What we do not claim
Four limits worth knowing before you build on this. Each is stated the same way — the limit, what still holds, what to do about it — because a limitation without its remedy reads as instability rather than candour.
1. Advisory-first is not enforcement
Section titled “1. Advisory-first is not enforcement”The limit. A verdict does not stop your agent. An agent that ignores preflight.action, or never
calls it, is ungoverned for that action.
What still holds. Authority is enforced server-side regardless: tenancy, scope, clearance and capability are refusals, not advice. Every call is in the audit trail.
What to do. Put the branch in your loop, test that you asked, and enable the CI/PR gate where you need a real stop.
Full detail: Advisory-first.
2. The scope lattice is v1
Section titled “2. The scope lattice is v1”The limit. Only two scope relationships are structurally computable today: an exact match,
and org:* covering everything in the organisation. There is no repository-under-team, no
service-under-repository, no arbitrary hierarchy. Anything else fails closed — it is treated as
“not covered” rather than guessed at.
What still holds. Both computable cases are exact and enforced, and failing closed means a mis-modelled scope loses you recall, never isolation. You will notice it as memory that does not surface, not as memory leaking somewhere it should not.
What to do. Model scopes flat — repo:payments, repo:infra, service:checkout — and use
org:* for genuinely organisation-wide policy. Do not encode hierarchy in the key and expect it to
be understood.
3. Degraded mode is a smaller answer, not an error
Section titled “3. Degraded mode is a smaller answer, not an error”The limit. Under load or partial dependency failure, context.assemble returns a reduced
result with degraded: true. You get less context than you would have.
What still holds. An outage never stalls your agent. A degraded assemble is a successful response carrying a real verdict, and it says so explicitly rather than pretending to be complete.
What to do. Treat degraded: true as a success and never as a retry trigger — retrying a
degraded response adds load to a system that is already telling you it is under pressure. Log it, and
consider surfacing the reduced confidence to your model.
See Degraded mode.
4. There is no SOC 2 attestation
Section titled “4. There is no SOC 2 attestation”The limit. Vectadyne has no SOC 2 report. If your procurement process requires one, we do not have it.
What still holds. The audit export is the artifact: a complete, exportable record of what was known, what was decided and by which principal, at the moment of each decision. Machine principals have their own identities and their own trails, so “which agent did this” has an answer.
What to do. Use the audit export for the evidence you need, and tell us if a certification is a hard requirement — it informs the roadmap.
Not built yet
Section titled “Not built yet”Things that do not exist. They are listed so you do not design around them:
- Multi-hop graph traversal (one hop of edges is returned; deeper walks are not)
- Bring-your-own-cloud deployment
- SAML SSO and SCIM provisioning
- Self-serve signup — Vectadyne is in private beta
Deliberately not our job
Section titled “Deliberately not our job”- Generating your code or your plan. Vectadyne is not a model.
- Managing your prompts or packing your context window. It answers what should be known; how you spend the budget is yours. It will measure the spend if you report it.
- Being your vector database. It uses retrieval internally; that is an implementation concern, and the backend is swappable without changing anything above.