# Nine questions to put to any pricing vendor

Published by World0, Inc. at https://world0.ai/diligence.html

Any serious vendor can answer the first eight from documentation it already
holds. The ninth needs measurement. World0 answers all nine for its own
measurement system on a call.

---

## Structure and entities

1. How many distinct pricing entities does your deployment create? Does one
   model instance, or one fitted model, ever price for more than one competing
   seller?

2. Which components are shared across client deployments? Weights, training
   data, feature stores, demand estimates, benchmarks. What isolates one
   client's data from every other client's estimator?

## Data and identification

3. Is any model fitted on data pooled across competing sellers? If pooling
   exists in any form, what does the vendor claim it identifies, and what
   validates that claim before deployment?

## Configuration

4. Which configuration constants change the system's pricing behavior, and how
   does each scale as data volume grows? Does regularization scale with sample
   count? Who reviews these constants before a deployment, and are they
   disclosed to the client?

## Action space

5. What price floors, caps or bands constrain the system? What does the system
   do at a band edge, and is time spent at band edges monitored, logged and
   reported to the client?

## Monitoring

6. Does production monitoring cover the full distribution of pricing outcomes,
   including extreme episodes, or does it report summary averages? What
   triggers human review of an extreme pricing episode, and who sees it?

## Learning dynamics

7. Does the system refit or adapt in production? Can fitted coefficients move
   in response to rivals' observed behavior? How does the vendor classify the
   system: a demand forecaster, an adaptive system reacting to competitors, or
   a strategic agent?

## For AI agent systems

8. Which base model and which dated snapshot does the deployment run? Are
   sampling settings fixed, disclosed and reproducible? Do competing
   deployments run the same base model, and what does the vendor do about the
   correlation risks of shared weights?

## Evidence

9. Can the vendor produce pre-deployment measurement of conduct-relevant
   behavior for the configuration being sold? A good answer shows a design
   locked and deposited in advance, a controlled synthetic market, and results
   a reviewer can reproduce independently from the logs. What acceptance
   criteria does the vendor set for its own system in advance, and what
   measured result would fail them?

---

## How to read the answers

A vendor with good answers documents entity separation, discloses its
constants, monitors its tails, and can hand over measurement a reviewer can
check. The ninth is the question no documentation answers today.

The nine turn what a pre-deployment measurement examines into questions a
buyer can put to a vendor.