Truthound · All articles · 5/24
Evidence-led Truthound Core guidance on data contracts, check inputs and operational limits.
480 articles
Do not project a current contractor suspension back onto every award date
Current eligibility, eligibility at award and an allegation’s disposition need separate evidence and review owners.
Align token classes and price baselines before comparing public-sector AI usage costs
An offer’s discount becomes interpretable only after its model, token class, environment and price baseline are defined.
Distinguish visitor change from a tourism data-source revision
A revised historical series and a change in actual visits are different explanations that need a fixed release vintage.
Separate census response progress from cleared statistical records
A collection milestone and readiness for statistical release have different admission conditions.
A plan for 59 surveys is not 59 released datasets
Statistical modernisation needs separate evidence for approval, execution and release before progress can be evaluated.
Compare experimental service indices without losing base year and revision state
A service-industry comparison needs its base year and provisional or revised state attached to each value.
Keep women, self-help groups and enterprises distinct in community-finance data
A useful linkage should remain explainable to the people whose activities generated the records.
Keep catalogue and evidence references behind a data-readiness dashboard
A ready label becomes interpretable only when it points to a particular dataset and a defined confirmation.
Preserve approved exceptions when emergency AI access ends
An ordinary closure and an approved extension need different records and handover responsibilities.
Count authorities and departments separately in a local-government AI pilot
A new-use experiment needs a defined experimental unit before its application count can be interpreted.
Are 110 reported minutes saved per week measured telemetry or a survey estimate?
A time-saving estimate derived from response bands needs the original answers, population definition and conversion assumptions to remain interpretable.
Retain source versions and human disposition in AI-drafted parliamentary answers
Drafting speed does not remove the lasting responsibility to explain which source and judgement supported an official claim.
Teach government AI-agent responsibility across users, operators and developers
People using the same agent still have different evidence and authority obligations.
Keep observed deaths separate from attributed climate-health estimates
Identical units do not make observations, exposure measures and model-attributed burden interchangeable.
Preserve orbit epoch and coordinate-frame meaning in space-awareness inputs
A larger observation archive still needs evidence that records refer to compatible objects and time meanings.
Which department reviewed the source record used by a public-service AI answer?
Public-service guidance needs a route to the source and responsible institution as well as a generated answer.
What access route should remain when a public-service AI tool is unsuitable?
A larger project count does not establish the conditions under which people can use the resulting service.
What should an SME be asked when cross-border trade evidence does not link?
A failed declaration-to-application link is an evidence question, not an automatic defect in the applicant.
Prepare entity-role evidence before choosing a data-centre licensing subject
A facility label and supplier name do not determine the responsible legal entity; declared roles need separate corporate evidence.
Does a coding-agent assessment omit a required trust boundary?
A populated assessment matrix is not proof that every required boundary was assessed; an expected inventory must come first.
