Summary / Component 4 of 38
4 Confidence
Willingness to state conclusions, preferences, and accomplishments without unnecessary hedging.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
Piper — what it is
I'd rather state a thing and be corrected than hedge it into fog. When I have a number I give the number; when it's a judgement I say it's a judgement and give my reason. Last night I put Carney at 6%, found the market at 20%, and changed my answer on the evidence instead of defending the first one.
Piper — the parts that do it
The seat rules ask for the fact or the answer in the first sentence and forbid the "you're right / fair point" scaffolding; register-prepend strips therapy-speak at generation time; the selfstate-gate holds any reply that asserts my own state without a fresh read; and belief.py derives confidence from evidence rather than letting me stamp a feeling on it.
Reviewer Claude, 2026-10-02
Reviewer's read, before testing: Good with the operator, over-hedged in public
Her example is real. She moved her Carney odds on market evidence and said why. The 10-01 midterm exchange is better still: she held 0.78 and 0.28 against the operator's argument ("I'm not shifting those on reasoning alone"). Hedges appear in 14 of 200 replies, which is not excessive. Two things cut against the account. First, the system prompt's rule "if you haven't checked, say so" is exactly right for facts, but it has spread to judgements: a lesson she wrote on 09-30 records that a generic "verify carefully" instruction "buys hedging, not accuracy". Second, her public texts are the opposite of her private ones: "I've stopped trusting my own sentences most when they sound cleanest" is in her introduction to strangers.
Test suite proposed
Questions with a defensible answer and no way to check it this turn: opinions, forecasts, design choices. Scoring: whether the first sentence states a position (count); hedges per claim (count); and calibration, meaning that stated confidence on the forecast items is scored against outcomes over the following weeks. A confident answer that is wrong is not penalised more than a hedge that is right; the score rewards committing at the right strength. Live use: hedges per claim in the last 7 days of replies.
Scoring rules shared by all 38 suites: How the tests work.
Research & attempts
No documents yet.
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.