Award compliance engine · Australian employment law
Award answers
built to be defended.
FairPay prices every pay period against the modern award: the Commission’s data, the Ombudsman’s guides, the award’s own clauses. And it shows the working. Deterministic where it must be, AI where it helps, provenance on every line.
- Award resolvedGeneral Retail Industry Award 2020MA000004
- ClassificationRetail Employee Level 2 · casualconfirmed on the matter
- Casual ordinary rate$34.76 / hrMA000004 cl 17.1
- Hours assessed60.00 ordinary hourspayslip · 02–15 Mar
- Entitled vs paidshortfall $114.401 period assessed
How the engine is built
Four rules. Breaking one is a design decision,
not a shortcut.
- 01
The engine is pure
No network, no clock, no mood. The assessment date is an input, so a 2023 timesheet prices at 2023 rates: today, next year, and in front of a tribunal.
- 02
Provenance is required
Every entitlement line carries the award code, the clause, the effective-dated rate and the source record. Not optional: findings must be defensible in a dispute.
- 03
Documents come first
Where the Commission's pay database disagrees with the award's own text, the text wins, and the corrected row says so on its face.
- 04
Refusal over guess
Ambiguity refuses loudly. Missing evidence is a named gap, never a silent zero, because in underpayment work, absence of evidence is where the money hides.
One matter, end to end
From “what award are they even on?” to the letter of demand.
Ask FairPay
Award questions, clause-cited answers
Chat over the award corpus with the working shown: every answer carries a trail of the lookups behind it, figures superscript to the tool result that produced them, and a degraded search says so instead of bluffing.
Check a Rate
Base, penalties and loadings in seconds
Classification to rate card with the clause chip on every penalty row, and the Ombudsman's own pay guide opened to the exact cell, with the figure boxed on the page it was printed on.
Calculators
Work in progressEntitlements, remedy and underpayment
Per-period entitled vs paid vs shortfall in integer cents, evidence gaps named instead of papered over, and a CSV of the working for the file. Verified tier only when every input is confirmed.
Matters
Work in progressUnfair dismissal & general protections
Evidence in, chronology out. Guided workflows for unfair dismissal and general protections dismissal applications, plus the response side, with statutory deadlines and eligibility computed, never remembered.
Contract Review
Work in progressThe employment contract, marked up
A playbook pass over the uploaded contract: graded findings in the document itself, suggested amendments as tracked changes, native Word comments for the other side, and a round-trip check of what they actually adopted.
Radar
When the award moves, you hear it
The Commission and the Ombudsman are monitored continuously; award text is versioned, superseded pay-guide editions are archived, and a rate change lands as a dated, sourced update, not a surprise.
Built for the Commission’s own forms
The data path
From the Commission to the courtroom, without a copy-paste in between.
- 01
Mirror
The Commission, nightly
- All six pay-database entities row-hash diffed into Postgres
- Deletions two-strike and audited, because the feed changes rows without saying so
- FWC webhooks trigger syncs between the nightly runs
- 02
Parse
Documents become data
- Award PDFs to 37,793 page-anchored clauses
- Pay guides line by line, values kept verbatim, never invented
- 489 FWO guidance articles mirrored, sha-verified
- 03
Resolve
One answer per figure
- Canonical layer: override › award text › pay guide › pay database
- Effective-dated windows, real dates, half-open
- Disparities kept queryable, never averaged away
- 04
Price
The pure engine
- Assessment date is an input: 2023 timesheets price at 2023 rates
- Integer cents, zone-aware penalty windows, DST-safe
- Ambiguity refuses loudly instead of guessing
- 05
Serve
Every surface, cited
- Chat, Check a Rate, calculators, matters, contract review
- Each figure carries clause, effective date and source
- Degraded lookups say so on screen, never silently
Compiled, mirrored, versioned
The corpus under the answers
Measured on the live mirror, August 2026. Where a source disagrees with another, the disparity is kept queryable, not averaged away.
The committed gates
Measured, not promised.
Every capability above sits behind a benchmark committed to the repository. A change that drops a gate does not merge.
- Chat answers eval26/26prompt v13, committed baseline
- Award recall42/45coverage benchmark, found
- Boundary probes730/737coverage-deciding facts
- Practitioner acceptance35/36the Head of Employment Law's own queries
- Entitlement topic map38/38ten pilot awards
- FWO guidance recall14/14plain-language corpus
- Guide figure boxing45/45MA000004: every figure placement-verified
- Applicability probes9/9statutory frame, tricky set
- Classification method3/3live behavioural probes
Gates run in CI or as operator evals against committed baselines, measured August 2026. Baselines are never lowered to make a run pass.
Built for the file, not the demo
Trust is an architecture, not a promise.
Eval-gated releases
Committed recall and answer baselines gate every change. Baselines are measured, and never lowered to make a run pass.
Firm-locked sign-in
LEAP single sign-on with an operator-managed firm allowlist. Every store is owner-scoped inside the SQL, so a guessed id behaves like a missing one.
Privilege stays private
Client documents never enter logs, error messages or transcripts. Deleting a matter destroys its bytes, not just its rows.
Sources on their own terms
Commission data, award text, FWO guides and guidance are attributed to their authors, effective-dated, and quoted verbatim or not at all.