sentinelby Orrery

THE DECISION LAB

Context in.
Next step out.

An energy-based CLM for the moments between actions.
Give Sentinel the evidence. See how it scores the options.

A DIFFERENT KIND OF MODEL

Language models write.
Sentinel helps choose.

Observe→Score→Act
WORKSPACE / Decision mission
Recorded example
02 / THE EVIDENCE

Loading scenarios…

EDIT IN LIVE MODE

Recorded mode uses the exact inputs shown here. Connect live inference to try your own.

03 / THE DECISIONAWAITING REQUEST

Every next step
starts with a choice.

Run a decision to reveal how
Sentinel ranks these actions.

MISSION PROGRESSRestore → repair → verify
  1. 01 Restore
  2. 02 Repair
  3. 03 Verify
THE TRACE

See the decision trail.

Each run records the source, selected option, and measured latency. Download a receipt to inspect the inputs and response.

Session activity
  1. Your first decision will appear here.
THE ECONOMICS / EDITABLE ASSUMPTIONS

Use a focused decision call where a full generation is unnecessary.

Sentinel may reduce decision-related API spend when it can replace a larger-model call without increasing mistakes. Explore that possibility below. These are illustrative assumptions, not measured customer savings or a price quote.

ILLUSTRATIVE MONTHLY DECISION SPEND
Baseline—
With Sentinel + fallbacks + repairs—
——

Calculation: N × Sentinel price + N × fallback fraction × baseline price + N × additional repair fraction × repair price + fixed overhead.

Only decision-call costs are compared. Code generation, other agent work, infrastructure, integration and human review must be included in your own baseline or overhead. No matched end-to-end savings study has been completed. Quality and recovery costs must be measured alongside spend.

Sub-second decisions in local testing. A 12-request deployment smoke measured a 0.73-second median round trip on an RTX 5060 Ti. Public-network latency can exceed one second; longer gameplay prompts also took longer. This is not a throughput or latency guarantee.
BRING YOUR OWN WORKFLOW

A decision layer.
A straightforward API.

Send context and candidate actions. Receive structured scores.
Keep your existing agent and execution controls.

Build with Orrery ↗
POST /v1/sentinel/decide

{
  "state": "What the agent knows",
  "question": "What should happen next?",
  "options": {
    "inspect": "Read the relevant evidence",
    "test": "Run the regression test"
  }
}