--- title: 'In the News: September 21, 2026 (Morning)' description: 'A local decision model ships with published calibration numbers and a frozen eval you can rerun, so a verification loop can be audited rather than trusted.' canonical_url: 'https://darkfactory.dev/news/2026-09-21-morning' markdown_url: 'https://darkfactory.dev/news/2026-09-21-morning.md' collection: news date_published: '2026-09-21T07:20:00-04:00' date_modified: '2026-09-21T07:20:00-04:00' --- # In the News: September 21, 2026 (Morning) The model that judges your agent's output is usually the least examined part of the loop. Jared Palmer published Kev this morning, a family of small decision models that ship with their accuracy, their calibration and a list of the cases where they are confidently wrong. Below it, Addy Osmani argues that most harness advice assumes a repository nobody has to be afraid of, and sets out what to do in the one you do. ## 1. A decision model that publishes where it is wrong **[Kev: small Jev-like decision models you can train and run yourself](https://github.com/jaredpalmer/kev)** · Jared Palmer · GitHub, surfaced on Hacker News, 2026-09-21 Kev is three models, at 0.8B, 4B and 9B, each a rank-16 LoRA adapter and a pointer head on a Qwen3.5 base. You hand it a piece of text and a set of yes/no, multiple-choice or rating questions, and it returns probabilities rather than a label. The repository publishes accuracy and Brier scores on datasets the models were not trained on: Kev-4B at 0.794 development and 0.832 test accuracy, Kev-9B at 0.812 and 0.837, against 0.857 development accuracy for the hosted Jev service the project benchmarks itself against. It also publishes where that breaks. On the new-source development set, the readme states, Kev-4B "assigns at least 0.9 probability to a wrong answer on 8.2% of questions (Kev-9B: 7.5%)". On records whose deciding evidence had been removed, Kev-9B still answers with at least 0.9 confidence 5% of the time, Jev 9%, and the previous-generation Kev-8B 26%. The evaluation data is frozen with recorded checksums and the benchmark command is in the readme, so those numbers can be rerun rather than taken. Palmer is explicit about what the comparison is not: "We don't know which datasets Jev was trained on, so this isn't a controlled comparison of the two architectures." The repository credits itself as "Built with Devin". The thread stood at 134 points and 58 comments at 07:10 EDT. **Why it matters:** an LLM judge sitting inside a verification loop is a component almost nobody can characterize. This one can be. It reports a Brier score, a per-question-type calibration error, the share of decisions you could automate at a 5% error budget, and its own failure rate at high confidence, all against a frozen suite you can run on your own data. You can ask a judge like that for evidence instead of taking its word, and it fits on a 32GB Mac. The caveat is in the readme too: probabilities are poorly calibrated on sources the model has not seen, so the threshold has to be set against your own labelled examples rather than borrowed from the table. ## 2. Draw the map before the agents start **[Brownfield Agentic Engineering](https://addyo.substack.com/p/brownfield-agentic-engineering)** · Addy Osmani · Elevate, 2026-09-14 Osmani's premise is that in an old system "the repository is no longer a complete description of how the thing actually behaves", and that agents make a change cheap to attempt without making it cheap to verify. His answer is to zone the codebase: green for well-tested, isolated, modern code where agents run in a tight loop; yellow for mixed quality, where characterization tests get written first; red for authentication, billing, permissions and payroll, where a human pairs on every step or the work does not happen. Three rules turn that from a metaphor into a procedure. A person draws the map, because an agent left to choose "starts in the scariest file, because the scariest file has the most interesting names". A zone only moves when the move is earned, meaning characterization tests exist and the module owner has reviewed the agent's first changes. And the zone decides the verbs. Three further rules sit on the same reasoning. Autonomy should track blast radius and recoverability, not model confidence, because "a model's confidence is a poor guide". The session that makes the tests pass should not also be the session that wrote them. And a migration is finished only when the old path is gone, since "if deletion is a future cleanup ticket, the migration unit is not complete". He cites SWE Refactor Bench for the cost of getting that last one wrong: across 520 agent runs, 28 passed its migration audit, behavioral tests and independent verification. **Why it matters:** the harness literature is largely written for repositories that were designed for it. Zoning is a way to get an unattended loop running in the part of the codebase that can absorb one, without pretending the rest is ready. The rule that will earn its keep soonest is the smallest: pin today's behavior in a separate pass, by a person or a different session, before an agent starts changing it. A green suite an agent authored to pass its own work is not evidence. --- ## Also this cycle - **[OpenSpec](https://openspec.dev/)** · Fission · The spec framework's front page names a five-stage loop, `/opsx:explore`, `/opsx:propose`, `/opsx:apply`, `/opsx:verify` and `/opsx:archive`, with verify defined as checking that the implementation matches the spec. Fission claims 69,700 GitHub stars and more than 265,000 developers a month, and lists compatibility with 39 coding agents including Claude Code, Codex, Cursor, Copilot, Gemini CLI and opencode. Those figures are the vendor's own and the page is a product page. The part worth noting is the shape: a named verification stage being distributed as a cross-harness convention rather than written once per team. --- ## Thread watch _Discussions gathering force. No primary read yet, so these are reported as discussions, not as findings._ - **[In September, AI generated code has made up 17.25% of all Linux Kernel patches](https://news.ycombinator.com/item?id=49784366)** · Hacker News · No primary has been read for this one, the submission points at a post on X by the Lunduke Journal, and neither the percentage nor the method behind it is reported here as established. What is reportable is the argument: 27 points against 52 comments at 07:09 EDT, a ratio that usually marks a contested claim rather than an endorsed one. The question underneath it, what share of accepted upstream work in a large open-source project is now agent-written, is one almost nobody has a measured answer to.