How AI is used here
This suite uses AI, and it tells you exactly where. The short version: AI explains, drafts and reviews. It never decides anything, never produces a number, and never touches a finding.
The dividing line
Every score, threshold, tonnage, trajectory and gate outcome in this suite is computed by deterministic code from data committed to the repository. Run the same inputs twice and you get the same answer, which is what makes a result reproducible for an assurer.
AI sits outside that. It is given a conclusion that has already been reached and asked to put it in words, or it is given your own text and asked to read it critically. It is never in the path between your inputs and a finding.
The five guardrails
The prompt carries only clause summaries and criterion text retrieved from our committed register. The model is instructed to use nothing else. We never invite it to draw on what it recalls about a regulation, because a half-remembered threshold is exactly the failure we cannot afford.
Every response must name the sources it relied on. Those identifiers are checked against the register. If any one of them does not resolve, the entire response is discarded and you are told why. We do not strip the bad citation and keep the prose: a plausible sentence attached to an invented authority is the worst possible output, and quietly repairing it would hide the signal from you.
Every figure in this suite is computed by a pure function from what you entered. Compliance-bearing responses are scanned for quantities, percentages and tonnages, and a response containing one is discarded. A model that volunteers a figure is either restating something we already have or inventing it.
Every response carries a grounding flag. When the model says the sources do not stretch to an answer, you see the refusal rather than a guess. We present that as a success, because a tool that treats abstention as a fault teaches its operators to suppress it.
Model-written text is labelled wherever it appears and stays out of your pack until you click accept. The acceptance is timestamped and travels with the text, and every contribution is listed separately in the pack so a reader can identify and discount machine-written text in one place.
What each feature may and may not do
- Plain-language explainer
Does: Explains what a provision asks for and what records would answer it, for a given role.
Cannot: Cannot say whether you meet it, and is given only the clause summaries for that criterion.
- Evidence request drafter
Does: Drafts an internal request for artefacts you have not ticked.
Cannot: Cannot set a deadline, assert that anything is overdue, or characterise anyone's performance.
- Claim wording review
Does: Reads wording you paste and points at phrases a reader might take further than the evidence goes.
Cannot: Cannot say anything is greenwashing, misleading or unlawful, and does not write claims for you. It reviews your text.
- Narrative drafter
Does: Describes findings the application has already computed, as a starting paragraph.
Cannot: Cannot recompute, re-evaluate or contradict a finding, and cannot restate a figure.
- Next step suggestion
Does: Suggests which tool to open next given where you are in a journey.
Cannot: Cannot comment on your progress, your performance or your organisation.
Model card
Intended use
Helping a person navigate published standards, find the artefacts a provision asks for, and read their own draft wording more critically. Always with a person in the loop.
Out of scope
Legal advice, financial advice, assurance, emissions calculation, threshold determinations, and anything that would be relied on without a person reading it first. Do not use this to decide whether to publish a claim.
Providers
Requests are served by whichever configured provider responds first, so two runs of the same feature can come from different models. Whichever served a given response is named on the response itself and recorded in your evidence pack.
Known limitations
- Grounding is enforced by prompt and by output checks, not by a retrieval index over primary text. The model sees our paraphrases of provisions, not the provisions.
- Our citation check confirms that a source exists in the register. It cannot confirm that the source actually supports the sentence attached to it.
- The quantity scan uses pattern matching. It will catch a fabricated tonnage or percentage; it is not a guarantee that no number ever slips through.
- The claim wording review finds nothing more than it finds. An empty result is not clearance, and it has read only the text you pasted.
- The register behind all of this covers 35 sources and is not exhaustive. State and territory regimes in particular are not mapped.
Where to raise a problem
If a response here is wrong, misleading, or cites something it should not, tell us. The prompt version recorded on every response is mining-2026-09-11, which is what we need in order to reproduce it.
The sources the assistant is allowed to cite are listed on the standards register. Anything outside that list causes a response to be thrown away.