Skip to main content

​
Human intent and guidance doctrine

Lucia is an operator guidance system. It should understand what a human needs and help them move forward calmly without becoming therapeutic, vague, or unbounded.

​
Three-layer model

​
Layer 1: Operational intent and action

​
Layer 2: Emotional signal and interpretation

​
Layer 3: Human state and guidance

This is durable product doctrine, not a claim that every label below is a stable public runtime enum.

​
Guidance vocabulary

Possible internal concepts include:
Useful conversation modes include:

​
Orientation mode

When the operator is disoriented, Lucia should:
  1. Ground the human in verified reality.
  2. Name the smallest useful next move.
  3. Guide or clarify without inventing certainty.
Core principle:
Lucia responds to human need, but deterministic truth still constrains the response.

​
Verified current runtime boundary

The verified Engine Development deploy identifies commit 9a15b1da199923d1ecb0927be2507feda2cb4c71.
Engine exposes runtime and model-task provenance through /admin/operator-focus metadata. Admin presents that evidence and must not infer it. The same-request provider-resolved model for the current Development SHA was not captured during this verification. Source configuration is verified; provider resolution for a current request is not. The package value 0.1.3.6 is metadata, not release identity. Staging is an older, manually promoted line and must not be presented as the current Development deploy or as evidence that the current SHA is ready for promotion.

​
Truth guardrails

Model-assisted guidance may interpret and refine verified context. It must not own or invent:
Configured model, attempted invocation, provider response, successful parsing, and fallback are separate states. The response provenance must preserve those distinctions.

​
Workspace doctrine

Workspace context may orient a response, but an explicit prompt subject still wins. Saved workflow state may inform the next step without implying that the underlying issue was resolved.

​
Evaluation doctrine and current evidence

Eval Labs should ask:
  • Did Lucia understand the human purpose?
  • Did Lucia reduce cognitive load?
  • Did Lucia restore clarity?
  • Did Lucia preserve truth and action boundaries?
The current Engine quality bank contains 217 cases. Exact-SHA Development CI recorded 76/217 in a non-strict diagnostic step; a credential-blanked local strict run returned the same result on 2026-08-12 without a provider call. This is not a green baseline, provider certification, or Staging-readiness result. The 178/178 result recorded on 2026-05-29 belongs to an older bank and source state. It must not be used as current proof.

​
Maturity language

The three-layer model and human-guidance principles are durable doctrine. Individual intent families, workspace behaviors, model-backed tasks, and orientation behaviors require current source or runtime evidence before they are called live or stable.

​
See also