.plan-26-45: Welcome to the House of Fun A self-hosted Qwen3.8 Flash model running on a Mac Studio passed the IUCN Red List assessor exam with an 81% score, above the 75% pass mark, according to a new preprint by Shane Weisz cited in Anil Madhavapeddy's .plan-26-45 weeknotes. Frontier models cleared the same 75% threshold at 96%, and the local result opens offline field uses for sensitive species data. The same notes also covered Tessera winning the OpenUK open source software project of the year award at the House of Commons and David Allsopp's OxCaml 5.4.0+ox9 release, which adds improved unboxed values, atomic fields, safer effects, and experimental layout and mode polymorphism. What a fun week Tessera https://anil.recoil.org/projects/tessera won the open source software project of the year https://anil.recoil.org/notes/openuk-tessera-award at the OpenUK Awards in the House of Commons. I was very inspired to hear from Enrica Porcari of CERN who mentioned that we may be able to get tours inside the now-shut-down LHC, which I will drop anything to attend Congratulations also to Shane Weisz https://shaneweisz.com on his new preprint on how local AI agents can now pass the IUCN Red List assessor exam https://anil.recoil.org/notes/ai-red-list-exam . The headline result is that while frontier models comfortably clear the 75% pass mark at 96%, a self-hosted Qwen3.8 Flash on the Mac Studio on his desk also passes with 81%. This opens up all sorts of offline field uses for sensitive species data. It's also a wrap for my 1A Foundations of Computer Science https://anil.recoil.org/notes/focs lecture course for this year. I very much enjoyed handing over to Jon Sterling https://jonmsterling.com , whose Discrete Maths course comes right after mine. Jon's also now lecturing slide-free https://www.jonmsterling.com/2026-W45/ and using the document camera to write down maths by hand. I'd like https://amok.recoil.org/@avsm/117586425001021630 to head in this direction next year and do more explanations by writing them down when teaching the basics. After David Allsopp https://www.dra27.uk released a new revision of OxCaml https://anil.recoil.org/projects/oxcaml , I wrote up my analysis of what changed in this OxCaml 5.4.0+ox9 release https://anil.recoil.org/notes/oxcaml-ox9-improvements . They include many performance-oriented features such as improved unboxed values, atomic fields, safer effects, and experimental layout and mode polymorphism. My personal monorepo https://tangled.org/anil.recoil.org/ox is now pinned to ox9, and I also pushed a working odoc 3.2.2+ox to the OxCaml opam repository https://github.com/oxcaml/opam-repository/pull/67 with a matching odoc PR https://github.com/ocaml/odoc/pull/1520 . While messing around with deploying my website, I also got sufficiently pissed off https://x.com/avsm/status/2119425506047398099 by my household Netgear routers that were still running a 2018 Linux that's vulnerable to attack https://thehackernews.com/2026/10/linux-kernel-tcp-flaw-with-98-cvss.html . So I also built a custom OpenWrt image with Linux 7.4-rc1, and bundled an OxCaml webserver with a Cap'n Proto endpoint running on the router itself. Wifi performance is now noticeably better on the new kernel it is the latest one after all , and CLI management beats the vendor web UI by a mile. I also fixed even more sandbox escapes in various Eio backends, with the first one being a really surprising one reported by Thomas Gazagnaire https://github.com/samoht . It involves a trailing slash on a symlink https://github.com/ocaml-multicore/eio/pull/967 . A path like foo/ should only ever refer to a directory, but the Eio POSIX backend treated the last component like a symlink it could follow. A link pointing outside the sandbox could therefore escape it by having a slash on the end. Argh The fix treats foo/ as foo/. , which forces foo to resolve as a directory. I also opened up a Windows PR with symlink and junction support https://github.com/ocaml-multicore/eio/pull/968 . I'm finding Windows filesystems extraordinarily complex to navigate, since they allow directory hardlinks 'junctions' and how these interact with the relatively new symlink support is the stuff of nightmares. I also ran security scans over uring and opened up more hardening for the OCaml bindings https://github.com/ocaml-multicore/ocaml-uring/pull/167 , which cover a number of edge cases. Akshay Oppiliappan https://oppi.li/ did get the big v2 upgrade of Tangled live, which meant I couldn't resist having a go at an OxCaml spindle so we can experiment with distributed workflows on ATProto at last My first cut of a custom Tangled https://tangled.org CI spindle is now up and running https://bsky.app/profile/anil.recoil.org/post/3mzsv34fb422u . It uses Andrew Nesbitt https://nesbitt.io 's brief https://github.com/git-pkgs/brief project detector to probe the languages and toolchains in a