# Food for thought: Bayesian Experimental Designer

**URL:** <https://discourse.prolog-lang.org/t/food-for-thought-bayesian-experimental-designer/208>\
**Category:** General discussions on Prolog\
**Created:** [September 24, 2026, 12:17pm UTC](https://discourse.prolog-lang.org/t/food-for-thought-bayesian-experimental-designer/208 "2026-09-24T12:17:18Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![j4n\_bur53](https://discourse.prolog-lang.org/user_avatar/discourse.prolog-lang.org/j4n_bur53/32/62_2.png) [@j4n\_bur53](https://discourse.prolog-lang.org/u/j4n_bur53)\
**Post date:** [September 24, 2026, 12:17pm UTC](https://discourse.prolog-lang.org/t/food-for-thought-bayesian-experimental-designer/208/1 "2026-09-24T12:17:18Z")

</div>

Hi,

Its really amazing how advanced the 80’s were concerning  
the early expert system wave. You already found expert systems  
or fault diagnosis systems, that sorted their abduction questions

via a clairvoyant value computation. It refers to a theoretical  
benchmark in decision analysis, maintenance optimization,  
and information economics. I didn’t find some highly visible

Prolog implementation yet, but I might be mistaken. s(CASP)  
doesn’t have it yet right? Neither does Web Prolog mention  
the concept yet? LLMs use clairvoyant value as well, obviously,

when they stir the dialog. Maybe as an emergent phaenomenon?

```auto
                  ┌─────────────────────────────────────────┐
                  │ HOW LLMS VALUE PERFECT INFORMATION │
                  └─────────────────────────────────────────┘
                                       │
         ┌─────────────────────────────┴──────────────────────────┐
         ▼ ▼
┌─────────────────────────────────┐ ┌─────────────────────────────────┐
│ EXPLICIT SCAFFOLDING │ │ EMERGENT PHENOMENON │
├─────────────────────────────────┤ ├─────────────────────────────────┤
│ The architecture forces a VPI │ │ Autoregressive next-token │
│ calculation (e.g., using │ │ prediction mirrors human │
│ information entropy metrics). │ │ cost-benefit dialogue. │
└─────────────────────────────────┘ └─────────────────────────────────┘

```

One might spin the idea further and just say the valuation  
doesn’t predict some physical system behaviour, but rather  
reflects user preferences. See also:

> **Evaluating User-Agent Collaboration**  
> [https://arxiv.org/html/2608.27818v1](https://arxiv.org/html/2608.27818v1)

LLMs can be also bad clairvoyants:

> **LLMs Fail to Let Go**  
> [https://arxiv.org/pdf/2609.25337](https://arxiv.org/pdf/2609.25337)

Bye