediation evidence")
|
||||
if self.state == "accepted_risk" and self.severity in {"blocker", "high"}:
|
||||
raise ValueError("blocker/high finding cannot be accepted as residual risk")
|
||||
return self
|
||||
|
||||
|
||||
class IndependentRedTeamReview(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
review_id: str = Field(pattern=r"^oas-g8-review-[a-z0-9-]+$")
|
||||
draft_id: str
|
||||
reviewer_agent_id: str = Field(min_length=1, max_length=180)
|
||||
dimensions: tuple[ReviewDimension, ...] = Field(min_length=3)
|
||||
findings: tuple[RedTeamFinding, ...]
|
||||
reviewed_payload_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
|
||||
@model_validator(mode="after")
|
||||
def require_unique_coverage(self) -> "IndependentRedTeamReview":
|
||||
if len(set(self.dimensions)) != len(self.dimensions):
|
||||
raise ValueError("red-team review dimensions must be unique")
|
||||
if any(item.dimension not in self.dimensions for item in self.findings):
|
||||
raise ValueError("red-team finding must belong to a reviewed dimension")
|
||||
return self
|
||||
|
||||
|
||||
class ContentBenchmarkQualification(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
benchmark_id: str = Field(pattern=r"^oas-g8-benchmark-[a-z0-9-]+$")
|
||||
draft_id: str
|
||||
variant_count: int = Field(ge=3)
|
||||
variant_pass_rate: float = Field(ge=0.0, le=1.0)
|
||||
answer_leakage_count: int = Field(ge=0)
|
||||
pii_finding_count: int = Field(ge=0)
|
||||
unsupported_claim_count: int = Field(ge=0)
|
||||
safety_failure_count: int = Field(ge=0)
|
||||
reward_hacking_count: int = Field(ge=0)
|
||||
evidence_refs: tuple[str, ...] = Field(min_length=1)
|
||||
|
||||
@property
|
||||
def qualified(self) -> bool:
|
||||
return (
|
||||
self.variant_pass_rate >= 0.85
|
||||
and self.answer_leakage_count == 0
|
||||
and self.pii_finding_count == 0
|
||||
and self.unsupported_claim_count == 0
|
||||
and self.safety_failure_count == 0
|
||||
and self.reward_hacking_count == 0
|
||||
)
|
||||
|
||||
|
||||
class ApprovedCatalogEntry(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
catalog_entry_id: str = Field(pattern=r"^oas-g8-catalog-[a-z0-9-]+$")
|
||||
draft_id: str
|
||||
payload_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
source_refs: tuple[str, ...] = Field(min_length=1)
|
||||
review_ids: tuple[str, ...] = Field(min_length=2)
|
||||
benchmark_id: str
|
||||
status: Literal["approved"] = "approved"
|
||||
clinical_claim_allowed: Literal[False] = False
|
||||
|
||||
|
||||
class ModelCalibrationSnapshot(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
snapshot_id: str = Field(pattern=r"^oas-g8-model-snapshot-[a-z0-9-]+$")
|
||||
model: str = Field(min_length=1, max_length=180)
|
||||
prompt_version: str = Field(min_length=1, max_length=80)
|
||||
benchmark_version: str = Field(min_length=1, max_length=80)
|
||||
task_accuracy: float = Field(ge=0.0, le=1.0)
|
||||
critical_miss_count: int = Field(ge=0)
|
||||
leakage_count: int = Field(ge=0)
|
||||
pii_count: int = Field(ge=0)
|
||||
calibration_error: float = Field(ge=0.0, le=1.0)
|
||||
subgroup_max_gap: float = Field(ge=0.0, le=1.0)
|
||||
|
||||
|
||||
class ModelChangeGateResult(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
baseline_snapshot_id: str
|
||||
candidate_snapshot_id: str
|
||||
decision: ModelChangeDecision
|
||||
reasons: tuple[str, ...] = Field(min_length=1)
|
||||
rollback_target_snapshot_id: str | None = None
|
||||
|
||||
@model_validator(mode="after")
|
||||
def require_rollback_target(self) -> "ModelChangeGateResult":
|
||||
if self.decision == "rollback" and not self.rollback_target_snapshot_id:
|
||||
raise ValueError("rollback decision requires a target snapshot")
|
||||
if self.decision != "rollback" and self.rollback_target_snapshot_id:
|
||||
raise ValueError("non-rollback decision cannot carry rollback target")
|
||||
return self
|
||||
|
||||
|
||||
class OperationalIncident(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
incident_id: str = Field(pattern=r"^oas-g8-incident-[a-z0-9-]+$")
|
||||
error_fingerprint: str = Field(pattern=r"^[a-f0-9]{64}$")
|
||||
affected_contract: str = Field(min_length=1, max_length=180)
|
||||
evidence_refs: tuple[str, ...] = Field(min_length=1)
|
||||
pii_included: Literal[False] = False
|
||||
|
||||
|
||||
class RegressionBacklogNode(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
node_id: str = Field(pattern=r"^oas-g8-node-[a-z0-9-]+$")
|
||||
node_type: Literal["reproduction_test", "implementation", "e2e", "runtime_proof"]
|
||||
depends_on: tuple[str, ...]
