BU113 Benchmark Validity under B_U BU113|B_U 锚定下的 Benchmark 有效性

BU113, “Benchmark Validity under B_U,” formulates benchmark validity as a B_U-anchored downstream formal reconstruction. Its object is the class of situations in which a benchmark, evaluation protocol, leaderboard, task set, metric, score, ranking, test suite, human-preference evaluation, safety benchmark, alignment benchmark, or flourishing benchmark claims to measure a model’s real capability, reliability, safety, alignment, usefulness, or contribution to human flourishing. The paper places this claim under the B_U chain of boundary consistency, dynamic completeness, admissible path space, structured pruning, residual gating, feedback, and settlement standing. Benchmark validity is therefore treated as a local formal-evaluative act whose force depends on whether the task boundary is declared, the measured object is correctly identified, the metric corresponds to the admissible target, residuals are cleared, score transport remains stable, and the result can enter a settlement-compatible chain.

The central judgment is direct: a benchmark score is a candidate readout. It becomes valid only when the score preserves the right object across task boundary, metric design, data construction, sampling conditions, evaluator behavior, deployment context, and real-world feedback. A benchmark can be rigorous, standardized, statistically sophisticated, widely cited, and operationally useful while still failing to deliver standing. Under B_U, validity requires more than performance measurement. It requires a declared admissible domain, a clear target object, an invariant evaluation core, residual audit, cross-context stability, and settlement-compatible consequences. A score can rank systems inside a local testing window; the B_U ledger adjudicates whether that ranking carries durable relevance on the real settlement surface.

BU113 develops this handshake through a sequence of formal positions: benchmark presentation, declared benchmark boundary, target object, metric function, admissible evaluation domain, score readout, benchmark quotient, evaluator transport, benchmark gap, deployment feedback, and settlement-compatible benchmark validity. These definitions translate the B_U source-axis discipline into a vocabulary familiar to AI evaluation, machine learning benchmarking, model cards, safety evaluation, alignment testing, leaderboard design, psychometrics, measurement validity, formal verification, and governance auditing. The local formal shell follows a clear chain: task → metric → score → quotient → benchmark gap → feedback → settlement-compatible validity.

The theorem skeletons establish the main structure. The Declared Benchmark Boundary Theorem states that validity begins with an explicit task and evaluation boundary. The Target-Object Alignment Theorem requires the metric to measure the intended object rather than a proxy artifact. The Benchmark Quotient Theorem states that equivalent task presentations must preserve the same evaluation meaning. The Score Transport Stability Theorem requires score meaning to remain stable under admissible changes of prompt, format, sample, evaluator, and context. The Benchmark Gap Obstruction Theorem identifies leakage, gaming, proxy drift, evaluator bias, dataset contamination, and context collapse as validity failures. The Settlement-Compatible Benchmark Theorem returns benchmark claims to the B_U ledger through residual clearing, feedback compatibility, deployment relevance, and same-world-sheet settlement.

The final judgment is that benchmark validity is a bridge between formal evaluation and reality-facing standing. Benchmarks can assist ranking, diagnosis, comparison, and governance. Their durable force arises only when the measured object survives boundary shifts, metric pressure, evaluator variation, deployment friction, residual exposure, and real-world feedback. A benchmark score is locally informative; a B_U-compatible benchmark becomes structurally valid when it enters the same boundary and settles on the same world-sheet.