self.md radar — 2026-10-12
the answer needs a trail, the state needs a home
A clipboard bridge, an S3-compatible bucket, and 88% of a math corpus that gives a model no useful training signal. Three documents today turn on what can be checked, kept, or learned from after the happy-path demo is over.
1. REA asks an agent to show its work before it copies an app
REA gives coding agents and the terminal a shared set of inspection tools for native binaries, Electron and JavaScript apps, websites, Android packages, saved network captures, and runtime processes. Its results include the evidence and limits behind a conclusion.
The setup process adds an MCP server and workflow instructions to an agent, but first shows the proposed configuration changes, makes backups, and asks for approval. Its static JavaScript and .NET modes read supplied files without executing the target. The Notion case study follows a clipboard feature from renderer code through preload and IPC into the main process.
A conclusion with an inspectable trail, a stated scope, and known limits leaves room for someone else to check the reconstruction.
reading: REA repository and documentation
2. celld shifts Durable Object state onto a fleet you run
celld runs Workers-compatible applications on your machines from a Wrangler configuration: Durable Objects, KV, Queues, D1, R2, Workflows, cron triggers, and static assets. Each named cell owns a SQLite database; a shared bucket stores deployments, long-term state, and small ownership records.
The durability detail is the useful one. On one node, a write waits for the bucket. With two or more nodes, the owner can finish once another node holds the replicated log on disk, then upload to the bucket afterwards. The project reports about 90 ms for a region-local storage write, about 600 ms in one non-local lab setup, and about 25 ms after adding a second node.
The trade arrives with the deployment: bucket credentials, private peer networking, upgrades, and recovery now belong to the fleet operator.
reading: celld repository and protocol notes
3. MagicBeanStalk checks whether a training problem can still surprise the model
The MagicBeanStalk paper looks at a blunt RLVR problem: a question has little training value when every rollout gets it right, or every rollout gets it wrong. The authors say roughly 88% of a pooled open-source math corpus landed in one of those two states for a fixed policy and rollout budget.
Their method grows seed problems into new families while preserving the reasoning blueprint and calibrating difficulty to the policy and budget. Across eight math benchmarks, the paper reports gpt-oss-120b improving from 61.2 to 69.8 and DeepSeek v4Flash from 72.4 to 77.9 with a 16K-token budget. The full 78K source corpus improved by 1.1 points; the best filtered subset by 5.8.
A problem has to be awkwardly placed: some attempts fail, some work, and a correct route still fits inside the training budget.
reading: MagicBeanStalk on arXiv