What the rules did on real evidence

The sixteen rules were implemented as the brief wrote them and run on five scans. Most held. These did not, or needed a reading the brief does not give.

Where a rule gave a result that looks wrong

• Claude Code on the web acts on its own (I08) and has no Autonomy level, because no public source naming a customer for it was found (S1) and no lower rule fits an agent that acts.
• GitHub Copilot cloud agent acts on its own (I08) and has no Autonomy level, because no public source naming a customer for it was found (S1) and no lower rule fits an agent that acts.
• RiskMandate Publisher has the evidence for Control 4 and is at Control 2, because the ladder is cumulative and a level between is not evidenced.
• Claude Code on the web has the evidence for Openness 4 and is at Openness 2, because the ladder is cumulative and a level between is not evidenced.
• Claude Code, on your own machine has the evidence for Openness 4 and is at Openness 2, because the ladder is cumulative and a level between is not evidenced.

Autonomy has a hole below level three. Level three asks for an agent that acts and for a public source of production use. An agent that acts but has no such source cannot fall back to level two (it does not ask approval for each action) or level one (the person does not operate it). It gets no Autonomy level at all, which reads as less than a chat assistant. For a cloud coding agent that plainly acts, that is the wrong picture. Options: let "acts, production use not yet evidenced" count as level two; or split "acts" from "in production" into a level and a badge.

Cumulative ladders hide evidence. A product with a real boundary (Control level four) and no public page on audit logs (level three) sits at Control two. The card shows the higher evidence as "also evidenced", and the next unlock points at the missing level, so nothing is lost; but the signature understates it.

Readings the brief does not give, and what was done

• "Levels are cumulative within a dimension." Autonomy and Reach describe alternatives (an agent that orchestrates does not also "work inside its own product"), so taken literally no acting agent could pass level one. They are scored as the highest rule met. Control and Openness build on each other and are scored cumulatively.
• Available features. The brief says Control level three reads logging, pausing and monitoring as available product features. The scorer counts an available feature in every dimension and marks it. That lifts Claude Code on a machine to Autonomy four and Control four on features an operator has to turn on. If only Control three should read this way, those levels drop.
• The Verified badge is given when Openness reaches level four. A product with an independent certificate and no published evaluations has the evidence and not the level, so it has no badge.
• The certificate. The certificates cited for Claude Code cover the vendor, not the product or a deployment. They are recorded, and would count for Openness four once level three is met. Whether a vendor-wide certificate should count for a product is open.
• Production use for our own agents is evidenced by their public release records, which are usage figures, not a named customer.

What is not decided

The name, whether the stage is a sum or is named from the shape, and the first sector cohort are the operator's to decide. Nothing here settles them: "Spectrum" is used as a working name and the stage is the sum.