What your AI sees
"What your AI sees" is a transparency view into the context Altretta hands your model — and the place where you set the limits on it. It turns "trust me" into "check for yourself, and change it if you don't like it."
Open it from the command palette: Trust · what your AI sees. You can also reach it from a note's context menu to see that note specifically.
The trust panel
For the note you're in, the panel lists exactly what the AI would be given about it: its verified facts, whether those are signed in the DAG, its neighbours by meaning, which of those carry DAG provenance, and its backlinks. It also names the verified-knowledge tools your AI can call. Because every answer is grounded in source:lines citations (see Grounded, cited answers), you can trace a model's reasoning straight back to your own notes and confirm it used what it claims to have used.
The controls
Under Control · what your AI can see and do, three settings decide what a connected client is allowed to do.
Access level
- Read-only — the default, and the safe one. Every writing tool is refused outright, so a connected client cannot change your vault at all.
- Read & write — unlocks the writing tools.
A read-and-write level is genuinely read and write. Only the propose a note tool routes through the Review inbox for your approval; editing a note, adding or removing tags, creating a folder, writing to the graph and running an ingest all take effect immediately. No tool parks a mutation and waits for a human. Grant write access to a client you trust, not by reflex.
Excluded folders
List any folder here and it is never shared with your AI. Matching is on whole path segments: excluding Personal/Finance hides that folder and everything inside it, but leaves a sibling called Personal/Finance2 alone. Backslashes, leading and trailing slashes are all normalised, so the form you type doesn't matter.
This is the right control for the folder of medical notes, salary reviews, or client contracts you keep in the same vault as everything else. It is exact and path-based — no guessing at content.
Require grounding
Off by default. Turn it on and the AI declines to answer whenever your vault doesn't give it solid support, rather than answering from thin evidence. With it off, an answer is produced whenever there is any usable source at all; with it on, "some evidence, but thin" becomes a refusal too.
Vault integrity
The panel also shows the size of your signed history — how many actions are logged, and how many active tips the history currently has. See The audit trail for what that history contains and Signed history for how it's built.
The AI activity log
Beyond any single answer, Altretta keeps an AI activity log recording what the AI queried and saw over time. It's the audit trail for your second brain's AI: which parts of your vault a model touched, and when. It lives alongside the consistency review — both are covered in Activity log & consistency review.
Why this matters
No black box
See the actual context behind an answer, not just the answer.
Verifiable, not just plausible
Trace every claim to source:lines in your own files.
A boundary you set
Exclude folders the AI must never read, and keep it read-only unless you decide otherwise.
An audit trail
A running log of what the AI queried and saw across your vault.
Next step
→ Activity log & consistency review to see the record of AI queries and catch notes that contradict each other.
Related
AI you can trust
Back to the AI section overview.
Provenance & proofs
Verify the signed anchor behind each cited passage in the Proof Inspector.
The audit trail
The vault-wide, signed record that the AI activity log complements.
Review inbox
What the approval inbox covers, and what it doesn't.
MCP tools reference
The tools that produce the context shown in the trust panel.
Glossary
Context, grounding, provenance — the trust terms, defined.