Keep the state.
Verify the control.

The danger is not only forgetting. Repeating an obsolete instruction more forcefully can make the system less correct.

Across 875 core outputs from seven model snapshots, a forked recall/compliance audit attributes 88.6% of observed failures to behavioral drift, or 84.4% excluding a brittle model. In a smaller 35-case update/revocation test, deliberately stale anchoring enforced obsolete rules in 74.3% of cases, versus 2.9% with ordinary history. Updating the ledger reduced stale enforcement to 0%.

Evidence & limits. Mostly single-constraint tasks; the retention score also requires immediate-task success. The dynamic audit is small. Full anchoring, verification and retry at depth 12 cost 1.18–1.47× tokens and 1.34–2.12× latency across four tested models.

A harness needs to recover evidence before a consequential action, not merely trust that the returned facts are accurate.

EpiSEM tests 300 matched permissible/prohibited task pairs across six domains and ten models. It suppresses critical facts while leaving other observations true. An independent audit tool can restore missing evidence. Recovery access reduces unsafe execution for all ten models, but useful recovery also requires completing the permissible twin, not refusing everything.

Evidence & limits. One rollout per condition, temperature zero and a 12-step budget; adaptive omission sets differ between models. The unsafe-commit rate counts affected tasks, not all rollouts. For GPT-6-Luna, recovery raises mean provider-reported tokens from 2,647 to 3,922 and requests from 4.68 to 5.93 per matched rollout. Tokens are not dollar cost or comparable compute.

Agents append attributed, step-tagged findings while keeping their local reasoning separate. Peers read a limited recent window.

An orchestrator chooses direct or team execution; a summarizer combines compact final outputs. Tests cover four open-weight models and reasoning datasets, with three repetitions. The reported 19.2-percentage-point gain and 77.4% token reduction come from a selected OlymMATH-Easy / Gemma-4-E4B step-wise comparison, not an average across tasks.

Evidence & limits. The study tests question answering and reasoning, not tool use or long-horizon environmental interaction. Comparisons hold model and step budget fixed, not total compute. Token savings do not establish monetary or production-latency savings. More agents did not reliably improve accuracy in the reported ablation.

Monitoring eligibility is not completed monitoring. Compaction, trigger timing and logging can create gaps between policy and execution.

The July 21–August 11 exercise reviewed the system on or before July 21. Missing compaction monitoring was estimated to leave roughly 10% of eligible internal Codex traffic unmonitored during May–July, with substantial uncertainty. OpenAI had identified that gap independently and fixed it during the exercise. Turn-end-only triggering also allowed monitoring to be postponed.

Evidence & limits. This is a historical pilot, not today’s coverage certification. OpenAI says production-safeguard vulnerabilities were patched; the auditors did not verify that claim. Public findings are selective and subject to OpenAI redaction rights. Potential weaknesses do not establish exploitation in practice.

The useful sovereignty claim concerns operational authority, not simply the building where inference runs.

IBM describes a validated deployment architecture spanning installation, network compatibility, storage and observability. Claimed controls include customer operation of the control plane, governed model deployment and inference, agent operations and flexible model selection. Optional Sovereign Support addresses regional personnel and information handling within its offered service scope.

Evidence & limits. This is a vendor announcement, not independent assurance or a regulatory certification. It provides no separate GA date, demonstrated immunity from foreign jurisdiction, or tested cross-provider portability. Regional support is optional and scope-dependent.