Citations, refusals and the "Why this draft" panel
Every claim an Axis Assure AI draft makes cites a record of your workspace. A claim that cites nothing, cites something it was not given, or quotes words its source does not contain is dropped before you see the draft. When too little is left, or your records do not support the request, you get no draft at all — you get what is missing.
TL;DR — The markers [S1], [S2] … in a draft are its sources. Open one to see the record, the version, the exact passage quoted and a link to the record. Why this draft shows what the AI was given: the framework, version and profile, your parameter values, the components, which framework words were withheld by their licence, the instructions' version, the model, and any instructions found inside your sources.
Where you see them. In Axis Assure: the AI draft of a statement in a requirement's details on Controls; the AI review of a submitted item in an evidence request's drawer on Evidence Requests; the AI mapping suggestions waiting on Crosswalk; and the copilot's answers. Each AI draft there shows its markers, Why this draft? and what it rests on.
How a draft is checked before you see it
- What the AI may read is chosen before it is asked: the records in scope of the request that you may see — and, for evidence, only a current version that passed the malware scan and that you may open. Where a redacted copy of a document exists, the AI reads the copy, never the restricted original.
- Each source gets a short marker (S1, S2, …). The AI never sees your records' identifiers.
- Every claim is held to its markers. A claim with no marker, a marker for a source the AI was not given, or a quote that is not word for word in the cited source is dropped. A value outside the list the request offered — a cause, a relationship, an answer choice — is dropped too, or, where the value is the draft's point, the whole answer is refused.
- More than half dropped, or nothing left: no draft.
- Links, images and formatting code are removed from every word of the draft.
When the AI refuses
You see Not enough evidence to draft this and what is missing. Nothing is saved as a draft. A refusal happens when:
- there was nothing in scope to ground a draft on (the AI is not asked at all);
- the AI answered that the sources do not support the request;
- too many of the AI's claims failed their citations.
A refusal is the system working: the AI does not write compliance text from general knowledge.
Citation markers
Open a marker to see:
| Shown | What it means |
|---|---|
| The record's title and kind | Evidence, a policy, an implementation statement, a framework requirement, a finding, a POA&M item, a test result, … |
| Version and date | The version the AI read, dated on your workspace's clock. A newer version exists when the record has moved on since. |
| The quoted passage, and its page when known | The exact words the claim rests on. |
| Open source | A link to the record in Axis Assure. |
| Text withheld | The source is a framework requirement whose licence does not allow its words here: the AI saw only its identifier, and nothing is quoted. |
| Source unavailable to you | You cannot see this record — someone with more access asked for the draft, or its access changed since. Checked each time you open the draft; no title or kind is shown. When a quote in the draft comes only from files you cannot open, the quote is hidden too and the draft says so. |
The "Why this draft" panel
| Section | What it shows |
|---|---|
| Framework | The framework, version and profile the draft was written for |
| Requirement | The requirement's identifier and label |
| Parameter values used | Your organization-defined values the draft was given — values only, never the framework's words |
| Components | The system's components the draft speaks about, and which are inherited from a provider |
| Words withheld from the AI | Requirements whose licence keeps their words out of an AI prompt, by identifier |
| Instructions | The version of the instructions the AI was given (they change only as a new version) |
| Model | The provider and model that answered, and whether a fallback model was used — or that no model was called |
| Instructions found inside sources | Any source — or part of the request, such as a customer's question — whose text reads like instructions to the AI, and what it looked like (telling the AI to ignore its instructions, to act as someone else, to send data elsewhere, …). The draft is still shown, marked for you to check, and is never offered as a suggestion card |
What a draft rests on
Each draft shows one of three labels, from the kinds of records it cites — never a number:
| Label | Meaning |
|---|---|
| Grounded in your evidence | It cites your evidence. |
| Framework text only | It cites only a framework's words. |
| Not grounded | It cites only your own descriptions (a system, a component, a control). |
How a mapping suggestion compares with the publishers' own mappings
Each AI mapping suggestion in the crosswalk's review queue shows how it compares with the mapping sets the framework publishers themselves issue (official tables, NIST's informative references, a publisher's own crosswalk) — decided by Axis Assure from those sets when the suggestion is made, never by the AI, and never a number:
| Label | Meaning |
|---|---|
| Matches an official mapping | A publisher's own mapping relates your control's requirement to this one the same way the AI does. |
| Contradicts an official mapping | A publisher's own mapping relates them, but differently from the AI. Check it closely. |
| No official mapping to compare | No publisher's mapping relates them: the suggestion came from your own or AxisSynapse's mapping sets, or from similar words. |
Every suggestion still needs its own decision by someone other than the person who asked for it.
What is never shown
No percentage, no confidence score, no "accuracy". What a draft rests on is shown as its sources; how well the AI does is measured offline and published as an evaluation summary — see Assure → Reports & Setup → About AI.
Related
Related