How do I build a normal weekly earnings evidence pack?

The trigger

An incapacity claim is accepted, or a pay variation, a change in hours or a payment query forces a recalculation. Before anyone opens the section 8 formula, the file has to show which records exist, which dates they cover and where each figure came from. Most errors here are evidence errors wearing the clothes of arithmetic, so the pack comes before the calculation.

The inputs

• The claim record: [CLAIMANT_NAME], [CLAIM_NUMBER] and [INJURY_DATE] as recorded on the claim form and the medical material
• Payslips covering the apparent relevant period, plus a wider payroll transaction export
• Rosters, and any overtime approval or authorisation log
• The employment instrument [EMPLOYMENT_INSTRUMENT] covering allowance and overtime terms
• A record of absences, leave and pay-rate changes around [INJURY_DATE]

The steps

1. De-identify every record before it leaves the claims system, using the placeholder set in the de-identification toolkit.
2. Run the prompt below to produce four linked layers: a period coverage map, a pay-element register, an authority trail and a list of open questions.
3. Open each cited record and compare it with the extraction. Anything without a file and page, row or cell reference goes to the gaps register, not the evidence table.
4. Make the legal classifications yourself. The pack leaves blank whether overtime was required and regular, and whether an allowance qualifies.
5. Select the relevant period, record why it was selected, then calculate. Where the figure carries downstream payments, run the arithmetic through a separate incapacity cross-check.
6. Issue the determination and notice through the normal human determination process, identifying the period used and summarising the data considered.

The prompts

You are preparing an evidence pack for a human decision-maker
under sections 8 and 9 of the Safety, Rehabilitation and
Compensation Act 1988. The material is de-identified.

Use only these placeholders: [CLAIMANT_NAME], [CLAIM_NUMBER],
[INJURY_DATE], [PAY_PERIOD_START], [PAY_PERIOD_END],
[EMPLOYMENT_INSTRUMENT].

Records: [PASTE DE-IDENTIFIED RECORDS]

Produce:
1. A period coverage map: the date of injury, the apparent
   two-week period before it, every overlapping pay cycle, and
   any absence, leave or pay-rate change touching those dates.
2. A pay-element register: ordinary hours, ordinary rates,
   overtime hours, overtime rates and each allowance, worded
   exactly as the records word them.
3. An authority trail: for each overtime entry and each
   allowance, the roster, approval, or instrument clause said
   to support it, and its stated purpose.
4. Open questions: missing records, conflicting sources and
   unexplained pay codes.

For every entry give the source file and page, row or cell.
Do not total or average anything. Do not apply the section 8
formula. Do not nominate a relevant period. Do not decide
whether overtime was required and regular, or whether an
allowance qualifies. Write "human classification required"
wherever legal characterisation is needed.

The checks

• Every factual entry carries a source location, and unsourced items sit in the gaps register rather than the evidence table
• The period map follows the statutory period, not the payroll cycle label
• Absences, leave and pay variations are visible on the map rather than averaged away
• Each overtime entry names the roster, approval or contract term said to authorise it
• Each allowance names the instrument clause and the purpose the record states
• No total, average or calculated figure appears anywhere in the output
• Where a record says overtime, the pack does not say regular overtime

Governance controls

• The delegate owns the decision. A person selects the relevant period, classifies every input, performs the calculation and signs the determination. AI extracts, reconciles, sources and flags.
• De-identification is mandatory. Names, claim numbers, dates of birth and diagnoses do not leave the claims system unmasked, and placeholders are re-attached only inside the system.
• Keep the hand-off free of arithmetic. The output schema carries a human classification field, never an included-in-earnings field, because a model given that column will fill it.
• A decision on earnings is a determination, so the notice, the reasons and the reconsideration and review rights follow the normal process. The pack supports the reasons and never supplies them.
• Record the AI's role and the verification step in the file note, so the evidence trail shows who checked what against which record.

The output

A source-linked evidence pack in four layers, with a gaps register beside it, filed with the claim: enough for a delegate to select the period, classify each input, calculate, and write reasons that trace back to records.

Time it takes

A first run takes about an hour while the record set and prompt are assembled. A practised run takes about 35 minutes, most of it a human comparing extractions against the original records.