Synaptwin

Closed-loop check

Is the loop around your AI ready for more scope?

Twelve questions on whether one AI initiative's whole loop is evidenced, from the outcome you are buying to the evidence that earns more authority. Enter the evidence before the score. Unknown stays unknown.

Your own answers, unverified. A proposed management rubric: not a Technology Readiness Level, not a probability of success, not a safety certification and not legal advice. A named person decides any release.

What this profile looks at
Verified outcome
A verified outcome
Loop
Inputs → specification → proposal → checks → authorised action → outcomes → correction
Measures
Independent outcomes, errors and full lifecycle cost
Control boundary
Permissions, intervention and reviewed release
Diagnose first
Define the intended users, the scope and the harms that are unacceptable.

Readiness index

—

Answered

0/12

Evidence complete

0/12

Critical-gate floor

—

Not assessed yet

Answer the questions, and enter the evidence before the score.

The three weakest questions

  1. 1. Outcome · Unknown
  2. 2. Specification · Unknown — Weak specification: clarify constraints and acceptance tests first.
  3. 3. Bottleneck · Unknown

We go through the three weakest questions with you. The result travels with the request as scores only, never your text.

The twelve questions

How to score

Score the evidence, not the ambition. Unknown is never zero and is left out of the average.

Unknown · 0 · fails or absent · 1 · concept · 2 · offline evidence · 3 · bounded pilot · 4 · sustained in scope

  1. Specify · Outcome· Critical gate

    1What independently verified outcome are we buying?

    Baseline, target, an independent measure of success and the accountable owner.

    Lever: Objective

  2. Specify · Specification

    2Can intent become an executable specification?

    Hard constraints, permitted actions and acceptance tests, written down.

    Lever: Requirements

  3. Diagnose · Bottleneck

    3What actually limits the end-to-end result?

    Errors traced to data, sensing, model, planning and execution, and compared.

    Lever: Allocation

  4. Learn · Data· Critical gate

    4Are the data relevant, representative and lawfully usable?

    Provenance, permissions, coverage and a held-out evaluation.

    Lever: Data

  5. Generate · Model capacity

    5Does more model or inference capacity improve deployed outcomes?

    Capacity or inference budgets compared with simpler alternatives at full cost.

    Lever: Compute

  6. Verify · Preflight

    6Does testing before deployment predict deployed outcomes?

    Simulation, replay, sandbox or scenario tests calibrated against independent outcomes.

    Lever: Validation

  7. Plan · Decision quality

    7Do better predictions lead to better decisions within the constraints?

    Downstream results, constraint violations and whether the action could be carried out.

    Lever: Optimisation

  8. Measure · Feedback· Critical gate

    8Is feedback independent, timely and informative?

    Outcomes traced back; delays, selection bias and the comparison design examined.

    Lever: Measurement

  9. Act · Control· Critical gate

    9Can harm be prevented or contained in time?

    Enforced permissions, protective limits, stop, fallback and recovery, tested.

    Lever: Safety

  10. Validate · Transfer

    10Does performance hold beyond familiar conditions?

    Held-out sites, tasks, populations and disruptions evaluated.

    Lever: Generalisation

  11. Evaluate · Economics

    11Does each verified outcome create net value?

    Integration, review, rework, operating and failure costs included.

    Lever: Economics

  12. Govern · Expansion· Critical gate

    12What evidence earns more scope or authority?

    Release gates, change control, an incident owner and withdrawal criteria defined.

    Lever: Release

The next allocation experiment

Pick the binding constraint, then test one change against a baseline, with a threshold and a stop rule. A rising score is not a scaling law.

Every profile you assessed

Each profile is its own assessment. Compare only where scope and evidence standards match.

Nothing assessed yet.

Saved in this browser only, so you can come back to it. Export the file to keep it. Do not enter confidential information.

The rubric adapts a map of closed-loop AI systems by industry. NIST AI RMF 1.0 anchors the lifecycle monitoring; it does not validate this rubric. NIST AI RMF 1.0

Stage 5 · Prove → Work with us

And what it costs to start.

Work with us Free 30-min call