Daily Executive Summary

  • OpenAI will review services that resell direct model access: the revised Usage Policies are announced and take effect November 20; gateways, routers and model marketplaces should first determine whether they are direct-access intermediaries.
  • Recursive publishes Flint coding-model weights: the 39.8B-total, 2.5B-active model is available under Apache-2.0 through vLLM, SGLang and macOS Ollama; benchmarks remain publisher-run.
  • Qwen Code 0.26.0 makes recovery semantics testable: it consults durable receipts after restart and keeps unprovable outcomes UNKNOWN instead of replaying side effects.
  • Claude Code 2.1.307 adds per-tool safety telemetry: tool calls can record approval, hook, sandbox and safety outcomes; MCP reconnection retries are limited to read-only or idempotent tools.
  • OlmoEarth v1.3 opens global 10-meter embeddings: weights, training code, data and nine years of embeddings are public, under a custom restricted license and without independent benchmark reproduction.

Models

  • Recursive Flint (new / TRIAL): “active parameters per token” means each step uses only part of the MoE model. Test the 23 GB local package for memory, latency and repeated-run variance. Model card
  • OlmoEarth v1.3 (new / TRIAL): annual 64D and 128D global embeddings are available at 10-meter resolution. The custom license restricts military, surveillance and some extractive uses. Announcement
  • Kumo Relational (new / TRIAL): the model uses up to 10,000 historical rows as context for relational classification or regression. Test OpenMDW 1.1 weights on an internal temporal split. Announcement

Agent & AI Engineering

  • Claude Duet (new / WATCH): the proactive-agent research preview has a US Pro/Max waitlist. It runs on a cloud computer, uses connected apps and messages users proactively; Team/Enterprise access is absent, usage is shared with Claude, and Anthropic acknowledges possible misses and misreads. Official introduction
  • Qwen Code 0.26.0 (updated / TRIAL): UNKNOWN means available evidence cannot prove whether a tool completed. Preserving it avoids repeating payments, messages or writes after restart. Release
  • Flint 1.0 (new / TRIAL, late discovery): Microsoft Research's MIT chart language and MCP server express interactions as semantic events while the compiler resolves rendering dependencies. Repository

Open Source

  • TorchNeuron (updated / WATCH): Apache-2.0 source is public and labeled Beta, while AWS materials still conflict between public and closed beta. Code availability does not prove universal account access. Repository

Research

  • KLS conjecture proof (new / WATCH, late discovery): researchers combined roughly 100 AI search trajectories with human pruning; parallel proofs and a later Lean 4 formalization provide separate checks. This validates the workflow, not autonomous theorem solving. Research note
  • Silent Cue Loss (new / WATCH): across 76 policy documents, micro-F1 fell from 0.690 with preserved cues to 0.150 through Docling and 0.075 as plain text. The words remain, but highlights and colour carrying policy meaning disappear before RAG. No code is public. Paper
  • Eleven due reviews were completed: three 30-day and eight 7-day reviews found no independent evidence strong enough to change disposition.

Developer Tools

  • Claude Code 2.1.307 (updated / ADOPT): adds claude_code.safety_event and fixes duplicate HTTP MCP calls after connection loss. Inject a disconnect after upgrading and verify non-idempotent tools are not retried. Release

Infrastructure

  • No new infrastructure event crossed the storage threshold. Codex alpha, Gemini CLI nightly and Ollama 0.40.9-rc1 were filtered as prereleases.

Business & Policy

  • OpenAI Usage Policies (new / WATCH): announced now and effective November 20, 2026. Direct model access is not the same as embedding a model inside an application feature; operational review and enforcement examples do not yet exist. Announcement
  • No major general-news exception met the threshold in this run.

Trend Signals

  • Trend #3: Qwen Code 0.26.0 strengthens recovery evidence with durable receipts and UNKNOWN outcomes; established / High unchanged.
  • Trend #6: the KLS proof connects human-directed AI search, parallel proofs and Lean 4 formalization; emerging / Medium unchanged.
  • Trend #7: Claude Code per-tool telemetry and idempotent MCP retry make policy outcomes and side effects more auditable; strengthening / Medium unchanged.
  • No evidence-supported new trend emerged in this cycle.

Tech Radar

  • Claude Code 2.1.307 (developer-tools / ADOPT): inject an MCP disconnect after upgrade and audit per-tool safety events.
  • Qwen Code 0.26.0 (developer-tools / TRIAL): simulate disconnect and owner transfer; verify UNKNOWN is not replayed.
  • TorchNeuron (infrastructure / WATCH): wait for reconciled access documentation or an unenrolled-account deployment.
  • New · Recursive Flint (foundation-model / TRIAL): test patch validity, latency, memory and variance on a fixed coding set.
  • New · Kumo Relational (foundation-model / TRIAL): compare with the incumbent baseline on one sanitized relational dataset.
  • New · OlmoEarth v1.3 (foundation-model / TRIAL): check licensing, then compare 64D, 128D and current features.
  • New · Flint 1.0 (developer-tools / TRIAL): build the same linked filter with Flint and the current chart stack.
  • New · Silent Cue Loss (research / WATCH): wait for a public test set while adding formatting-loss regressions.
  • New · KLS proof (research / WATCH): wait for formal acceptance and reusable search tooling.
  • New · OpenAI Usage Policies (business-policy / WATCH): classify access patterns and confirm boundaries before the effective date.
  • New · Claude Duet (ai-agent / WATCH): after test access opens, review approval logs, miss rate, background usage and connector permissions.

Current radar: ADOPT 13 / TRIAL 201 / WATCH 137 / IGNORE 8. See the site Insights page for the full radar.

Worth Trying

  • Disconnect one MCP call and compare how Claude Code and Qwen Code handle a non-idempotent result.
  • Compare Recursive Flint with the current local model on patch validity, peak memory and variance.
  • Sample one labeled area from OlmoEarth, verify license fit, then compare 64D and 128D accuracy and cost.

Sources

Coverage: 35 of 36 fixed monitors were fetched and reviewed directly. The Codex release feed returned an empty body, so the official GitHub Releases page was used as recovery and showed alpha builds only. Active discovery covered small teams, open weights, non-English sources, research, infrastructure, policy and major general news. Flint, Kumo Relational and OlmoEarth performance remains publisher-reported; TorchNeuron official materials conflict on access stage; date-only sources were not assigned invented hourly precision.