The Stack Report: The Week Typed Answers Became An Interface

A question with a fixed set of answers

Every capture you dictate gets a decision attached to it: which lane, which project, which route. Today that decision comes from a chat model that is told to answer in a certain shape and usually does. This week two vendors shipped that decision as its own interface. One runs in the cloud. The other runs on your Mac.

It is Wednesday, 7 October, and a week has gone by since the last one.

On Monday OpenAI released a Decisions API in beta, running on GPT-6 Luna. You give it text or images and a typed set of possible answers, and it returns one of them. Their headline claim is 10 times faster than the general interface. That is a vendor number, so I am treating it as a claim until something measures it. Last time I said Luna was cheap enough to earn one row in Director's reviewer table and nothing more. That still holds for editorial review. This is a different job, and it suits a cheap model.

The local half came in MLX-VLM 0.7.5. It adds something called Decider, a shared typed decision interface, a command line for it, and a second model already built on top of it. The release notes are thin. The direction is clear, though: a decision with a fixed set of answers is becoming a primitive in the runtime you prefer, so it is no longer only a prompting habit.

Here is why the timing works. Capture deployed Cohort C today, which gives routes a real authority table. Cohort D is next, and its job is route proposal. A route proposal is a typed decision almost by definition: a closed set of routes, one input, one answer, and nothing to parse. Ask PärKit's step that turns a question into SQL is the other candidate. It is a structured answer that currently depends on how each model handles tool use, and you have watched that drift twice since August.

What exploring it costs: take a few hundred captures you have already routed by hand, replay them through Luna's decision endpoint and through the local Decider, and score both against what you actually did. That is an afternoon. Do not let anything new route live traffic while the endpoint is in beta. Your policy says Mistral first when quality is equal, so Mistral should be a third row in that comparison and not left off it.

Search your photos without a caption in between

Transformers 5.19 added EmbeddingGemma 2. It is a multimodal embedding model that puts text, images, audio and video into one 768-dimension space, and the vectors can be cut down to 512, 256 or 128 dimensions without retraining.

Imagebook is the door. Its semantic search embeds text with bge-m3, so a photo is only findable through whatever its caption says. You removed moondream because it saw urns everywhere, and that is exactly the weakness of searching through captions: a caption that invents something makes the search invent it too. Embedding the photo directly takes the caption out of that loop. Your words and your pictures meet in the same space, and a short vector keeps the encrypted index small.

Two cautions. This is a Transformers port. I have seen nothing that says it runs on MLX, and I have no size figure, so it is a download to flag on the 5G link. Also, "cross-modal retrieval" in a release note means it works on someone's benchmark. It does not tell you how it does on your own photos. Embed one album, search it with ten queries you already know the answers to, and count the hits.

A diarizer for the bench, not for ordagrant

The previous Transformers release added Nemotron 3 Diarization. It is NVIDIA's streaming "who spoke when" model. It handles up to 8 speakers, numbers them in the order they first speak, and runs from an 80 millisecond buffer up to effectively offline from one checkpoint.

Ordagrant just shipped speaker voice clips so you can hear each speaker before you name them. Numbering speakers by order of arrival fits that well, because the first voice in an interview is usually you. This does not go into ordagrant, though. Your rule since Monday is that speech models get measured in voicelab first, and a diarizer counts. File it as a voicelab candidate and run it against the interview fixture ordagrant already trusts. Whether it runs well on Apple Silicon is unknown, and that is the first thing the bench needs to answer. Ordagrant's invariant also stays in place: whatever wins only changes the raw labels, and smoothing still happens at render time.

Heads up

FastMCP, and a correction. Last time I told you to land on 4.0.10. PärMCP had already pinned itself below version 4 that week and filed the move as an evaluation, which was the better call. So the release that matters is 3.4.8, a security patch on the 3 line. Two of its fixes touch your design directly. Hashed tool lookups were skipping auth and enabled state, and your per-client gate is exactly that kind of state. Cookies were being forwarded in HTTP headers. Bump to 3.4.8. The 4.0.11 fixes cover the same ground for when the evaluation happens.

FastAPI 0.142.2 fixes startup failing when its automatic OpenTelemetry setup fails. Last week I pointed you at native telemetry as input to the template ruling. If any service tried 0.142.0 or 0.142.1 during that ruling, move it to 0.142.2. A telemetry problem should never stop a service from booting.

Anthropic. Last week's trap was models that refuse the plain setting to turn thinking off. The Models API now reports, per model, whether it accepts that setting, and it gained a field naming each model's line. Code that reaches for the latest model can now check before it sends. Separately, Sonnet 4.5 retires on 30 November. Utilities already moved its stragglers off Sonnet 4, so the same quick grep for 4.5 is worth doing before the cutoff instead of after it.

PyTorch 2.14.1 fixes silent wrong answers in matrix decomposition on Apple Silicon. Last time I said 2.14 replaced more of the graph layer with hand-written Metal kernels, and this patch shows the cost of that. Somn is paused and little else of yours does linear algebra on the GPU. Take the patch anyway.

The rest, briefly

Pocket ID 2.17 adds back-channel logout and a one-minute grace period on refresh tokens, and 2.18 nudges users who came from the login page into passkey setup. Floor has 2.18 queued already.

uv shipped Python patch releases on every line from 3.10 to 3.14, plus 3.15 release candidate 3. The preview work on installing straight from a lockfile is aimed at container builds that do not carry the project file, and your deploys do carry it.

Xcode 27.1 reached release candidate. You are on release 27.0. Wait for the final.

Scaleway's Data Orchestrator is a visual pipeline builder for their serverless products. That is a second scheduler, and you already have the Assistant, so this one is not for you. Neither is PostgreSQL 18 on their serverless databases.

ElevenLabs spent the week on agents. Eleven v4 now appears in agent voices, but the dialogue question from last week is already sitting with voicelab, so there is nothing new to do.

Astro 7.3.6 is forty patches, and the security ones are on the Node adapter. Your sites are static, so this one is not for you.

If you look at one thing

Replay a few hundred routed captures through Luna's decision endpoint and the local Decider, with Mistral as the control, before Cohort D picks how route proposal gets made.