The last batch is the one that does not fit anywhere else, and it is mostly about the parts of working with models that are not the model: the record of why a decision was made, the measurement of a decision the model makes, the explanation of how any of it works, and the plumbing that lets an agent reach a tool without a rewrite.
None of these are model improvements, and that is the point. A stronger model does not tell you why a decision was made, does not tell you what a hidden state already contains, and does not tell you which job to apply for. Those are records, measurements and surfaces, and they have to be built separately — usually by people who are not training anything.
That split also explains why so many of these ship a second interface. ADRs for AI, Rhyven and Frontier Roles each publish an MCP server, Decido returns its answers in someone else's wire format, and Search2o exposes agents over a REST API with a GUI alongside it. A directory an agent cannot query is a directory the agent will not read, and the launches that understand this treat the protocol as the primary interface and the web page as the fallback.
Keeping the why
ADRs for AI starts from an observation that is hard to argue with: the most valuable byproduct of working with an AI is the thinking, and in ordinary practice all of it evaporates with the session. It ships a repository starter with ADR and seed templates, the directives, an inbox protocol and task-runner recipes you can adopt whole or feature by feature. A VS Code extension, also on Open VSX, gives tree views of your open questions, diagnostics that catch malformed records as you type, deep links, and an MCP server so your agents can query the record. An iOS client lets you review and one-tap Accept so an answer lands as a real commit.
Decisions you can measure
Decido returns calibrated yes/no, pick-one and rating decisions about a document on a laptop CPU. You hand it a state and typed questions, and it returns a probability for every answer in the same wire format as the Jev System One API, so existing clients need little more than a new base URL. The model is a LoRA fine-tune of Gemma 4 26B-A4B shipped as an 8-bit GGUF for llama.cpp, and install.sh builds llama.cpp, downloads the weights and runs an example.
LexFrame measures something harder: what a hidden state already contains. It is a Python toolkit for interpreting language-model residual streams in their native semantics, reading a hidden state the way the model reads it, in the model's own vocabulary — the state runs through the model's own remaining layers with nothing fitted to the output — and separating what the state carries by itself from what the context adds later, layer by layer and position by position.
Learning how any of this works
MEF LLM Studio is a Windows desktop app that teaches language models by having you build one. It follows the path from tokens through training to generated text with explanations and small experiments, no scripting needed: take your own text through a tokenizer and dataset to a small model, follow a recorded generation in Demo Explorer from context to next token, and explore image generation on the Vision path. There is a browser wiki with Explain for looking up terms before installing anything, and CPU-only starting paths.
Finding the tools and the jobs
Rhyven is a marketplace of headless apps built for agents rather than for people. You connect once through MCP and then add apps your agent can use without a graphical interface; the agent can search the marketplace, ask you to approve an installation, and run the app on your own computer or server, with data staying in your environment. Apps share results inside a collection so another agent can continue the work, and an agent can turn repeated work into an app, test it and save it.
Frontier Roles is a job board for AI developers, listing first-party roles in research, agents, evals and inference, filterable by remote, hybrid or on-site, seniority, country and city. It also publishes an MCP server, so an agent can search the board on your behalf instead of you clicking through filters.
Running agents as a service
Search2o is a stateless, asynchronous Python server that runs agents inside your organization, with a REST API and a bundled GUI. You create an account for a license key, install with pip and run the server with the key in the environment; the agent server and the GUI are developed in separate repositories and there is an accompanying skill. It is in open beta.
What connects these seven is that none of them make a model better. They make the surrounding work — deciding, measuring, explaining, hiring and serving — something you can point at, which is usually the part that was missing.