Any decision can be explained, traced and contested.

Trust in a decision comes from being able to check it. Luminnovate records what was true when each decision was made, uses models whose logic can be read, and lets AI write only from evidence it can point to.

The decision record

The trace shows where a number came from. The record shows everything that was true when the decision was made, months or years later.

Harbour Joinery, credit decisionUW-2026-0917, 10 Sep 2026

The data snapshotThe rows used, as of the decision date, fingerprinted. Designed to be sealed so it can be proven later.
The signalsEach figure's definition and version, traced to the raw data.
The model and the policyThe model in force, with who approved it. The policy version is designed to be recorded beside it.
The decisionInputs, outputs, reasons and uncertainty. Overrides are designed to carry their justification.
The narrativeWhat the AI wrote, with every figure checked against the evidence.
The outcomeWhat actually happened, joined back to the decision.
ReplayDesigned to run the decision again later and give the same answer.

Illustrative example. Harbour Joinery is a synthetic business.

Explanations

The explanation is the model

For high stakes decisions on structured data, the final decision uses a model whose logic can be read. Complex models are used to discover signals, not to decide.

ScorecardChosen

Months trading	+18
Revenue against peers	+24
Undeclared debt	−40
Overdraft fees, 90 days	−22

Dense networkNot chosen

Illustrative example.

0.674Public lending data

The simpler model that matched

for the readable model, against 0.667 for gradient boosting (AUC)

What was observed: Two models built on the same loans: a scorecard style logistic regression and gradient boosting.
What a generic approach says: Use the more complex model.
What the engine read: The logistic regression scored 0.674 against 0.667: the readable model matched the complex one.
What it meant for the decision: The readable model was chosen, so the explanation is the model itself.

Reasons, and what would change them

Not only what drove a decision, but what would have had to be different for it to change. A decline becomes a list of evidence the customer can supply.

Harbour JoineryRefer

If both short term facilities were declared, with statements and payout figuresreassess
If the largest customer's share matched the 20% declaredconcentration clears
If a current company extract named the current directorsidentity clears

Illustrative example. Harbour Joinery is a synthetic business.

Judged against the right peers

A contractor's irregular income is normal for a contractor. Explanations that favour familiar variables can quietly exclude whole groups, so each customer is judged against businesses that operate the same way.

Against cafesAlarming

Against contractorsOrdinary

Illustrative example.

25 / 100Synthetic business

The contractor and the cafe

what a generic threshold scores 34 negative balance days in 180

What was observed: 34 days with a negative balance in the last 180.
What a generic approach says: A generic threshold scores it 25 out of 100: high risk.
What the engine read: For a contractor waiting on progress claims, that is normal. For a cafe settling card payments daily, it would be alarming. A collapse in trading across January matched an industry wide construction shutdown, and was expected.
What it meant for the decision: Each customer is judged against businesses that operate the same way, so an ordinary pattern is not read as risk.

AI writes. It never reasons.

Engines compute the evidence; AI writes from it; an automatic check rejects any figure the evidence does not support.

Grounding checkMemo draft

RejectedMonthly revenue is $128,000, comfortably above the request.

PassedMonthly revenue is $106,379 at the median, from customer payments only.

Illustrative example. Harbour Joinery is a synthetic business.

RejectedLuminnovate's own testing

The number that passed the check

any explanation with a figure the evidence does not contain

What was observed: AI writes the explanation for each decision.
What a generic approach says: Trust the text the model produces.
What the engine read: Every number in an explanation must appear in the evidence record, or the explanation is rejected. During development the check caught a hard coded number in our own writer.
What it meant for the decision: AI explains the evidence; it never invents it.

Explanations are tested like models

Whether reasons track the model, whether similar customers get similar reasons, and whether credit officers find them useful are each tested, not assumed.

Reasons track the model

Similar customers, similar reasons

Useful to credit officers

Illustrative example.

3.65% to 0.04%Anonymised invoice data

The collections fix that looked like a model

invoices paid 30 or more days late

What was observed: Invoices paid 30 or more days late collapsed from 3.65% to 0.04%.
What a generic approach says: A book wide change in payment behaviour.
What the engine read: 65% of the fall sat in one region, where 39% of invoices had run late and then none did.
What it meant for the decision: This corrected our own earlier reading of a book wide switch. Monitoring should show where a change happened before anyone explains why.

Regulatory context

Reviewed October 2026

Our clients carry the obligations; our job is to make them easier to meet. Every decision comes with its evidence, its reasons and its record, which is what a board, an auditor, AFCA or a regulator eventually asks to see. The platform is built to fit the obligations our clients carry under the rules below.

The obligation, and what we built for itObligations rest with the client

Who it applies to	Obligation	What Luminnovate provides
Banks, insurers, super funds	APRA CPS 234, information security, including suppliers handling their information	Software that runs inside the client's environment with no network access by default, and an evidence pack that returns only aggregate results
Banks, insurers, super funds	APRA CPG 235, managing data risk	Data lineage, every figure computed as of its date, and the decision record
Banks, insurers, super funds	APRA CPS 220 and CPS 510, risk management and governance, where model risk expectations sit	Model cards, validation on later periods held out from tuning, and monitoring against outcomes
Banks, insurers, super funds	The Financial Accountability Regime	Decisions an accountable executive can defend, each kept with its evidence
Consumer lenders	NCCP and ASIC RG 209, responsible lending and verification	Income and expenses read from transactions rather than trusted as declared
Consumer lenders	AFCA and ASIC RG 271, dispute resolution	Reasons, and what would have changed the decision, with a record that can be traced
Anyone using bureau data	Privacy Act Part IIIA, credit reporting	Bureau data used inside the client's environment, only for the client's permitted purpose under their contract
Everyone	Privacy Act and the Australian Privacy Principles, including the automated decision making disclosure that applies from 10 December 2026	Documentation of how each model works, written so a client can lift it into its own privacy policy
Open banking users	Consumer Data Right	Data received only within the arrangement it was collected under
Banks and payment providers	AML/CTF obligations and the Scams Prevention Framework	Account level monitoring for mule and scam behaviour, with alerts calibrated to the team's capacity
Everyone	Anti discrimination law	Proxies for protected attributes screened before a signal is used, and each customer judged against peers that operate the same way

AI governance

Australia has no AI Act. The reference points are the Voluntary AI Safety Standard's ten guardrails and the Department of Industry, Science and Resources' Guidance for AI Adoption, with ISO/IEC 42001 as the international management standard. Mandatory guardrails for high risk settings were proposed in 2024 and have not been legislated. For clients with operations in the European Union, the EU AI Act applies to those operations.

This section describes how the platform supports clients' own obligations. It is not legal advice, and it does not imply endorsement or approval by any regulator.

Governance, security and privacy

Where it runsInside your environment, with no network access by default. Evaluations return only aggregate results.
Who decidesA person with approval authority decides. AI drafts and recommends.
MaturityPilot ready, and tested on public and synthetic data. Not yet certified to SOC 2 or ISO 27001.
DetailSecurity and governance and the privacy policy.

Bring your model risk or security checklist to the first conversation.

Book a conversation