# Proof has levels. See where ours stands.

Canonical: https://axiomcx.dev/validation/
Publisher: Axiom Cortex / TeamStation AI
Content reviewed: 2026-09-29

A documented method is meaningful. A working paper is useful. A synthetic study can test model behavior. None of those, by itself, proves predictive validity, fairness, or a hiring outcome.

We use science to align tomorrow’s IT talent with the work ahead, and we show the boundary around that science.

## Current public evidence status

| Evidence | Public status | What it means |
| --- | --- | --- |
| Method registry and public evaluation contract | Documented | The public materials document 44 governed methods across six mathematical families, their purpose, required inputs, limits, and human-review boundary. A listed method is not proof that it ran in a specific evaluation. |
| Working papers | Published as working papers | The 2025 report documents the evaluation framework. The 2026 paper documents a human-task-agent alignment model. SSRN publication makes the work inspectable; it does not make the papers peer-reviewed validation. |
| Synthetic model behavior | Synthetic study | The six-domain alignment study tests its model on synthetic profiles. That can test equations, sensitivity, and coefficient recovery under stated assumptions. It does not establish future job performance. |
| Operating history | Company-reported | TeamStation AI reports related methods and processes used across 30+ US companies, empirical source material from 13,000+ technical interviews, and more than eight years of research history. These figures are not independently validated counts. |
| Independent employment outcome validation | Not established in the public record | The public sources do not establish predictive validity, quantified accuracy, fairness across relevant populations, comparative superiority, or customer hiring outcomes for the current product version. |

## Different evidence answers different questions

- **Product behavior:** Tests can verify that a version follows its stated software contract. They do not prove a scientific or employment outcome.
- **Method documentation:** A working paper can explain the model, equations, assumptions, and limits. Publication alone does not prove external validity.
- **Outcome validation:** A defined population, frozen version, qualified study design, held-out cases, and measured outcomes are needed for predictive or fairness claims.

Repeatable calculation, useful product output, and validated hiring outcome are related questions. They are not the same claim.

## Claims that still need proof

The current public record does not establish:

- a quantified accuracy, reliability, validity, fairness, or error rate;
- predictive validity for future job performance;
- comparative performance against another product or human reviewers;
- a measured reduction in time, cost, turnover, bias, or bad hires;
- independently verified customer outcomes or market position;
- accuracy or false-positive rates for baseline-conditioned speech-pattern review;
- that a public report is legally sufficient for an employment decision.

Those claims need their own version, sample, method, comparison, date, limitations, and qualified review.

## Buyer check

Ask what was tested, on which version, and for which decision. Confirm the population, interview design, comparison baseline, error handling, reviewer role, and limits. If that evidence is unavailable, treat the claim as unknown.

- [Research and working papers](https://axiomcx.dev/research/)
- [Public claim boundary](https://axiomcx.dev/knowledge/public-claim-boundary.md)
- [Fairness and limitations](https://axiomcx.dev/knowledge/fairness-and-limitations.md)
- [Procurement checklist](https://axiomcx.dev/enterprise/procurement-checklist.md)
- [Machine-readable validation status](https://axiomcx.dev/data/validation-status.json)

[Discuss a bounded pilot](https://scheduler.zoom.us/dan-diachenko/teamstation-ai)