We run the detection. You keep the decision.
The capability runs as a managed service: collection on a schedule, one operator resolved across registers, scored and explained cases, an analyst decision on each one, and an assurance record a reviewer can follow.
Intelligence, not evidence
Every output answers one question: is this worth investigating? It never answers the other one. We do not determine that an operator has committed fraud, and we do not ask you to accept that we have.
That line is deliberate. It keeps us outside your statutory decision-making, so the determination and the liability stay with the agency that holds the power to make it. It also means every referral has to carry its reasoning, because it will be read by someone deciding whether to act on it.
A closed loop, not a delivery
Outcomes from investigations return to the models. That is the step that is almost always missing, and it is why a capability that worked in its first month stops producing by its sixth.
What we operate
Five stages, each one observable. You can ask at any point what ran, what it found, and what an analyst did about it.
Collection
Agents pull public registers, regulatory actions and corporate records on a schedule, rather than when somebody remembers to look.
You get a source inventory: what was collected, how much of it, and when it was last refreshed.
Entity fusion
One operator is resolved across registers by ABN, name similarity and shared address — including trust structures that carry no public director record.
You get a single view of an entity, with its aliases and its related parties.
Scoring
Documented, weighted rubrics, one per typology. Every score breaks down into the components that produced it.
You get ranked flags, each carrying the signal that raised it.
Analyst triage
An analyst reads the flag, works through the components, and decides whether it is worth an investigator’s time.
You get referrals with a named decision behind them, not alerts.
Governance
Bias testing, model-version-stamped logging and retention run alongside the other four stages, not after them.
You get the assurance record: what was decided, by whom, against which version of which model.
What one pass over public registers produces
These figures come from our own analysis of published Australian registers in September 2026. They are not the results of a client engagement, and they are not a recovery claim.
Entities resolved across the disability, aged care, corporate and charity registers in a single pass.
Registered providers scored against a documented phoenix-risk rubric, with the components shown.
Related-entity pairs surfaced for triage, from roughly 4.4 million pairwise comparisons.
Related-entity review across 2,980 registered providers is roughly 4.4 million pairwise comparisons. It completes in under two minutes. It has never been run as a routine compliance check.
Assurance and measurement
A company selling detection to a government agency should expect its own work to be audited. The evidence trail is part of the service, not an add-on.
- Bias testing runs fortnightly across six dimensions: rurality, provider size, specialty, state, and practitioner age and gender. A disparity ratio of 1.25 triggers review; 1.5 is treated as a breach and stops the model.
- Every referral is logged with its model version, its confidence, the components behind it and the analyst who actioned it. The log is immutable and held for seven years.
- Precision and recall are tracked per model version, and reported to you whether they improved or not.
This is what lets an answer hold when it is asked for at estimates, under FOI, or by an auditor reading backwards from a decision.
What each side brings
You provide
- A named contact who can say whether a referral was useful.
- A route for referrals into the investigation workflow you already use. We do not ask you to work in our tool.
- Program data, once we move past public sources, under an agreed data-sharing arrangement.
- Outcomes, so the models learn from what your investigators actually found.
We deliver
- Ranked referrals, each with its score components and a specific next step.
- The entity picture kept current, rather than rebuilt for each request.
- The assurance record, and the bias testing behind it.
- Documentation and training, from the first week, so the capability can be handed to your team.
How it is priced and measured
A proof of concept is a fixed fee for a defined question, a defined dataset and an agreed measure of a useful answer. The managed capability is a base subscription with part of the fee held at risk, paid against service metrics you set, you weight and you measure. We would rather be paid on whether the capability works than on how long we spent running it.
A base fee buys continuity. It pays us to keep the capability running, and costs us nothing if it stops being useful to you. A component that moves with your measures pays us to improve it, and keeps our attention on your program in year three rather than on whatever else might grow our revenue.
We are not paid a share of what is recovered, and not paid per referral. Both would pay us to push more work at your investigators whatever its quality, and both would put a commercial interest inside a referral that has to survive review.
Holding part of a fee at risk is awkward for a Commonwealth buyer. The money is committed when the contract is signed either way, so the question is whether an unearned component becomes an underspend to explain or an earned one arrives without headroom. We set out three arrangements and which is easiest to approve rather than leaving it to the first year it matters.
Fees are quoted after a scoping conversation and fixed before work starts. We will send indicative figures on request.
We report monthly, and the same measures settle what is at risk:
- Referrals accepted by your investigators, as a share of referrals made.
- Administrative and compliance actions that began with a referral from us.
- Movement in an agreed typology, against a baseline set before we start.
- New typologies surfaced and validated by your analysts.
- Time from a signal appearing in a public source to a referral reaching you.
- Cost per accepted referral, and disparity ratios against their thresholds.
We do not report value recovered. We have not measured it, and a number we cannot evidence is worse than no number.
Where the capability is running
Cross-register provider integrity
Finding operators who are non-compliant in one payment system while still being paid by another, using public registers only.
Where the same capability applies next
Both of these use the machinery that already runs. Neither is in service yet, and we will say so in the room rather than imply otherwise here.
- Procurement and supplier integrity. The same entity resolution applied to supplier registers and contract data instead of provider registers.
- Quality and safety early warning. Outlier detection on published quality indicators, ratings and coronial findings, to surface a service that is drifting before a complaint arrives.