What we record about every AI action

Recorded for every AI action

• Provider and model, as identified at the time of the call
• Action type — ask, analyze, brainstorm, create, rewrite, optimize, audit, translate, repurpose
• A reference to the instruction, stored internally
• The source facts and content assets that went into the context packet
• The generated output as a diff against the current version
• The human who edited it afterwards, and what they changed
• The reviewer and the approval, as separate events

Deliberately not exposed publicly

Provenance is an internal audit record. Publishing all of it on a customer's website would leak their private context.

• Prompts and instruction text
• The retrieved source content in a context packet
• Internal fact evidence links
• Anything belonging to another tenant, under any circumstances

Why the separation matters

AI Suggested is not Draft

A reviewer reads a machine draft differently from a colleague's draft, and should. Folding them into one state removes exactly the signal that determines how carefully the text gets checked.

Generated is not approved

The model produced it, a person edited three sentences, a second person approved it. Three actors, three events. A single 'last modified by' field would attribute all of it to whoever saved last.

Inputs matter more than output

When a claim is later disputed, the useful record is which facts and which existing assets the generation was working from — not the wording of the prompt.

Committed is not written

For imported content, Git tells you who committed a line. That is evidence about authorship, not a determination of it, and Narraxio records it as 'introduced in commit' rather than as a byline.

Risk classes

Not every AI action deserves the same ceremony, and pretending otherwise means the important ones get rushed.

• Low — meta descriptions, internal link suggestions, brainstorming
• Medium — homepage copy, feature descriptions, translations
• High — pricing, legal statements, regulatory and compliance claims, product availability, comparative claims

Training

Our intent is that customer content is not used to train shared models without an explicit customer agreement, with enterprise controls over which providers and models may process content at all.

Where that intent rests on a provider contract rather than on something enforced in our own architecture, the Security page says so. A vendor selling provenance should be the last one to blur that distinction.