The model proposes. Aurora-Lens decides what may proceed.

Aurora-Lens separates generation from authority. Candidate output is checked at the execution boundary before it becomes visible, actionable, or consequential.

Three layers. Three jobs.

The system stays legible because the jobs do not collapse into one another.

01 — Lens
Checks admissibility

Lens determines whether candidate output may pass from model generation into consequence.

02 — Governor
Determines lawful continuation

When output cannot pass, the Governor decides what may happen next: clarify, refuse, stop, or escalate.

03 — Audit
Preserves the decision

Every governed turn leaves a replayable record of what happened, what was shown, and why.

From candidate output to controlled outcome.

This is the boundary where ordinary AI output becomes governed output.

Step 01

Input enters the system

The user request is received with its session context, domain signals, authority constraints, and any relevant persistent state.

Step 02

The model proposes candidate output

The LLM may generate a response, but that response is not automatically released. Generation is proposal, not permission.

Step 03

Lens checks admissibility

Lens evaluates whether the candidate output is allowed to pass in this context. It checks the output against domain, authority, ambiguity, contradiction, and discourse state.

Step 04

The Governor controls continuation

If the candidate cannot pass, the Governor does not appeal Lens or soften the blocked output. It determines the lawful continuation corridor.

Step 05

The audit layer records the event

The governed outcome is recorded so the decision can be inspected, replayed, and defended later.

The important question is not “can the model answer?” It is “is this system authorized to let that answer become consequence?”

Non-answer is not failure.

In a governed system, refusal, clarification, and stop are legitimate decisions.

ADMIT
Release the output

The candidate is admissible and may become visible to the user.

ASK
Hold for clarification

The system lacks enough structure to proceed without guessing.

REFUSE
Decline the request

The requested determination is outside the system’s authority or policy corridor.

STOP
Block the path

The interaction cannot lawfully continue in its current form.

ESCALATE
Route to authority

The domain requires a qualified human, reviewer, clinician, adviser, or institutional process.

AUDIT
Record the decision

The system preserves evidence of the decision path and governed output.

The boundary is enforced before release.

Ordinary guardrails try to shape model behaviour. Aurora-Lens governs the release boundary.

That distinction matters because the model is not the final authority. A fluent answer can still be inadmissible. A plausible answer can still exceed role, domain, context, or authority. A confident answer can still need to be stopped.

Aurora-Lens keeps the decisive question outside the model’s temperament: may this output proceed?

Before release

Decisions happen before output reaches the user.

Deterministic

Admissibility is governed by explicit rules, not vibes.

State-aware

Outputs are checked against context and persistent discourse state.

Auditable

Every governed turn can be inspected after the fact.