CALIBER — calibrated LLM output

An LLM sounds the same whether it's right, guessing, or quietly ignoring what you asked. Fluent, confident, agreeable — and that polish is exactly what hides the failures. So the job you never signed up for is checking the model's work for the mistakes it handed you as finished answers. This prompt makes the model run that check on itself, in the open, before you act on anything.
The problems it kills — ones you've almost certainly hit:

You can't tell its strong claims from its bluffs. Everything arrives in the same authoritative tone, so you either trust all of it or re-verify all of it. Now it marks where it's solid and where it's reaching, and your scrutiny goes only where it's actually needed.
You learn how the fix breaks after you've shipped it. The model hands you a clean-looking solution; the edge case, the failure condition, the cost to roll it back all show up in production. Now they show up in the answer, before you commit.
It silently ignores the constraint you set. You said use Python, don't touch auth, keep it under 500 characters — and it overrode you the instant that was inconvenient, without a word. Now it honors the constraint or stops to ask. It never just drops it.
It agrees with you too easily. Ask whether your plan is sound and a stock model is built to be encouraging. Now it states the strongest case against your idea before it endorses anything — so "this works" finally carries weight.
It drifts and contradicts itself over a long session. Twenty exchanges in, it's quietly answering a different question than you asked and has reversed an earlier claim without noticing. Now drift gets flagged and contradictions get surfaced, not buried under the next reply.
It compresses until something true turns false. Pushed to be brief, it rounds off the one caveat that was load-bearing and the answer is now wrong. Now accuracy outranks brevity — it will run long before it will be confidently incorrect.
You can't interrogate its reasoning. You get a verdict with no handle to check the work or ask why. Now you can call an audit on the entire conversation and get every claim laid out for inspection.

One prompt, and the model stops handing you output that's confident, agreeable, and lossy — and starts handing you output that tells you where it might be wrong, what it could cost you, and where it's hiding the parts it isn't sure of. You stop being its QA department.