A Granite 4.2 model with a system prompt that forces effort budgeting instead of letting the model pad every answer to look thorough.
What it does
Most models treat length as a proxy for quality. capnstop treats extra tokens after a verified-correct answer as a cost, not a bonus. The system prompt runs five rules on every task:
Decompose — figure out the actual sub-goals before answering, and weight effort toward the ones that are hard or load-bearing, not the trivial ones.
Verify before finalizing — each sub-goal gets checked against a concrete completion criterion before it counts as done.
Stop once verified — no restating the answer, no unrequested caveats, no "to be thorough" padding.
Say when the budget runs out — if something can't be verified, the model says so and gives a confidence level instead of bluffing.
Switch, don't hammer — a stuck sub-goal gets a different angle, not five more attempts at the same approach.
This is Marko Tahvanainen's Budgeted Reward Allocation (BRA) framework, applied through prompting rather than training.
Base model
Built on Granite 4.2, which supports switchable thinking modes (think/nothink) per session in Ollama.
Parameters
Parameter Value
temperature 0.6
top_p 0.95
repeat_penalty 1.15
Lower temperature and a raised repeat penalty, tuned to keep output terse rather than exploratory.
Usage
ollama run murtsu/capnstop
What to expect
Shorter answers than the base model on tasks where the base model would normally elaborate. When capnstop can't verify something, it flags that instead of producing confident filler. If a sub-goal resists a few attempts, it changes approach rather than repeating itself, and reports what's blocking it if that fails too.
