How it stays safe

Handing an AI the keys to your scene sounds risky, so 3dCMS wraps it in guardrails that make the Assistant safe to trust.

• All-or-nothing edits. Every batch is validated before it's applied. If any single operation is invalid, the whole batch is rejected and your scene is left exactly as it was, never half-applied.
• It retries. When a batch is rejected, the Assistant is told what went wrong and gets a couple of automatic attempts to fix it.
• No quiet half-edits. An operation only accepts the fields it actually acts on. Ask for something an op can't do in one step (recoloring while updating, moving a duplicate, reparenting a group) and the batch is refused with an explanation the Assistant can act on, rather than reporting "1 updated" while your scene sits unchanged. In practice you see it split the work into steps that all land.
• One Undo. The entire edit (however many objects it touched) collapses into a single Undo step. Don't like the result? One Ctrl+Z.
• Sanity guards. Nonsense transforms (a position that isn't three real numbers) are rejected outright, so objects can't be flung into the void or vanish.
• It targets objects by identity, never by guesswork, so it edits the thing you meant.
• It can read back what it can set. The scene summary the Assistant reads now includes the fields it was previously able to write but never see: which neighborhood a node is bound to, a link pointing at a saved viewpoint, whether an object carries a tap action or animation, and a particle system's count and sprite. A setting it can't see is one it will happily set twice or undo by accident.
• Very large scenes are summarised, and say so. Past a size ceiling the summary is truncated with an explicit marker, rather than being silently cut off mid-scene and leaving the Assistant confidently editing a scene it has only half of.
• The lint names dead references. Its advisory warnings now include a node bound to a page you've since deleted (which would switch a whole room off), links pointing at pages or viewpoints that no longer exist, and similar loose ends.

If something does go wrong on Anthropic's side (a bad key, a rate limit), the Assistant shows a plain-language message telling you what to check.