A correct answer can still be missing the part that makes it safe.

011. Evaluate the workflow as an executable artifact

WorkflowArena lets an agent decide whether and how to delegate on 145 long-horizon tasks across eight domains. Every delegation is recorded, execution shares one time budget, and a hidden verifier scores the integrated deliverable. The change from earlier fixed-protocol or offline workflow tests is that the model authors a workflow and then runs it against real task artifacts. [1]

In the authors' fixed-executor transfer experiment, workflows written by Claude Opus 5.5 raised the continuous scores of four weaker execution models by 6.41 points on average in static mode. That is evidence that organization can affect performance, not proof that one model or harness is the best orchestrator.

Each configuration ran once per task, some outputs used an LLM judge, and the harness is confounded with the model in the main comparison.

Decision affected: keep the worker model fixed when comparing orchestration policies. Read Sections 3.3 and 4.5 for the recorded workflow interface and transfer design, then the limitations before using the reported gains.

022. A memory graph can be an index instead of the record

MIRIX-2.0 stores complete typed records in an evidence plane, then uses a provenance hypergraph only as an index. Graph facts cite source rows, retrieval has explicit budgets, and consolidation merges graph anchors outside the write path without rewriting stored memories. A failed graph write does not roll back the evidence, and a failed graph read can fall back to direct retrieval. [2]

The authors report leading four of five compared QA benchmarks and higher mean success than compared memory methods on ALFWorld and WebShop. They also report 3.8 to 4.0 times slower document ingestion than flat memory, and the available latency measurements describe the graph-without-consolidation configuration. Several component effects remain unisolated, and some baselines are author-reported.

Decision affected: if a graph layer is added, keep the original evidence addressable and benchmark ingestion, answer latency, tokens and correction behavior separately. Read Sections 2.2 to 2.5, Section 3.4 and the limitations.

033. Reference agreement is not evidence sufficiency

A controlled memory-evaluation study constructs pairs of legal histories that produce byte-identical visible contexts but different correct answers. This lets the authors separate matching the hidden reference from having enough visible evidence to determine it. In their synthetic device histories, two hosted readers often gave concrete answers when the evidence was certified insufficient. [3]

This is an evaluation diagnosis, not a new memory architecture. The questions are short and synthetic, only two hosted readers are tested, and a deterministic baseline solves the constructed grammar. The useful change is the scoring contract: report sufficient, insufficient and unresolved cases separately, and test the evaluator on executable controls.

Decision affected: add an evidence-sufficiency label and an explicit undetermined outcome to memory tests. Read Sections 7 and 8 for the evaluation consequences and limitations.

044. MCP server safety depends on the implementation version

The official MCP servers 2.4.0 release fixes six vulnerabilities across filesystem, git and fetch servers. File reads now stay attached to the checked file on Linux and macOS, repository-owned Git configuration can no longer launch programs during read-only calls, and fetch rejects additional private IPv6 addresses and normalized paths that bypass robots rules. [4]

This is a shipped security release, not a claim that every MCP server is now safe. Windows retains stated gaps, custom servers are outside the release, and package managers may not offer the semver filesystem fix to users of older date-versioned packages. The release says to install version 2.3.0 explicitly in that case.

Decision affected: inventory the exact server package and version before reviewing scopes or tool prompts. Read the v2.4.0 security-fix table and the package-specific notes, then test the denial and confinement paths in your own environment.

055. A local runtime upgrade can change memory use and API behavior

Ollama 0.40.16 reports that Qwen3.8 Flash Next uses about 27 GiB less memory on Apple Silicon. The same stable release makes the Responses API respect the truncation parameter instead of silently truncating by default, returns an empty model list rather than null, updates llama.cpp and changes Claude Code launch routing. [5]

These are project release notes, not an independent benchmark or a result for MLX-LM, LM Studio or exo. A lower-memory model path and a changed truncation contract affect different decisions. Pin the version and replay the complete request, response and tool-turn fixture before adopting either.

Decision affected: separate capacity, compatibility and quality gates. Read the v0.40.16 release notes and full changelog, then measure peak memory and terminal API behavior on the exact model artifact and machine.

06Use one boundary worksheet

Swipe or scroll to compare

This worksheet is a proposed implementation experiment. No local benchmark, security exploit, memory deployment or production workflow test was performed for this article.