Decide what good looks like.
Outcome — what should change?
Owner — who will close the work?
Boundary — where may AI act?
Evidence — how will we know the result is right?
HOW HELMARK WORKS · STD 3.0
Four decisions keep AI delivery under human control.
Helmark does not add a second development process. It adds a control model inside the one you already have.
THE DELIVERY LOOP
FRAME → RUN → CHECK → OWN
The loop becomes especially important when AI or agents perform a substantive part of the work.
Allow autonomy inside the Boundary.
A human, an AI assistant, one agent or multiple agents can execute. Helmark does not require approving every AI step.
Verify two different things.
Right Thing? Did we solve the right problem?
Works Right? Is the implementation actually correct?
Make acceptance a human act.
The Human Owner should understand what changed, why the result is trusted and what the most important failure risk is.
WHEN FRAME IS REQUIRED
Stronger AI involvement or higher risk means a clearer control envelope.
Full FRAME is required.
Full FRAME is required.
Full FRAME is required.
A shorter note about AI use and verification can be enough.
PROPORTIONAL VERIFICATION
Do not treat all AI use as equally risky.
NONE / ASSIST / CORE describes how the result was created. LOW / HIGH / CRITICAL describes how serious failure could be.
ASSIST / LOWShort record of AI use + how the Owner checked the result
ASSIST / HIGHFRAME + test/scenario + real human review
ASSIST / CRITICALHIGH controls + independent human verification criterion
CORE / LOWFRAME + test/scenario + real human review + OWN
CORE / HIGHCORE / LOW + deliberate risk analysis
CORE / CRITICALCORE / HIGH + independent human verification criterion
FLOW CONTROL
CHECK Load is the queue to watch.
If AI generates work faster than people can verify, understand and accept it, the system is building unfinished control work.
CHECK FIRST
Finish verification before adding more AI output.
Reduce parallel work
Fewer active Work Units or agents can improve actual throughput.
Improve Evidence
Better Evidence shortens responsible acceptance.
LONG-TERM CONTROL
Keep the organisation able to understand its own system.
Software can work while understanding decays.
Helmark treats this loss of understanding as a delivery risk.
Know who can genuinely explain important components.
An Anchor is a visible source of real system understanding.
Recognise when independent judgement has been lost.
Rebuild knowledge, reduce risk, limit changes or explicitly accept the risk.
Watch Energy, Focus, Pressure and Clarity.
Use trends as a delivery-system sensor, never as an individual productivity KPI.
FULL SPECIFICATION