Skip to content

Architecture

From the outside Vectadyne is an endpoint and a token. Inside, everything runs through five stages, and each one is where a different kind of quality problem gets solved.

outcome becomes newevidence1 · EvidencePRs · CI · incidents ·reviews2 · Derivationtyped, scoped memory3 · Guideswhat matters for this topic4 · Retrieval + admissionranked, then filtered byauthority5 · Governanceverdict · citations · audit

Raw engineering activity: a merged PR, a failed CI run, an incident, a review comment, a correction from a human. Evidence is immutable and append-only — the record of what happened does not get edited later because someone disagreed with it.

Correction is expressed by adding, not by overwriting: a correction supersedes an earlier claim and both stay in the record. That is what makes the audit trail worth having.

Evidence becomes typed memory. A PR description is not a memory; the conclusion drawn from it is. Each memory carries:

  • a kindfact, decision, lesson, policy, guide, evidence
  • a scope — where it applies
  • a statusactive, disputed, stale, deprecated
  • a sensitivity — the clearance needed to see it
  • links to the evidence that supports it

Kinds are not decoration; they carry precedence. A policy outranks a lesson when they conflict. See The memory graph.

A topic guide is the standing answer to “what should anyone touching this know?” — assembled from the memory that applies to a topic, and resolved fresh rather than stored as a stale digest.

This is the layer that makes retrieval feel like knowledge rather than search. Without it, every query starts from scratch and the same five facts are re-derived on every run.

Two distinct steps, and conflating them is the mistake that makes a system leak:

Retrieval ranks candidates — hybrid lexical and dense, plus one hop of graph edges to pull in what the top hits are connected to.

Admission then filters by authority: tenant, scope, clearance, status. It runs at read time, on every request, against the calling principal.

Admission is not a re-ranking pass with a low weight for “probably shouldn’t see this”. It is a filter. What comes back is what this credential is allowed to see, which is why an empty result is a legitimate answer rather than an error.

See Retrieval and admission.

The action-facing layer. Given a proposed action, produce a verdict with reasons and citations, and record what was known at the moment of the decision.

The record is the part that outlives the run. “The agent did the wrong thing” is a very different conversation from “the agent did the wrong thing and here is what it had been told”.

See Governance and audit.

The dotted line in the diagram is the whole point. An agent’s outcome — what it did, whether it worked — goes back in as evidence, so the corpus improves every run. A system without that line is a retrieval index that never learns.

Only the endpoint. Seven verbs map onto these stages:

StageVerbs
Evidenceevidence.ingest
Derivationmemory.remember
Guidesguide.resolve
Retrieval + admissionmemory.search
Governancepreflight.action, context.assemble, context.report

context.assemble spans stages 3 to 5 in one call, which is why it is the verb most integrations end up centred on. Full schemas: MCP reference.