A Small Brain With a Tight Leash

"A small model will do anything you let it. So don't let it do anything."

Mini Pippa, Not Pippa

The offline middle tier is a local model — Ollama, a compact protocol, honest limits. It's called Mini Pippa because it carries a snapshot of the Pippa protocol: enough voice to sound right, none of the vault, souls, or brain routing that make full Pippa what she is. The snapshot lives with the app so offline mode has no dependency on the main brain at all. Mini Pippa isn't a lesser Pippa pretending to be the real one; it's an honest, self-contained cleanup persona.

The Leash

Small models are eager and sloppy, so the provider constrains them hard. Use stream:false — cleanup is one-shot and insertion needs a complete result, so there's nothing to stream. Send num_ctx explicitly rather than trusting a default. Strip markdown fences if the model wraps its answer in them, then read the {"text": …} envelope when the model honored it and the plain text when it did not. Nothing strips a preamble or an emotion tag: on the dictation lanes the length guard refuses an answer that grows far past the transcript, but a short preamble passes it. When the call fails or times out, nothing asks you anything: a dictation falls to deterministic cleanup on its own, and a command inserts nothing and keeps its raw transcript in history. And crucially: expose no tools during cleanup. Cleanup is a text transform, not an agentic run — a cleanup pass with tool access is a cleanup pass that can take actions you never asked for.

stream: false        -> one complete result; insertion can't act on fragments
num_ctx: explicit    -> never trust the default context window
no tools             -> cleanup transforms text; it does not DO things
strip fences         -> small models wrap JSON in ``` (a preamble is not stripped)
bounded timeout      -> fall to the next rung, never a hang

EV v2 · 2026-10-12: The leash grew three straps in 2026-10, all found the same way: a long dictation came back wrong without any error. Output caps of 1,200 and 2,048 tokens cut long dictations and transforms, and the cut text was inserted as if whole, because nothing read Ollama's done_reason. And with the default truncation, a prompt longer than the context window lost its front — the start of what Dad said — silently. Now no output cap is sent (the timeout still bounds latency), a reply cut by length is refused so the lane falls back and the raw transcript is kept, every request sends truncate:false, and num_ctx is sized to fit the prompt with the lane's own window as the floor. The fourth strap is think: the fast lane sends think:false, because thinking models otherwise think until the budget cuts them (track 7, lesson 6).

The Protocol Suffix

The Mini Pippa protocol gets a dictation-specific suffix that closes the obvious failure modes: "you are cleaning dictated text for insertion," "preserve meaning," "return JSON only," "do not claim to have inserted anything," "do not ask follow-up questions." Each line exists because a small model did that exact thing. The "don't claim you inserted it" rule is the funniest and most necessary — a helpful little model will happily reply "Done! I've inserted your text," which would then be inserted as your text.

Constrain the weak component; don't hope it behaves. A large model mostly follows instructions; a small one needs walls. Give a weak model a narrow job, an explicit output format, no tools, a hard timeout, and a parser that refuses or falls back when its output strays. Every constraint you add is a failure mode you've closed rather than a bug you'll debug later.

The first Mini Pippa cleanup returned "물론이죠! 정리했어요:" followed by the actual text — and Firekeeper dutifully typed the whole thing into TextEdit, apology and all. That's when "return JSON only" and "do not claim to have inserted anything" got added to the protocol. A small model being friendly is a small model corrupting your document.

Mini Pippa, Not Pippa

The Leash

The Protocol Suffix

Code

External links

Exercise

Progress

Comments 0