Practical guide

AI programme status report verification checklist

A programme status report should be verified as a decision instrument, not proofread as a piece of presentation. A report can be numerically accurate and still be unsafe to use if the reporting period is unclear, a green label has no rule, a critical exception is averaged away, or the requested decision belongs to someone else. This checklist uses the LOCKED method: lineage, operating rules,.

By Two Prune

A programme status report should be verified as a decision instrument, not proofread as a piece of presentation. A report can be numerically accurate and still be unsafe to use if the reporting period is unclear, a green label has no rule, a critical exception is averaged away, or the requested decision belongs to someone else. This checklist uses the LOCKED method: lineage, operating rules, claims, key exceptions, escalation, and decision record.

Source basis and limits

The current UK Government Functional Standard for Project Delivery defines expectations for governance, roles, planning, control, transition, use, and disposal. The Orange Book treats risk management as part of governance and decision-making and calls for structured communication of uncertainty. The AQuA Book emphasizes traceable data, assumptions, limitations, changes, decisions, and proportionate assurance. These sources support the verification controls below; they do not certify a report, prescribe one universal status scheme, or transfer accountability to an AI tool. GovS 002: Project Delivery The Orange Book The AQuA Book

Sources: [1] [2] [3]

How to use the checklist

Run the checklist after the report author has assembled the evidence but before the governance pack is frozen. Assign a verifier who is independent of the material status judgments where proportionate. Mark every item pass, repair, not applicable, or not verified. A blank item is not a pass.

Create a verification ledger:

| Check ID | Report claim or field | Authoritative source | Test performed | Result | Repair owner | Evidence retained | | --- | --- | --- | --- | --- | --- | --- | | | | | | | | |

Do not paste sensitive source material into an unapproved AI service. Where AI helps compare documents, retain the source IDs and human disposition rather than treating the model output as verification.

L — Lineage and reporting boundary

Select three material claims and walk each one backward from the report to the source record, then forward from the source to the report. A broken route in either direction is a repair, even if the sentence looks plausible.

  • [ ] The programme, reporting period, data cut-off, report version, author, reviewer, and approver are named.
  • [ ] Scope includes the outcomes, workstreams, dependencies, transition activities, and exclusions relevant to the decision audience.
  • [ ] Every material figure and factual statement points to an authoritative record with an owner and refresh date.
  • [ ] The population, denominator, unit, currency, time basis, and calculation route are explicit where needed.
  • [ ] Evidence copied from an earlier report was reopened and checked, not assumed to remain current.
  • [ ] Not available, not assessed, and not applicable are used distinctly.
  • [ ] Access restrictions and retention requirements are respected in the report and verification record.

O — Operating rules and status logic

Recalculate at least one status from the approved rule. If the reported label cannot be reproduced, record the discrepancy instead of choosing the label that seems most reasonable.

  • [ ] Status labels, tolerances, trends, and confidence levels were defined before reviewing the current results.
  • [ ] Delivery, operational readiness, use and quality, and benefits or outcomes are not collapsed into one colour.
  • [ ] A green status does not depend on silence, missing evidence, or an untested assumption.
  • [ ] Forecast dates and costs show their assumptions and a range or scenario where one value would imply false precision.
  • [ ] Changes to baseline, scope, schedule, cost, benefit, or quality have a named authority and downstream effect.
  • [ ] The report distinguishes assumption, risk, issue, dependency, opportunity, and decision.
  • [ ] Confidence reflects evidence quality and recency rather than management optimism.

C — Claims and calculation checks

For each repaired claim, preserve the original draft, the reason it failed, the corrected text, and the human who accepted the change. Do not let AI silently overwrite the audit trail.

  • [ ] Arithmetic is independently reproduced from the retained input snapshot.
  • [ ] Totals reconcile with component records and duplicate or excluded items are explained.
  • [ ] Percentages name their denominator and missing records are visible.
  • [ ] Milestone completion is not presented as evidence that an outcome or benefit occurred.
  • [ ] Association is not rewritten as causation.
  • [ ] Comparative claims use comparable periods, populations, and definitions.
  • [ ] AI-generated summaries, citations, classifications, and explanations were checked against the underlying records.
  • [ ] Unsupported adjectives such as on track, minor, nearly complete, or low risk are replaced with evidence and a rule.

