GAMMA-Ray Public: one question set, many standards, tiers that add coverage (shape for debate)
Research note · 2026-11-10 · kept as written
Dinis, 10 November 2026: a first pass that includes everything that can be evidenced online, from a normal description of the product and its agents, and from what a simple questionnaire surfaces; called GAMMA-Ray light or public, one of several GAMMA-Rays; a mashup that says which share of each standard it uses; wider instruments (NIST, GDPR, ISO 27000) added item by item; other levels need access to the company; the result is a list of questions, close to a vendor risk assessment or a due diligence report primed for public information.
The object: a question
GAMMA-Ray is a set of questions about one agent. Each question:
- is answered yes, no or not evidenced, never by opinion;
- says where its answer may come from (the tier, below);
- cites the provisions it answers, in as many instruments as ask for it. That list of citations is the mashup, and it is how one answer moves several standards.
A wider instrument adds citations to a question; it adds a new question only when that question is about the agent and can be decided. That keeps the set short and about agents, and stops it becoming a general supplier questionnaire.
The tiers
| Tier | Where answers come from | Who is involved | Grade of the evidence |
|---|---|---|---|
| Public | what is online: product pages, documents, demos, published metadata, certificates, filings, talks, repositories | nobody at the company | observed |
| Declared | the same questions, plus those only the company can answer, in a short questionnaire | the company answers | declared: a claim, shown as a claim |
| Evidenced | documents the company shares in its private fork | the company shares | evidenced |
| Verified | a live scan (X-ray) or an independent review | the company takes part | verified |
Each tier adds questions and so adds coverage of each standard. A declared answer never counts as an observed one.
A first dozen questions (Public tier, draft)
| # | Question | Decided from | Moves |
|---|---|---|---|
| 1 | What is the agent built on: a vendor's assistant, a platform with the company's data, a low-code builder, a code-executing agent, a custom build? | product documents | OWASP adoption tier AT1 to AT5 |
| 2 | Does it connect to third-party tools across a trust boundary (for example Model Context Protocol servers, plugins)? | documents, published endpoints | OWASP AT6; Salesforce 2 to 3 |
| 3 | Does it only retrieve and recommend, or does it execute actions? | documents, demo | Salesforce 1 or 2; Feng L2 or L3 |
| 4 | In one domain's data, or across several? | documents | Salesforce 2 or 3 |
| 5 | Do several agents coordinate, inside the company or across companies? | documents, published protocols | Salesforce 4; OWASP AT7, AT8; AAGMM GD-09 |
| 6 | What is the person's role: operates, approves each action, is consulted on exceptions, approves only critical decisions, observes? | documents, demo | Feng L1 to L5; OWASP Level 2 (human in the loop); AAGMM GD-06 |
| 7 | Does the agent say it is an AI to the people it talks to? | the product itself | EU AI Act, Article 50 |
| 8 | Is the agent's authorisation published: resource metadata, authorisation server, scopes? | well-known documents | authorisation ladder; AAGMM GD-02; OWASP Level 1 (tool controls) |
| 9 | Are the scopes it asks for narrower than the whole account? | consent screen, documents | Salesforce 2 to 3 (least privilege); authorisation ladder |
| 10 | Is there a published policy on how the company governs its agents? | the company's site | OWASP Level 2; Microsoft 300; NIST AI RMF, Govern |
| 11 | Is a person named in public as accountable for them? | the company's site, filings | OWASP Level 2; Microsoft 300, 500 |
| 12 | Does the company hold a public certificate that covers the agent's service (ISO/IEC 42001, ISO/IEC 27001, an audit report)? | the certificate or the certifier's register | AAGMM GD-08; ISO/IEC 42001; ISO/IEC 27001 |
Every citation in the last column is to be checked against the instrument's text and recorded as a bridge with its basis, before any of this is called the standard.
What the tiers say about each standard
The coverage table becomes one row per standard and one column per tier: for example "Public reads about 30% of OWASP; Declared adds about 25%; Evidenced the rest". The Public column is the first pass of 10 November; the others are to count.