Skip to content
JKCAHome

Knowledge

The stack, the contract, and the later check.

Each drawing answers one load-bearing question before a capability is claimed. Where a figure is a simulation rather than a measurement, it says so.

Infrastructure

The stack

There is not a research pile and a product pile. Apple Silicon runs the models. Cloudflare Workers run the public tools. Rust owns the contracts between them. The boundary has to be showable: a capability that cannot name where data stops is not claimed.

Five layers from silicon to the evidence a reviewer readsA vertical stack: Apple Silicon, MLX on-device runtime, Rust contracts, Cloudflare Workers, then the evidence layer a reviewer reads.05EvidenceProbes, receipts, and records a reviewer can read04EdgeWorkers, D1, R2, and the publish path03ContractsRust types, request structs, and tests02Local runtimeMLX inference on the machine that holds the data01HardwareApple Silicon unified memory
A layer can only claim what the layer under it can show.

Framework

Why Leptos

One language across the browser and the server contract removes a whole class of drift. Leptos compiles the same Rust types to the host and to WASM. A field the interface can edit has to exist on the request struct; otherwise the edit is theatre and the store silently keeps the previous value.

One request type feeds both compile targetsA shared request type feeds both the SSR host and the WASM hydrate target. The browser cannot invent an editable field the server never declared.TYPERequest structEditable fields live hereSSRHost renderSame types, first HTMLHYDRATEWASM clientSame types, interactivity
One declared type, two compile targets: the browser cannot invent a field the server never accepted.

Authority

How agents should work

An agent can propose work. It cannot approve it, execute an unbound change, or certify the result. The loop that holds is: name the operation, bind it to an exact plan, require a person, run only that plan, and read back what actually happened.

Propose, bind, approve, run, read backFive stations in a line. Only the middle station, approve, is marked as a person. The agent may propose and help draft; it does not authorize or certify.01Proposemachine02Bindmachine03Approveperson04Runmachine05Read backmachine
Approval is the one station that cannot be handed to the machine.

Local inference

Apple Silicon and MLX

Unified memory is why local inference is interesting on a Mac: the weights and the working set share one pool instead of paying a copy tax to a discrete card. MLX is the library when the job should stay on the machine that already holds the data.

Model and working set share one memory poolA Mac box contains model weights and the working set in one unified-memory region. A remote inference box is present but disconnected.DEVICEApple SiliconModel weightsWorking setREMOTENot requiredNo inference hop
Weights and working set share one pool: no copy tax, no inference hop.

Constraints

Stop rules

A model can draft a whole subsystem in an afternoon, which moves the binding constraint from build speed to what can be accepted without a check. Four stop rules carry most of that weight.

Evidence before narrative
A claim that matters operationally has a check, a probe, or a receipt behind it. A confident summary is not evidence, least of all when a model wrote it.
One path
Local shortcuts may exist. They do not redefine how the system is built, published, or rolled back.
Narrow scope
A smaller tool with a named job beats a broader one that hides responsibility in integrations.
Inspect after the fact
If what the system did cannot be reconstructed, the next layer does not ship on top of it.

Simulation

Token speed

Five output rates, run side by side, so the difference is visible instead of quoted. The streams are simulated at fixed rates; they are not model benchmarks.

Five streams start blank, then generate and scroll continuously. Each small book fills and turns pages with the generated words. The comparison runs for 30 seconds.
Download the video (8.2 MB)
Equivalent output per minute
Output rateEquivalent
5 tok/s261 words/min
10 tok/s5.22 reference paragraphs/min
50 tok/s26.09 reference paragraphs/min
120 tok/s25.04 book pages/min
300 tok/s62.61 book pages/min

One reference paragraph is 100 words; one book page is 250 words. These illustrative rates repeat a measured 100-word, 115-token sample using cl100k_base. They are simulations, not model benchmarks.