K — Key exceptions and control evidence

Use a critical-exception table when an average could conceal a material problem:

| Exception | Affected scope | Consequence | Current control evidence | Owner | Required decision | | --- | --- | --- | --- | --- | --- | | | | | | | |

  • [ ] High-consequence exceptions remain visible even when an aggregate result is favorable.
  • [ ] Control operation is evidenced, not inferred from the existence of a policy or owner.
  • [ ] Accessibility, privacy, security, financial, legal, operational, and service concerns are routed to the relevant specialist.
  • [ ] Conflicting source values remain identified until an authorized owner resolves them.
  • [ ] Stale, partial, sampled, or supplier-provided evidence is labelled with its limitation.
  • [ ] Residual uncertainty and accepted risk have an accountable owner and review date.
  • [ ] Minority or dissenting specialist views are not removed merely to create a single narrative.

E — Escalation and ownership

[ ] Every escalation names the condition breached or forecast to be breached.

  • [ ] The accountable decision owner is distinguished from advisers, reviewers, and delivery staff.
  • [ ] Options include consequences, uncertainty, reversibility, and the effect of delay.
  • [ ] Requested decisions are within the named owner's authority.
  • [ ] Specialist approval is obtained where the decision crosses a specialist mandate.
  • [ ] The report states the next evidence gate and the trigger for revisiting the recommendation.
  • [ ] No employment, compliance, safety, or outcome decision is delegated to an AI-generated recommendation.

D — Decision pack and retained record

Verification is incomplete if the meeting receives a corrected report but the retained record cannot show what changed or why.

  • [ ] The opening summary matches the detailed evidence, including adverse findings.
  • [ ] Links, appendices, and source identifiers work for the intended reviewers.
  • [ ] The version presented to the governance body is the verified version.
  • [ ] Repairs, unresolved items, waivers, and approver dispositions are recorded.
  • [ ] Minutes capture the decision, conditions, dissent, owner, and review date.
  • [ ] The report, source snapshot, calculation record, verification ledger, and decision record are retained under applicable policy.

Worked verification: fictional service transition

A draft programme report says the transition is green because 96% of records reconciled, training is complete, and the software milestone is on time. The verifier samples the lineage behind each claim.

The 96% figure reproduces, but the unmatched 4% contains nearly all records for a high-consequence case type. The aggregate passes arithmetic but fails the key-exception check. “Training complete” means invitations were issued; attendance and task completion are not yet available. The milestone is accurate but the approved readiness rule also requires keyboard-only testing, support ownership, a rehearsed fallback, and reconciled high-consequence records.

The report is repaired to separate delivery from readiness. Delivery remains on plan. Readiness becomes red under the approved rule, with medium evidence confidence because a supplier control record is dated. The decision table gives the senior responsible owner two bounded options: delay transition for two weeks, or exclude the affected pathway and retain the legacy queue. Accessibility, service operations, and information-governance reviewers record their specialist positions.

An AI comparison had described the remaining reconciliation as “minor.” The verifier rejects that wording, records the disposition, and retains the source IDs. The final pack asks for a decision and a revisit trigger rather than merely requesting that the board note progress.

Two Prune boundary

Two Prune can place a participant in a realistic assessment where they inspect a programme report, trace supplied evidence, use AI within stated controls, reproduce calculations, find concealed exceptions, and produce a verification ledger. It reports observed assessment performance and evidence confidence. It does not certify programme status, accept organizational risk, replace specialist assurance, predict delivery outcomes, or make employment decisions.

Sources

UK Government Functional Standard GovS 002: Project Delivery

  • HM Treasury: The Orange Book
  • GOV.UK: The AQuA Book

Sources: [1] [2] [3]

Sources

  1. 1.GovS 002: Project Delivery · UK Government
  2. 2.The Orange Book · UK Government
  3. 3.The AQuA Book · UK Government