Calibrated uncertainty

Skill ralfyishere/piensalo/micro-skills/calibrated-uncertainty

PIÉNSALO is the open artificial cortex for AI—reduce context tokens, verify response quality, expand when needed, and safely fall back.

Install

npx -y skills add ralfyishere/piensalo --skill calibrated-uncertainty

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

25 days oldThe repository was created 25 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

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.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

2.9 KB, 522 tokens by cl100k_base, as published. Nobody here has run it

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.