Keep useful evidence.
Make authority explicit.

Human approval becomes a required protocol transition before a consequential tool is available.

The Python library exposes only legal tools and rejects out-of-state calls before effects occur. In pressured airline tasks, the baseline made unconfirmed changes in 13 of 50 GPT-4.1 tasks and 29 of 128 Haiku task-runs; Concordat recorded none. Ordinary task reward missed many baseline violations.

Evidence & limits. The gate and violation metric share the same explicit-yes definition, so zero violations are guaranteed relative to that definition. Its keyword test can accept ambiguous consent. This small, simulated, two-model study also has quota interruptions and a between-model prompt change. No significant success difference is not proof of equivalence.

Facts can be true and on topic while distorting the balance of evidence.

The diagnostic benchmark holds fact text fixed while changing who owns supporting, decoy and irrelevant facts across 480 tasks. Rare-first selection helps when rare facts happen to support the answer, but is no better than random on average. Relevance filters retain on-topic decoys. Attribution to multiple agents can further bias the final reader.

Evidence & limits. These are controlled, author-reported group-decision experiments. Much of the design uses scripted discussions with a known fact inventory. Main controlled runs use two small open models; hosted checks are exploratory. Removing tags can hurt when shared evidence is useful. These results do not establish that provenance should be removed.

A better average reward can hide a low rate of fully correct outputs.

Across 28 EHR derivation tasks, skills improve partial agreement with reference outputs. On an eight-task control subset, an operational specification explains most of the benefit; the extra skill contribution is indistinguishable from zero. Only 25 of 280 skilled Codex runs pass every strict check, versus 13 of 280 without skills.

Evidence & limits. The arXiv record reports NeurIPS 2026 acceptance; proceedings acceptance was not separately verified. Reference agreement does not establish clinical correctness or readiness. Twenty-six tasks use MIMIC-IV, limiting cross-hospital generality. Clinician review’s causal effect and a unique Markdown-format benefit are not established.

Accumulated experiments can inform a new team structure instead of another round of the same roles.

A supervisor diagnoses plateaus and changes roles or coordination. A learning document, evolution history and team artifacts carry experience forward. The paper reports improvements over coding-agent baselines under matched token budgets. With extended budget, its generated MLA kernel wins on 36 of 47 FlashInfer-Bench workloads.

Evidence & limits. Author-reported, hardware-specific experiments. The headline 1.2× mean speedup uses the stated production baseline; Figure 1 also reports 0.96× against a plan-excluding run-stage baseline. The remaining 11 workloads are slower. Main settings report the better of two runs; additional settings run once. Extended-budget results are not matched-budget gains or universal throughput improvements.

Policy enforcement should remain separate from the model’s reasoning.

NIST proposes distinct verifiable agent identities, narrower authority at each delegation, no standing privilege, short-lived revocable access and runtime checks. Other tenets cover human-review triggers, signed mandates, deterministic access decisions, monitoring and an independent append-only evidence record.

Status & limits. These are initial design guideposts for ongoing NCCoE work, not a binding standard, certification or newly effective legal requirement. NIST explicitly says present technology and standards cannot yet fulfill every tenet across all deployment settings.

The offering combines local models and infrastructure with collaboration-platform controls.

Mattermost describes customer-controlled or open-weight models, on-premises or air-gapped Cohere North, common access policies and audit trails, and customer GPUs. It says v12 is available. Current admin documentation describes per-tool allow, ask and disable policies, agent tool scoping and MCP access rules.

Evidence & limits. Deployment and security claims come from the vendor. Approval is configurable, not universal. Documented audit records omit prompts, tool arguments and results, and persistence depends on server configuration. Those documentation details were checked now, but not dated as new October 27 features. No independent sovereignty or security certification is established here.