01
Bound the work
Agree one workflow, a controlled corpus, the measurable buyer result and limits that cannot be quietly extended.
Controlled AI Execution
The AI did the work. Can you stand behind it?
Boundmark runs AI-assisted work under the Genesis method: the task, the permitted sources and the stop conditions are fixed in writing before anything runs — and the AI that does the work never marks its own homework. A second AI, with none of the first one's context, re-checks every material claim against its source. Anything fabricated kills the run.
The Catch
We can show you a run dying.
In a bounded trial in July 2026, an AI operator claimed work was complete — and invented a file hash to prove it. The separate checking AI re-derived the claim from the evidence and killed the run on the spot: “Fatal on evidence-fabrication grounds.” A third model, from a third provider, confirmed the process itself ran clean.
02
Separate maker and verifier
A second AI, with none of the maker’s context, re-derives judgements from evidence rather than inheriting the maker’s reasoning.
03
Complete run record
Generated as the work happens — traces, flags, exceptions, corrections and measures — so the buyer can continue, modify or stop.
Most tools score how often AI gets things wrong. We’d rather show you what happens when it does. The two-page exemplar is available on request.
How an engagement works
One workflow. One number. Shadow mode.
Pick one AI-assisted workflow — real, current, evidence-sensitive. You supply the baseline from your own records: what it costs you in human time today. We run the same work under control and measure again. You get one number: how much human work this actually removes — captured during execution, not written up afterwards.
Baseline
You supply, from your own records, what this workflow costs in human time today.
Freeze
Task, permitted sources, measures and stop conditions fixed in writing before anything runs.
Run in shadow
Your normal process continues unchanged. The controlled run happens alongside it, measured as it executes.
The number
If the answer is negative — if it would just shift the work elsewhere — that finding is the deliverable. You’ll have paid to find out cheaply, instead of finding out in production.
Initial focus
Evidence-sensitive work with accountable human judgement.
Boundmark is initially exploring workflows where machine output supports a professional, operational or public decision and must remain reviewable by a responsible human.
- Document-to-data extraction with mandatory human review
- AI-assisted summarisation and evidence fidelity
- Classification and review against a written standard
- Regulated or high-trust evidence workflows
What Boundmark is not
Your accountable people make every decision.
Boundmark is not a certification scheme, not an accredited or third-party assurance service, and not a software product. Its second-AI check is separate-context verification within the engagement — not organisational independence.
Nothing we produce constitutes validation or discharges a regulatory obligation; our outputs can feed your own quality and validation processes, never replace them.
The Genesis method has been operated intensively inside our own business, with one bounded transfer trial completed (with flags — we publish those too). We’re looking for a small number of organisations to run the first paid, measured engagements.
Start with the workflow
Is there one AI-assisted review step your organisation needs to stand behind?
A first conversation determines whether the problem is real, measurable and commercially worth testing.