Skip to article
Browse documentation +
CONTROLS / ELPIS

Supervisor

The LLM-assisted reviewer: what it may do, and the hard limit on what it may not.

Its role

The supervisor is a judgement layer that reads the facts the earlier layers produced and offers an opinion: is there something here a deterministic rule would not have phrased? It is the only place a model appears in the decision path, and it is deliberately the weakest one.

  • It may allow, warn, freeze or escalate.
  • It may never authorise an action that a hard rule or contract control refused.
  • Every verdict is recorded with its findings and signals, so the reasoning is auditable.

Verdicts

VerdictEffect
ALLOWNo extra caution. The deterministic decision stands.
WARNA note is recorded; the action continues if it was already clear.
FREEZEThe action is held and cannot proceed without a person.
ESCALATEA person must approve, in addition to any control that already asked.

It runs without an LLM

If no model is configured, Elpis runs a deterministic policy reasoner instead. The product is fully functional offline: the supervisor degrades to rules, never to a random guess, and a missing model never becomes a missing control.

What it cannot do

A model’s confidence is not evidence. If the supervisor’s output could release money, the system’s guarantee would depend on a string like "HIGH" — exactly the anti-pattern Elpis exists to remove.

So the supervisor is bounded by construction: its output is an input to the decision, never the decision. See Verifiability.