Sense and decide
Provenance by Default
Every piece of AI-assisted work arrives marked with what the model produced, what it drew on, and how confident it was.
Developed by Superadditive in our own review of the research. There is no outside source to credit.
Changes a rule or role · An hour or more to set up · Anyone on the team can start it
Try this first
Start with one requirement. Mark which parts came from the model. A reviewer then knows where to look hardest.
Use it when
People are signing off on AI-assisted work without knowing which parts the AI wrote or how sure it was.
Skip it when
Trivial output does not need it and people will ignore the requirement anyway. Spend it where someone is accountable for the result.
How to introduce it
Set the standard that AI-assisted work arrives with three things attached. What the model generated. What it drew on. How confident it was. Whoever owns the outcome cannot evaluate the work without them.
How to show up
Agree a light standard for what gets recorded and make it the default rather than an extra step. Enough to judge the work, not so much that people route around it.
How long it takes
Some upfront work to build it into tools and templates. Then it costs little per use.
What makes it hard
It feels like admin and dies under deadline, so build it into the tools and templates instead of relying on discipline. Capturing the model's real confidence is the hard part, because fluent output always sounds sure of itself.
What it looks like when it's working
Reviewers can see what the model contributed and how sure it was, and they catch weak spots they would have missed. Work still arriving as a seamless block means the standard did not stick.
How long until it sticks
Expect a few weeks of use before recording provenance feels normal rather than like extra work.
How you know it stuck
Work arrives marked and reviewers expect it without asking.
The idea behind it
Fluent output buries its own origins and its own uncertainty. Without the record, whoever signs off is trusting blind.
Where it comes from
our review. Fluent output hides its origins and its uncertainty, so the checker and the accountable owner can't tell what they're signing off on. Every AI-influenced artifact carries what was AI-generated, its sources, and its stated confidence, so the person who owns the outcome can evaluate it.
Evidence
No outside source. This move came out of our own review of the research, so treat it as reasoned judgment rather than tested practice.
Also fits: Handoffs drop, between people and between people and AI
Home problem: Confident claims go unchecked