LGD-Eval Evidence Grader v1.2: Evidence-Graded Assessment Engine for Agentic AI Trust — Authorization Validity Dimension

Abstract

Standalone technical note and runnable reference implementation of the LGD-Eval Evidence Grader v1.2, released under CC BY 4.0. It turns structured evidence into domain-level, defensible assessment conclusions — no composite scores, evidence-first, reproducible output — and serves as the reference implementation for ITU-T FG-TIDA contribution C2 (Assessment Criteria and Benchmarks for Trustworthy Agentic AI Systems), ToR 3.4 / 4.5. What is new in v1.2 Authorization validity dimension (new in v1.2): an independent four-state result per evidence item / domain — verified / unresolved / rejected / quarantined — with machine-readable reason codes and optional relying-party actions. Key semantics: missing authorization (unresolved) is never equal to confirmed out-of-scope authorization (rejected). Evidence grading V/S/M/L/P (GRADE-adapted): V = verified/independently reproduced, S = strong converging method-level evidence, M = moderate (partial reproduction or single method), L = limited (observation or single source), P = pending (not yet testable). Payload digest: end-to-end reproducible hash over the whole input, now including the case-level authorization object. Failure semantics (v1.1): per-item / per-domain failure_class = quarantined (malformed input) / rejected (grade=P) / unresolved (link not confirmable) / none; aggregate takes the most severe class (quarantined > rejected > unresolved > none). Worked examples B1 & B2: B1 = unverifiable issuer/scope conclusion (authorization unresolved); B2 = confirmed out-of-scope issuer (authorization rejected). Input and conclusion JSONs are included in the attached bundle. Bilingual summary / 中文摘要 LGD-Eval 证据分级器 v1.2 的独立技术说明与可运行参考实现(CC BY 4.0)。将结构化证据转换为域级、可辩护的评估结论——禁止复合分、证据优先、输出可复现;供 ITU-T FG-TIDA C2 贡献稿使用(ToR 3.4 / 4.5)。v1.2 新增授权有效性独立维度(verified / unresolved / rejected / quarantined 四态 + 机器可读原因码 + 依赖方动作);证据分级沿用 V/S/M/L/P;payload digest 已纳入 authorization,端到端可复现;附 B1(issuer/scope 不可核实→unresolved)与 B2(确认越权→rejected)两组 worked examples 的输入与结论 JSON。 Independence statement: this is a standalone evaluation-line record (LGD-Eval, neutral methodology perspective) and is not a new version of the Lifecycle Governance Doctrine (LGD) theory record; it shares no version chain with it.