Preserve the evidence.
Own the intervention.

Rather than repeatedly asking an agent to improve itself, give the editor structured evidence about what previous edits actually did.

HGM-Dual combines lineage-aware change records with general and error-targeted candidate generation. On DeepPlanning, authors report travel accuracy of 64.38% versus 56.25% for the strongest listed HGM baseline at 4,000 evaluation cases, and matching baseline endpoints with two to four times fewer evaluations.

Evidence & limits. These are author-reported preprint results. The 120-example evaluation includes the 60 optimization examples, so headline scores are not wholly held out. Only travel and shopping domains are tested, with domain-specific error categorization. Evaluation counts are not total dollar costs.

Before paying for stronger reasoning, check whether the agent can see the information the document author encoded in color and highlighting.

On 76 real policy documents, formatting-aware ingestion reaches micro-F1 0.690, versus 0.150 for Docling and 0.075 for plain text. Scoring uses deterministic matches against OOXML markings. Cue ablation attributes the effect mainly to color and highlighting, not to bold or italics.

Evidence & limits. The task explicitly asks for author-marked requirements, not general document QA. Rendered-image agreement is not human label validation. PDF tests use native text, not OCR; the representation costs about twice as many tokens as plain text. Semantic cue labels do not help every tested model.

A caption alone can lose the detail needed to tell two objects apart or establish how the same object changed over time.

The memory hypergraph links entity mentions, facts, state changes and source observations. Agentic search can traverse these layers and inspect original images. Across three benchmarks, authors report average LLM-judge improvements of 8.20–9.98 percentage points over the strongest baseline.

Evidence & limits. Results cover three backbones and three runs, but measure judged QA, not autonomous end-to-end completion. Token use is roughly doubled; preserving images also brings storage and processing obligations. The cheaper top-k retrieval ablation retains gains without the full iterative-search cost.

A model can identify danger without the organization being able to contain it. The new proposal assigns responsibilities across the actors operating the system.

The blueprint calls for monitoring during training, evaluation and internal use; communicating relevant alerts; escalation and shutdown within each actor’s control; and explicit criteria for safely restarting. It also calls for evaluations of representative agent systems as tools, permissions and environments change.

Evidence & limits. This is a policy proposal, not enacted law, independent certification or proof the proposed safeguards work. It is a new October 21 publication, distinct from the earlier incident disclosures.

Governance needs both an organizational owner and an observable technical control. Neither a policy document nor a model score supplies both.

PwC’s new guidance pairs enterprise responsibilities, escalation and independent challenge with system-specific evaluations, deployment gates, access restrictions, change control and recovery. It recommends generating evidence during operation, including exceptions, interventions and remediation status, rather than reconstructing it only for periodic review.

Evidence & limits. This is professional guidance, not an evaluated deployment or a new legal requirement. The accord discussed in the article is older and voluntary; the October 21 development is the advisory analysis. Model-behavior evaluation and assurance over organizational controls answer different questions.

The meaningful claim is not merely where data sits: customers are meant to control model placement, version and access, with inference inside their own infrastructure.

TrendAI’s announcement describes Nemotron-powered analysis, guided investigation and policy-based response for Vision One Sovereign and Private Cloud. It claims no dependency on external inference services and continued security-team oversight. The existing SPC platform is available; the Nemotron integration is a planned capability, with TITECH a design partner ahead of general availability.

Evidence & limits. These are vendor claims and a preview, not an independent audit or a GA integration. The announcement does not establish key custody, cross-vendor portability, jurisdictional immunity or a tested exit path.