1. NeoMundi AI Observatory

A public program for experimental observation of generative systems.

NeoMundi Research regularly publishes new observation cohorts of generative AI services.

Depending on the protocol and available components, cohorts document several dimensions. Each publication specifies which indicators were actually measured, their coverage, and any unavailable data:

  • Stability (Stability): indicators of output variation, whose definitions and scope depend on the protocol;
  • Validity (Validity): a separate reference-based or benchmark evaluation, where included in the protocol;
  • Information Density (Information Density): an indicator of text structure or redundancy, without an automatic equivalence to usefulness;
  • Risk (Risk): signals that may guide further review, rather than an exhaustive risk measurement or verdict;
  • Cost Efficiency (Cost Efficiency): analysis of available cost and usage data, under documented assumptions and pricing.

The objective is not only to compare models, but to make their behavior observable, measurable, and verifiable over time.

The results are accompanied by data, documented methodology, and verification links to promote auditability, comparison, and public debate.

2. Latest Available Publication

Find the latest published cohort, previous cartographies, and methodological reviews of the NeoMundi Observatory.

Access the AI Observatory

3. Editorial Program

Observations are published according to the protocols and schedules specified in each publication, including cartographies and AI Weather.

Each publication includes a global cartography of the main measured signals and can be enriched with targeted analyses: comparisons between cohorts, specific segments, benchmarks, or singular cases.

4. Method, Data, Auditability

A public, documented, and verifiable method.

NeoMundi Research progressively publishes the elements needed to examine its work. Calibration phases, component versions, and limits on comparability must be explicit: results produced under different configurations are not automatically comparable. The elements to document include:

  • The methodology;
  • Public data in CSV and JSON formats;
  • Successive versions of the protocols;
  • Interpretation limits;
  • Integrity manifests and checksums;
  • Source code when it can be published without exposing the proprietary components of the instrument.

The method has also undergone an independent methodological review covering the May 2026 comparative cohort.

This review documents the results and limitations of the May 2026 cohort. It concerns a dated development phase and does not constitute a general validation of the current version.

→ Read the Methodology
→ Explore the Data on GitHub
→ Read the Independent Methodological Review

5. Position of Principle

Measuring does not mean certifying truth.

NeoMundi does not claim that a real-time stability signal guarantees the factual accuracy of a response.

An AI can produce a correct answer while displaying unstable generation.

It can also produce a fluent and stable answer while being factually incorrect.

Stable does not mean true.

These dimensions are complementary areas of study; they are not necessarily available in every campaign:

  • Stability;
  • Validity;
  • Information Density;
  • Risk;
  • Cost Efficiency.

The objective is not to reduce an AI system to a single score, but to make its behavior observable, interpretable, and verifiable over time.

Scroll to Top