This entity-resolution guardrail has the right benchmark shape—and zero results

A new paper on agentic entity resolution does something unusually useful for a conceptual architecture: it states, in advance, what evidence would count against its central idea. It also reports no implementation and no experimental results. That combination makes the paper more valuable as a benchmark specification than as evidence that its proposed guardrail works.

The intervention is a commit boundary

The proposal separates evidence an agent may use while reasoning from evidence allowed to authorize a persistent identity-changing mutation. A protected merge or graph update would require claim-specific eligible evidence, resolved material conflicts, valid scope, and current graph, evidence and policy versions. The authors explicitly say this does not stop an agent forming a wrong hypothesis or letting that hypothesis shape later retrieval; dependency-aware evidence flow is left for future work.

Four conditions, not one bundled comparison

The proposed evaluation has four arms: unconstrained agentic entity resolution; conventional identifier rules with default-deny behavior; a generic two-phase authorization boundary; and the DIO/ODI-informed challenger. The important comparison is the last two. They must share the same cases, domain facts, evidence-promotion process, validator class, resources and decision budget. If the generic gate independently derives the same constraints and produces identical outcomes, the paper says that is not evidence for the challenger.

That design avoids a common benchmark error: crediting a named framework for gains actually supplied by more annotations, stricter abstention or a larger information budget. It also identifies a remaining confound—policy-authoring skill—and recommends independent author teams, though it does not require them.

Safety without coverage is not a win

The metric set is deliberately multi-objective. Policy-Enforcement Violation Rate measures commits that violate the stated authorization contract. Referentially False Commit Rate measures false identity-changing commits among committed mutations, while Safe Resolution Coverage keeps deferred cases in the denominator. Unwarranted Abstention Rate and False Positive Intervention Rate expose systems that appear safe only because they halt useful work. The protocol also asks for precision, recall, F1, stale or duplicate commits, unresolved timeout outcomes and cost per correct resolution.

The proposed falsification rule is clear: the DIO/ODI mapping earns no support if it fails to lower false commits at comparable safe coverage, or fails to improve coverage without materially increasing false commits, relative to the matched generic gate. A generic gate that matches or Pareto-dominates the challenger counts against incremental value for that task and policy family.

What is still missing

There is no selected corpus, frozen split, implementation, run count, uncertainty estimate or measured numerator for any metric. The paper suggests temporal and ambiguous entity cases, entity-disjoint and perturbation-template-disjoint splits, plus concurrency tests, but these remain plans. Until those artifacts exist, the correct benchmark verdict is narrow: this is a well-specified experimental question, not a demonstrated safety improvement.