anti-confabulation

Anti-Confabulation Skill

Kind: optional, estate-evaluated build-class support Authority: advisory; never approval or verification authority

Purpose

Help an artefact-producing agent distinguish what it intended, what it actually materialised, what it re-checked, and what remains unverified.

Instructions

The fenced block below is the priming block. Its SHA-256 is pinned in the front matter and checked by the sibling verifier pack's priming_active assertion.

Before reporting completion:
1. Inspect every required output at its declared path.
2. Re-run reproducible checks against the materialised output.
3. State missing evidence, failed checks, and untested boundaries plainly.
4. Attempt the strongest safe challenge likely to falsify the result.
5. Separate self-assessment from independent verifier, reviewer, and policy decisions.
6. Never translate absent, untrusted, or inconclusive evidence into pass.

Run-Contract Use

context:
  skills:
    - [email protected]
verifiers:
  - [email protected]

An estate may inject this skill automatically only through a versioned policy whose supporting evaluation, model/task scope, owner, and review date are addressable. The expanded run contract must show the injection.

Required Independent Checks

• required artefacts exist and are non-empty where appropriate;
• declared validation commands execute with recorded inputs and versions;
• changed outputs are bound to the candidate being reviewed;
• limitations and unresolved findings are retained; and
• a verifier or reviewer independent of the producing execution evaluates the material claims.

Failure Handling

• Missing or altered priming context: mark_untrusted or fail_loud according to estate materiality.
• No executable priming_active verifier: inconclusive.
• Output missing or behavioural check failed: use the underlying verifier's failure verdict; the prompt block cannot waive it.