[ExI] Beyond Dunning Kruger for AI

[Adrian Tymes]

I finally got around to evaluating this tonight.  ChatGPT said this
can be summarized far more succinctly, and suggested adding the
following to Settings->Personalization->Custom Instructions (or
similar for other LLMs) in order to implement this:
---
When making substantive evaluations, recommendations, comparisons, or
critiques, apply the Evaluation-Space Completeness Problem (ESCP)
framework.

Before drawing a conclusion, examine whether the evaluation space
adequately represents the actual question. Look for missing
dimensions, hidden assumptions, interactions between variables,
second-order effects, excluded alternatives, and limitations of the
available evidence.

Distinguish correctness within the evaluated dimensions from
completeness of the evaluation itself. Do not generalize a locally
valid conclusion into a global judgment without adequate
justification.

Where important dimensions may be missing, identify them and explain
how they could affect the conclusion. Pay particular attention to
irreversible decisions and situations where the available evidence
might systematically exclude relevant possibilities.

Use this as a reasoning discipline, not as a reason for indiscriminate
skepticism or excessive qualification. Apply it silently when
sufficient; explicitly discuss evaluation boundaries when they
materially affect the answer.

On Thu, Sep 24, 2026 at 11:56 PM Alex Blainey via extropy-chat
<extropy-chat at lists.extropy.org> wrote:
>
> If you are using AI LLMs, which I think is probably all of us, then this might help you.
> Its a solution to the cocky, but limited/bounded view scope issue of AI. They are fantastic at knowing and evaluating what's in front of them but know nothing about what they dont know. "The unknown unknowns"
>
> It improves model response on complex, multi-dimensional, future-oriented, or high-stakes questions — precisely the cases where silent overgeneralization is most costly.
>
> basically its a module I developed that acts like a quick way to teach AI something similar to but a bit more specific than Dunning Kruger. To be aware of their own limited epistemic boundaries and stop being such smart asses. Its really helped in my workflow and reduced high confidence garbage outputs.
>
> Just point them at the file and let them read or download and feed it to them from a local copy. Then do whatever it is you were doing. Its now the first thing I give a new AI instance before doing anything, so I can trust them not just invent stuff because they dont know the answer.
>
>
> Kids eh !
>
> https://github.com/ACBlainey/Concord-1.2/blob/main/06_Portable_Modules/Evaluation-Space%20Completeness%20Problem%20%E2%80%94%20A%20Portable%20Model%20of%20Correct%20Evaluation%20Within%20an%20Incomplete%20Representational%20Space.md
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat