Bounded tools
The agent may decide what to try. Registered capabilities, identity, policy, and environment decide what it can do.
Trust model
The agent proposes. The platform constrains. Evidence decides.
ConvergeQL is designed so reviewers can reconstruct what happened without asking a model to narrate its private reasoning.
IdentityCapabilityPolicyEvidenceHuman approval
Deterministic gates
Model confidence is not an acceptance predicate. Required validators must return evidenced results against one candidate.
Explicit state
ChangeRun state belongs to the control plane. Agent session memory is advisory and is never the workflow system of record.
Evidence over narration
Tool results, validator outputs, diffs, approvals, and Git identifiers are execution truth. Explanations point back to them.
Human delivery gate
The initial authority boundary stops at a human-approved GitHub change. No autonomous merge or production deployment.
Provider separation
BigQuery and Snowflake evidence retain different meanings. Unsupported, delayed, or inaccessible evidence fails closed.
Current public boundary
Marketing site live does not mean product integration live.
- No customer warehouse or repository connection exists.
- No BigQuery or Snowflake adapter has passed conformance.
- No authentication, tenancy, retention, or regional control plane is implemented.
- No agent runtime is connected to the public marketing experience.
- No customer outcomes, savings, accuracy, or reliability improvement is claimed.
Decisionchange_join_strategy
ReasonCandidate produced duplicate customer grain
Evidencevalidation://candidate-02/grain
Resultcandidate-03
Synthetic decision record. ConvergeQL records explicit rationale and evidence references, not hidden chain of thought.
Design partner preview
Help shape the first governed model-change workflow.
Bring one existing dbt-managed model, its declared invariants, and a review problem worth solving.