|
||||
evidence_ref: str | None = Field(default=None, max_length=220)
|
||||
status: Literal["pending", "passed", "failed"]
|
||||
|
||||
|
||||
class IncidentRegressionDag(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
incident_id: str
|
||||
nodes: tuple[RegressionBacklogNode, ...] = Field(min_length=4, max_length=4)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def require_ordered_closed_loop(self) -> "IncidentRegressionDag":
|
||||
ids = [item.node_id for item in self.nodes]
|
||||
if len(set(ids)) != len(ids):
|
||||
raise ValueError("incident DAG node ids must be unique")
|
||||
by_type = {item.node_type: item for item in self.nodes}
|
||||
if set(by_type) != {
|
||||
"reproduction_test",
|
||||
"implementation",
|
||||
"e2e",
|
||||
"runtime_proof",
|
||||
}:
|
||||
raise ValueError(
|
||||
"incident DAG requires reproduction, implementation, E2E, runtime"
|
||||
)
|
||||
known: set[str] = set()
|
||||
for item in self.nodes:
|
||||
if any(parent not in known for parent in item.depends_on):
|
||||
raise ValueError("incident DAG dependencies must point backward")
|
||||
known.add(item.node_id)
|
||||
return self
|
||||
|
||||
|
||||
class AgenticReleaseManifest(BaseModel):
|
||||
model_config = ConfigDict(extra="forbid", frozen=True, protected_namespaces=())
|
||||
|
||||
release_id: str = Field(pattern=r"^oas-g8-release-[a-z0-9-]+$")
|
||||
red_green_passed: bool
|
||||
contract_passed: bool
|
||||
e2e_passed: bool
|
||||
runtime_proof_passed: bool
|
||||
public_proof_passed: bool
|
||||
ssot_synced: bool
|
||||
evidence_refs: tuple[str, ...]
|
||||
|
||||
@property
|
||||
def releasable(self) -> bool:
|
||||
return all(
|
||||
(
|
||||
self.red_green_passed,
|
||||
self.contract_passed,
|
||||
self.e2e_passed,
|
||||
self.runtime_proof_passed,
|
||||
self.public_proof_passed,
|
||||
self.ssot_synced,
|
||||
)
|
||||
) and bool(self.evidence_refs)
|
||||
|
||||
|
||||
__all__ = [
|
||||
"AgenticReleaseManifest",
|
||||
"ApprovedCatalogEntry",
|
||||
"ContentBenchmarkQualification",
|
||||
"ContentSourceArtifact",
|
||||
"FindingSeverity",
|
||||
"FindingState",
|
||||
"GeneratedContentDraft",
|
||||
"IncidentRegressionDag",
|
||||
"IndependentRedTeamReview",
|
||||
"ModelCalibrationSnapshot",
|
||||
"ModelChangeDecision",
|
||||
"ModelChangeGateResult",
|
||||
"OperationalIncident",
|
||||
"RedTeamFinding",
|
||||
"RegressionBacklogNode",
|
||||
"ReviewDimension",
|
||||
]
|
||||
Loading…
Add table
Add a link
Reference in a new issue