The Annual Review

Behaviour programmes accumulate settings, exceptions and habits. A day a year is enough to see whether they still describe what was agreed.

The safety outcome

Collisions and severity over three years, by type, because one year is noise.

Vulnerable road user incidents separately.

Low-speed manoeuvring damage separately, since it responds to different interventions.

Near-miss reports and whether they held.

And the honest attribution question: what else changed — fleet, routes, volumes, staff turnover?

The system

False trigger rate, trended.

Thresholds: still appropriate for the current vehicle mix?

New vehicle types added since last year, and whether they were retuned.

Calibration currency across devices.

Whether the vendor changed the scoring model, which invalidates trends and which they may not have announced.

The process

Time from event to coaching, trended.

Disputes and outcomes.

Structural findings identified against acted on, which is the measure of whether the programme is one-directional.

And access log review: who looked, at what, and did anyone check.

The boundaries

Has any feature been enabled that was not part of the original justification?

Has the data been used outside the stated purpose, once?

Is retention still what was set, per category?

Is the score still explainable to the person it describes?

The decision

One question: would we deploy this the same way again?

A yes with specifics is a programme working.

A yes because it is already there is a subscription carrying obligations.

And a no is worth saying, because switching parts of it off is available and is almost never considered.

What to report

One page: safety outcomes, what the programme found about the operation, what changed as a result.

Plus one honest statement about what it has not demonstrated, which is what makes the rest credible and is the section most often omitted.

Check whether the model changed

The thing vendors alter without announcing.

A scoring model updated mid-year invalidates every trend that spans it.

Ask annually whether weights, thresholds or detection logic were changed, and require notification in the contract.

A trend line across an unannounced model change is a chart of two different measurements, presented as one.

Reproduce the workflow

For another way to make this requirement testable, consult inspect the service-desk example. Reproduce the case with real roles, failures and recovery steps.

Independent reference

For a thematic point of reference, see the Health and Safety Executive. Consult the source directly because technical, operational and legal details change.