Optimisation improves execution quality without quietly rewriting intent.

Once governance is operational, improvement becomes continuous: reduce violations, increase alignment, and shorten time-to-correction.

That distinction matters. In weak AI programmes, “optimisation” often becomes a justification for changing outputs until they look good in the moment. In a governed model, optimisation should improve how reliably the system executes the intended policy, not quietly rewrite the policy itself.

What to optimise

• Evaluation thresholds and scoring
• Test coverage for new channels/use-cases
• Escalation rules and ownership
• Policy clarity (remove ambiguity)
• Retrieval quality for approved knowledge
• Cost and carbon efficiency where model choice affects operations

It is also sensible to optimise the surrounding operating model:

• how quickly incidents are triaged
• how easily teams can identify the right policy bundle
• how much manual review is still needed
• how often recurring exceptions can be eliminated through better specification

What not to optimise

• Core intent (that belongs to strategy)
• Legal constraints (those are non-negotiable)
• Approved claims without a formal policy change
• Human accountability for high-risk decisions

Why feedback loops matter

The value of optimisation comes from shortening the distance between observation and correction. If assurance detects drift but nothing changes upstream, the organisation has measurement without control. The point of a feedback loop is to move evidence back into mapping, specification, testing, and release.

This is why optimisation belongs inside the Brando lifecycle rather than outside it. It is not a postscript. It is part of how a governed system becomes easier to run, safer to scale, and more dependable in live work.

What a useful feedback loop contains

A good feedback loop connects four things:

1. Signal: what happened in production?
2. Diagnosis: which rule, asset, model, or workflow caused it?
3. Correction: what definition or control needs to change?
4. Release: how is the corrected version tested and deployed?

That loop keeps optimisation honest. It prevents teams from chasing isolated outputs and instead improves the system that creates them.