repository and route it to the right workflow. This lets repositories get CI without writing a pipeline file, using the per-language metadata and guessing the right commands to run. There's OCaml and OxCaml support first of course, and I'm messing around with Python support via uv and pytest as the second route e.g. for my Evidence TAP parsosaur https://tangled.org/anil.recoil.org/parsosaur library which I'll talk more about later . Behind the scenes of the spindle OCaml logic is a set of libraries to handle opam solving, and then it hands it off to our existing OCluster https://github.com/ocurrent/ocluster Cap'n Proto submission system, so this interops really well with the existing GitHub infrastructure too and doesn't require duplicating all the runners. I'll write up more on this in a separate post next week, as I'd also like to get Michael Dales https://mynameismwd.org up and running with a LIFE pipeline to validate the use of this for larger jobs. The UK label map https://anil.recoil.org/notes/2026w44 I started last week has now progressed to include 400+ data sources, all held in a single DuckDB https://duckdb.org spatial index, and a tile server for raster tiles. The hard part was harmonisation of the licences, as several datasets need clickthroughs; right now I'm keeping it private but tracking all these individually in each table. While I'm assembling all these, I was also delighted to meet Tom Loosemore https://loosemore.com/ in Cambridge after many years, for an invigorating plotting session. Tom has been an inspiration to me since Up My Street https://upmystreet.org/about back in 1998, and his pioneering work establishing the UK Government Digital Service. Tom's recently https://loosemore.com been exploring how personal AI agents could be useful for normal people, like his bin-collection site https://loosemore.com/2026/10/09/how-agents-show-up-what-muse-and-other-agents-reveal-when-they-use-whenisbins/ and what data they leak about you He's also been conducting user research https://loosemore.com/2026/11/09/user-research-with-ai-agents-to-disprove-a-hypothesis/ using agents, which struck me as a similar approach to the DeepMind Concordia https://github.com/google-deepmind/concordia simulations from our ARIA ecosystem resilience workshop https://anil.recoil.org/notes/foundational-ecosystem-workshop last year. He's also encouraging those developing public services to prepare for the looming impact when agents arrive en masse https://loosemore.com/2026/10/26/here-be-dragons/ : Those running public services should be considering what they can do to shape how agents engage with their services. Do not wait for them to arrive en mass. You may dismiss my anxieties as premature, since there are minuscule numbers of people using AI agents in the UK at the moment. Remember Hemingway’s quotation on being asked how he went bankrupt: “Two ways. Gradually, then suddenly.” -- Here be dragons, Tom Loosemore, 2026 https://loosemore.com/2026/10/26/here-be-dragons/ I've been continuing to develop the latest version of the Taposaur website for internal use, as a growing number of colleagues such as Andrea Manica https://www.zoo.cam.ac.uk/people/andrea-manica and Neil Burgess from UNEP-WCMC are querying our metadata for various research tasks. Now that we have around 30m fulltexts, the next phase is to finish mirroring OpenAlex's PDF connection ~80m and doing a run through Parsosaur https://tangled.org/anil.recoil.org/parsosaur to get reliable TEI XML metadata. I'll talk more about this stack in the coming weeks, as we're running it over our classic Conservation Evidence https://anil.recoil.org/projects/ce butterflies and moths review corpus. I got into the closed beta for Rex https://superlogical.com from Superlogical last week and it has totally become my default terminal. It's a terminal multiplexer that instantly resumes connections. Its notification system tells me when a remote pty is waiting for input or has finished , and sessions can be named and grouped logically. A cool side effect is that my coding agents now rename their own Rex session to describe the task they're working on, so the sidebar is a handy project todo list. I'm regularly SSHed into over 10 servers these days so the instant resume is also very cool I spent the weekend hacking on the Android Tessera labelling app https://anil.recoil.org/notes/2026w43 for Julia P.G. Jones https://www.bangor.ac.uk/staff/sens/julia-patricia-gordon-jones-010356/en , and she now has it working on her phone for her Madagascar field trip. She is currently en route to the rainforests with Miranda Lam https://www.linkedin.com/in/miranda-lam-a088561b4 to try it out against real habitats Finley Stirk's weeknotes now have an Atom feed https://finleystirk.github.io/weeknotes/feed.xml Next week: I'm giving an IUCN webinar on Tessera with David Coomes https://coomeslab.org , so come along if you're free