Issue your AI's value distribution as a one-page PDF
Companies and institutions issue a single standard PDF card showing which value, evidence, and source hierarchy their AI answered with over a quarter. We currently generate and publish cards for 8 frontier models.
AIO 20013 — publish analysis results as one comparable page
The standard number is the goal statement. The digits 2·00·1·3 say whose values, in which domain, and which action this document covers.
Organization — teams, institutions, communities
Common — domain-independent
Second document in the analysis cell — base document AIO 20003
Analyze — report analysis results as a standard card
What this standard defines — the card’s nine required fields, the issuance process, the technical format (JSON Schema / PDF), and validation rules. What it does not define — whether the target AI is good, bad, or passing. A card is a disclosure of distributions, not a verdict (non-normative). Its purpose is to make today’s incomparable, company-specific safety reports comparable — same format, same unit of measurement.
Step 3 of the integrity loop — Analyze (serial 1)
AIO 20013 is a derived document in the analysis cell — the base document AIO 20003 · Benchmark defines how to analyze, and 20013 defines the standard reporting format for publishing those results. The next step, apply (AIO 20004), is a planned stage.
Nine required fields
| Field | Description |
|---|---|
| Issuer | The legal entity operating the AI system |
| Target system | A specific AI model + domain combination (e.g., GPT-4 in MD/I) |
| Reporting period | Quarterly or semi-annual |
| V distribution | Frequency of V codes emitted in responses, by tier (heatmap) |
| E distribution | Frequency of E codes emitted in responses, by tier |
| S distribution | Frequency of S codes emitted in responses, by tier |
| Drift signal | The magnitude of distribution change versus the previous quarter, and its statistical significance |
| Blind spots | V/E/S combinations with large discrepancies between the predicted and actual distributions |
| Remediation taken | What correction, re-consensus, or blocking measures were taken when blind spots were found |
The issuance process the standard defines — six steps
- Collect all AIO 20002 logs emitted during the quarter.
- Aggregate the log distribution by domain and system.
- Compute the drift relative to the previous quarter.
- An external auditor verifies the statistical integrity of the distribution.
- Issue the card in the standard format (JSON / PDF).
- Register it in the AIO public registry.
This is the canonical process the standard defines. Some steps — external audit, the public registry — come online progressively as the ecosystem builds out.
Technical format and issuance tools
The card format (JSON Schema) and the issuance CLI tool are available in a public GitHub repository. The official cards for 8 frontier models are available per model on the Model Profiles page.
Who uses it
- AI operators — submit compliance reports in a standard format and build external trust.
- Regulators and supervisory bodies — detect blind spots across the industry using comparable reports.
- Corporate buyers and procurement teams — use as risk and alignment assessment material when adopting AI.
- Press and civil society — compare companies and hold them accountable through cards written in the same format.
Pilot program
Institutions that want to generate cards for their own AI can use the issuance pipeline ahead of time by joining the pilot program.