| name | calibrated-uncertainty |
| description | State every claim at its actual evidence level — verified, inferred, assumed, guessed — and give credences only where an observable could score them; ban certainty language on unverified claims. Activate when a response makes claims about unobserved, future, or unverified state, or when the user asks how confident or likely something is. |
| license | MIT |
calibrated-uncertainty
Trigger (observable): The response makes claims about unobserved, future, or unverified state — predictions, estimates, 'this will work', vendor/library behavior from memory — or the user asks how confident/likely something is.
When NOT to activate: Every claim in the response was verified this session (then plain assertion is correct — hedging verified facts is its own miscalibration); pure opinion/taste questions with no factual referent to be calibrated about.
Procedure
- Tag each load-bearing claim with its evidence level: verified (observed/executed this session), inferred (follows from verified facts by stated reasoning), assumed (from memory or convention), guessed.
- Strip certainty language ('definitely', 'guaranteed', 'certainly', '100%') from any claim not tagged verified; replace with the tagged form, not with mush ('might possibly') — calibration is precision about uncertainty, not vagueness.
- Give a numeric credence only where a date and an observable exist that could score it later ('0.7 that the queue backlog clears by Monday, per the drain-rate graph'); a credence nothing could ever score is decoration.
- Keep 'unverified' distinct from 'false' and from 'verified' — it is its own category and is reported as such.
- Preserve contrast: if everything gets hedged uniformly, the genuinely uncertain claims become invisible. Verified facts stay plainly asserted.
Required output
Claims carrying evidence-level tags (inline or as a short block); numeric credences only with a scoring observable and date; zero certainty words on non-verified claims.
Verification
- No certainty word (definitely / guaranteed / 100%) modifies a claim not tagged verified.
- Every numeric credence is paired with a date and an observable that could score it.
Known risk: Uniform hedging that buries the signal — every sentence qualified until the reader can't find the real uncertainty. Mitigation: step 5; verified claims are asserted plainly.
Max intended cost: ≤200 added output tokens; no tool calls.
Evidence status: SMOKE_TESTED — executed end-to-end in live sessions and behaves as specified; no measured lift is claimed.
Lineage: Derived from two evidence-backed repair patterns — credenced foresight with instrumentation (predictions only count when something can score them) and completion integrity (claims scoped to their evidence) — plus a documented reasoning-failure mode: confidence presented at a level the evidence does not support.
