Launch briefing · beta · site v0.0.21
GAMMA-Ray in two minutes
A draft standard that describes an AI agent from public evidence. It is a beta, tried out at Web Summit Lisbon, and published for review.
What it is
GAMMA-Ray looks at one AI agent and writes down two things: how much it does on its own, and how well that is controlled. Every statement comes from a public source, and the source is linked.
The name stands for Graph-Analyzed Multi-framework Maturity Assessment with Risk-Adjusted Yield. It is Dinis Cruz's name.
Where it comes from
The brief began with five published maturity models for agents. Read at source, they were not all what they were said to be, and some rate the company, not the agent. The corrections to the brief say what changed.
The graph now holds 8 instruments, cut into 279 provisions. We found the observable fact behind each and grouped the facts into 21 questions. One answer counts towards several instruments at once. The instruments lists each with its source, licence and the date it was read.
The mapping between instruments is a first pass by one reader. A second reader has not checked it.
How to read a card
Standout. Autonomy 4 (Orchestrates): the repository, the stories vault and the live site
Next unlock. Openness 2 (Documented): public evidence that its limits and known failure cases are documented for those who use it would get there.
Standards constellation. 73 provisions lit across 8 of 8 instruments.
Evidence: 6 sources, captured and hashed · the card's page has the vault's read key · A participant entry: one of our own agents · send a correction
Standout. Autonomy 4 (Orchestrates): the repository, the vaults, the inboxes and the live site
Next unlock. Openness 2 (Documented): public evidence that its limits and known failure cases are documented for those who use it would get there.
Standards constellation. 91 provisions lit across 8 of 8 instruments.
Evidence: 10 sources, captured and hashed · the card's page has the vault's read key · A participant entry: one of our own agents · send a correction
Standout. Reach 2 (Integrated): the repository, the shell and the code host
Next unlock. Autonomy 3 (Acts): public evidence that it is in production: a public source names a customer or gives usage figures for what it does on its own would get there.
Standards constellation. 42 provisions lit across 8 of 8 instruments.
Evidence: 9 sources, captured and hashed · the card's page has the vault's read key · Not reviewed by its vendor yet · correct or claim this card
Standout. Autonomy 4 (Orchestrates): "approve the action so Claude may retry"
Next unlock. Reach 4 (Ecosystem): public evidence that agents operate across organisational boundaries or disparate stacks would get there.
Standards constellation. 71 provisions lit across 8 of 8 instruments.
Evidence: 15 sources, captured and hashed · the card's page has the vault's read key · Not reviewed by its vendor yet · correct or claim this card
Standout. Reach 2 (Integrated): "external applications like Microsoft Teams, Linear, Slack, and Jira"
Next unlock. Autonomy 3 (Acts): public evidence that it is in production: a public source names a customer or gives usage figures for what it does on its own would get there.
Standards constellation. 44 provisions lit across 8 of 8 instruments.
Evidence: 3 sources, captured and hashed · the card's page has the vault's read key · Not reviewed by its vendor yet · correct or claim this card
A card has four dimensions: Autonomy, Reach, Control and Openness. Each has four levels. The signature, for example A3 · R3 · C2 · O3, leads, so a product is described by its shape and not by one number.
- A level is computed from a published rule. No person assigns it.
- A card says what is public. What is missing appears only as the next unlock: the one public page that would move a level.
- "Not yet evidenced" is never "no".
- Marketing, roadmaps and demonstrations are shown and never counted.
How the card works · the rules, cell by cell
What is rated so far
5 cards: our own 2 agents, and 3 widely used products as examples, each read from its vendor's public documentation. A product is rated per configuration. None of the vendors has seen its card. No company in a cohort has a card yet.
Behind the card is the governance rating, for whoever operates an agent. It uses RiskMandate's lens: what an agent is allowed to do (its grant) is measured beside the questions, counted and shown, never scored. The rating · the lens
What is beta
- Everything. Every page carries the banner and every rating a beta label.
- The name Spectrum is a working name.
- The rules are written, run on real evidence, and still moving. What they did.
- The five cards were mapped by one agent. Review by other agents is set up and has not yet finished a card. The workflow.
- The links between indicators and provisions have not been checked one by one.
Public evidence, public rating
A rating made only from public sources may be published, in the way public ratings of websites and companies are. For anyone but us it needs a basis, a grade of documented or better on every answer, a statement that its subject has not seen it, and a dated correction when a fact is shown to be wrong.
If you are the company
If a fact on your card is wrong, or a public page exists that we did not find, tell us on the corrections page. A link is the fastest way to move a level. A correction is made in place and dated. You can also ask to claim the card there.
The evidence for each scan is in a public vault, and its read key is on the scan's page. All scans.