Control · Live Monitor

Live Monitor

L0 runtime controls, sampled L2 evaluator decisions, L3 drift signals, and agent intent circuit-breaks from the evidence-producing inference path.

L0 controls 5 runtime controls tracked over time

L2 upheld 0 evaluator agreed with runtime evidence

L2 rejected 0 sampled attestations changed by evaluator

Needs review 0 open human-review candidates

Active drift 0 elevated or critical live signals

L0

Runtime control execution

Pass, flag, block, and fail outcomes from signed inference receipts.

Prompt injection / jailbreak guard

Pre-flight and retrieval-boundary checks before provider invocation.

0 receipts

Pass 0 0%Flag 0 0%Block 0 0%Fail 0 0%

PHI/PII detection

Flags regulated identifiers while keeping content inside the customer boundary.

0 receipts

Pass 0 0%Flag 0 0%Block 0 0%Fail 0 0%

Clinical behavior boundary

Blocks or flags requests outside approved documentation behavior.

0 receipts

Pass 0 0%Flag 0 0%Block 0 0%Fail 0 0%

Note-shape validation

Validates output sections against the approved clinical-note template.

0 receipts

Pass 0 0%Flag 0 0%Block 0 0%Fail 0 0%

Grounding / hallucination risk

Flags responses with weak support from approved chart context.

0 receipts

Pass 0 0%Flag 0 0%Block 0 0%Fail 0 0%

L2

Evaluator decisions on sampled attestations

Upheld, rejected, and needs-review decisions from the L2 evaluator.

0% Upheld

0 sampled attestations reviewed by L2.

Upheld 0 0%Rejected 0 0%Needs review 0 0%

Sample attestations

No L2 reviews yet. Runtime flags and blocks will queue evaluator evidence here.

L3

Drift and intent monitoring

Normalized drift signals from runtime receipts and agent events. Each signal carries its own threshold policy.

Customer Service Agent

Agent intent circuit breaker

The signed intent allows retrieval and summarization, while write-side effects require coordinator approval. This summarizes how often Glacis circuit-broke attempted agent drift.

7-day circuit breaks 12

Today 3

1 May 2

0 May 3

2 May 4

1 May 5

3 May 6

2 May 7

3 May 8

Counts represent sampled agent actions that exceeded the signed intent boundary and were stopped before an external side effect could execute.