The Mould Loft

Instrument 03 · runs in your browser

The wardroom

The wardroom is where a ship’s specialists sit. You do not go looking for the navigator, the surgeon and the gunner in three different places — you go to one room and the right one stands up.

Finding F2 — this instrument had to be rebuilt to work at all

wardroom.html routes a task by POSTing to api.anthropic.com from the reader’s browser. The request carries no x-api-key and no anthropic-version header, and the endpoint does not allow a browser origin without an explicit opt-in header.

Read the fetch call in the page source. Every click lands in the catch branch, which prints “That did not come back cleanly. Try again with a shorter description.” — wording that reads as a transient fault rather than a missing capability.

What changed. Routing this task does not need a model. The rules that decide it are already written down in the page: a sensitivity gate, five bench entries with a stated speciality, ten assistants with a stated host. Those rules run in the browser with no key and no network. The version below calls nothing and sends nothing. Open the network tab and route a task: there is no request.

Route a task

Try one:

1 reconcile prices · 2 make icons · 3 look up a holding · 4 hours of recordings · 5 intake records · 6 two tasks in one sentence

The bench

What each is actually for, and where it runs. The distinction that matters is the second one — cloud models see your data, local ones do not.

ClaudeCLOUD
Long documents, careful reasoning, code, anything where being wrong quietly is the real risk.
Use it as: Sitting in the writing seat and in the build loop.
ChatGPT / GPTCLOUD
Broad general work, image generation, the widest add-on ecosystem, voice.
Use it as: General-purpose second opinion and multimodal work.
GeminiCLOUD
Very long context, video and audio input, tight integration with Google files and mail.
Use it as: Anything where the source material is a pile of Drive documents or hours of recording.
PerplexityCLOUD
Search-first questions where you need current sources with citations attached, not prose.
Use it as: The look-it-up seat. It answers with links; the others answer with paragraphs.
Local model (Llama-class)ON YOUR MACHINE
Anything that must not leave the machine. Weaker than the frontier models, and that is the trade.
Use it as: Restricted material. Not a general workhorse — a compliance capability.

The assistants

Seeded from the ones on record. The router matches a task against the host and the job, and says “direct — no in-house assistant matched the wording” rather than guessing.

ARIEL
The Whiteboard — dean / tutor guide
AF GFly
Pro Box — profession box engine
FOOLSCAP
The Shrine — manuscript companion
Lamplight
Sophia project — companion
Kindler
Sophia project — companion
Epinoia
Sophia project — companion
ENTRY bots ×5
ENTRY — local concept bots
Portfolio assistant
FAFO — primed with full roster + gates
Queue assistant
Queue console — research assignment
BB.assistant
BB2G CORE — the standing AI panel

Where the money should sit

Three different kinds of spend get conflated into one number. They behave differently and only one of them scales with what you build.

Seats — for you

Chat subscriptions you personally sit in front of. Flat cost. Does not grow when your apps get users. Two good seats beats six mediocre ones — you cannot be in six windows at once.

API — for the apps

What your shipped apps consume when someone uses them. Metered by tokens, not by seat. This is the bucket that should grow, and it can be recovered in licensing rather than absorbed.

Local — for what can’t leave

Hardware plus setup, then effectively free per query. Only worth it if you actually hold material that cannot lawfully go to a cloud provider — but if you do, nothing else substitutes.

The gate is harder to talk out of than the dropdown. Her page routes on the sensitivity setting alone. This one also reads the task: describe intake records, a case file, a minor’s record, an export-controlled part or a privileged document and it goes local whatever the dropdown says — and it tells you it overrode you. Try option 5 with sensitivity set to Public.