Capability needs
controls.

Dory repeats a shared layer block in latent space. Increasing the loop count spends more computation inside the model without adding an equivalent sequence of visible reasoning tokens.

NVIDIA reports SWE-bench Verified scores of 50.3%, 59.5% and 61.8% for one, two and four loops using the BF16 artifact with NeMo Gym and the Evaluator SDK. Last-loop KV caching avoids keeping a separate cache for every loop.

Evidence & limits. These are vendor-reported results with incomplete full-run configuration, not an apples-to-apples leaderboard comparison. Fewer tokens do not establish lower cost. Serving requires pending vLLM integration or a Dory-enabled llama.cpp build. The Q5 GGUF is a separate artifact; its results cannot be inferred from BF16 scores.

Reflection released Beam weights alongside Mirror, an open-source CLI harness. The company says the model was trained with the harness: a useful reference point for testing model–harness co-design.

Mirror combines a Python frontend with a Rust execution core. It supports parallel worktrees, resumable conversations, skills, MCP, custom providers and self-hosted endpoints. The source is Apache 2.0 licensed.

Evidence & limits. Manual approval is the default, but approval does not sandbox filesystem or network access: tools inherit the user’s authority. Mirror is developmental and native Windows is unsupported. Self-distillation recipes were promised, not shipped in this release.

When an account’s shared quota is exhausted, AgentGrid now skips all models in that account instead of failing through each one.

The release also preserves the lead account when switching harnesses, makes Codex task errors visibly fail rather than disappear, and lets scheduled loops stop themselves on completion.

Evidence & limits. These are vendor release notes; the behavior has not been independently tested here.

An October 16 update to footnote 2 says METR can publish risk assessments without an Anthropic publication veto. Anthropic can still require redactions within predefined categories and ultimately decides what belongs in those categories.

The categories include customer data, attorney–client privileged information and model architectures. METR retains rights to make certain statements about the access and information it received.

Evidence & limits. This is a clarification of disclosed terms in a voluntary arrangement, not evidence that the contract changed October 16. It is neither a published evaluation outcome nor proof that risks are resolved.

This preprint evaluates stored memory facts alongside downstream question answering. FEB-30 contains 30 trajectories across six categories and three difficulty levels.

Across 11 runs of one evaluated extraction, embedding, classification and synthesis system, the mean absolute gap between QA and memory scores was 16.8 percentage points, with a standard deviation of 2.5 points. Answer quality alone did not explain where the memory pipeline failed.

Evidence & limits. Small and largely synthetic, with one evaluated implementation and sensitivity to judges and thresholds. It does not establish results for procedural or episodic memory. Submitted October 15 at 12:58 EDT and announced in Thursday’s cohort; this is an older reading, not an October 16 release.