Meet the model where it writes: a palette either way round, a row separator it can type, several conversions in one call, and a date unit that defaults

Read a table a model wrote on one line, because escaping the escape is what a small model does

Keep the sentence a refused tool wrote, so a finished run says why rather than tool_error

Adjust a ragged row and an unreadable colour rather than throwing the file away, and notice a tool call that ran out of room

Give the grammar one branch per tool, and say out loud when llama.cpp refuses it

Make the first move of a file-producing run a tool call, and stop a model repeating one

Hold the sampler to the action schema, so a small model calls the tool instead of describing it

Measure the GPU offload instead of estimating it, and keep the development scratch out of the repository

Offload what the card can hold, survive a dropped connection, and stop sending prompts the model cannot read

A worker without a readable energy counter refuses to start, and names every backend it tried

Measure GPU and CPU energy from hardware counters and sum them; drop the modelled fallback

Tighten every user-facing string: UK English, no em dashes, cards summarise

Pentest fixes: pin the runtime source host, evict rather than refuse; say on the page that llama.cpp is fetched

Workers print their own code and are claimed from the site; fetch and verify llama.cpp

Harden every component declaration against Gradio's reset; denser library grid