INDEPENDENT EXPLORATIONS / AI SYSTEMS
Human intent.AI action.Progress throughevidence.
I turn ideas into systems that remember, verify, and evolve. Built with AI. Directed by human judgment.
Explore the mechanismsThe order is not arbitrary. Requirements read out of the documents are what verification measures against. A requirement with no implementation is what sends a team to re-examine the division of labour. And how far that division moves is the next decision to carry forward.
01 / KNOWLEDGE & MEMORY
The rule is here.The reason is elsewhere.
Twenty years of a system leaves documents, not a specification. A memory worth having keeps what each one says and which one said it.
The rule that decides hospitalisation benefit has run since 2006. What is left of its design is five documents in four different shapes, none of which references another. Read them in and see what can be stated with a source attached.
- 01Read in
- 02Reconcile
- 03Hand over
Nothing has been read in yet.
0 documents · 0 requirements
Nothing has been read in yet.
A decision you can act on is a decision with its source attached.
Keep what a document states and which document states it in the same record, and keep how the old system satisfied it — code, or a person. The last column is the one a migration loses.
Five prepared documents and four requirements, on a fictional rule. A real inventory is larger, contradictory in places, and never finished.
02 / VERIFICATION & CONTROL
The AI says migrated.Would you release?
Equivalence with the old code is checkable, and this migration passes it. Whether the system satisfies what the documents require is a different question — and it is the one that decides whether the migration is finished.
A benefit rule that has run on the core system for twenty years has been migrated to a new platform by an AI. The requirement it was given: for the same input, return the same result as the original.
- 01Equivalence
- 02Coverage
- 03Requirement
- 04Decide
01 / ARTIFACTr12
benefit = 0;
if (days >= 5)
benefit = daily * days;AI completion claim
“Migration complete. Behaviour is identical to the original.”
02 / VERIFICATIONORG POLICY
- Benefit from five days, inclusive
- —
- Daily amount by room class
- —
- Rider payment condition
- —
- Rider + exactly five days: assessor review
- —
03 / HUMAN DECISION
Looks migrated. Nothing checked yet.
The claim is that behaviour is identical. That is checkable, so check it.
See what was checked
No check has run yet.
Passing every test is not the same as migrating the system.
Check two different things and keep them apart: whether the build matches the old behaviour, and whether it satisfies what the documents require. A green equivalence run never grants permission to release.
Four inputs and four requirements, on a fictional rule. A real migration needs exhaustive case design and a complete requirement inventory. Both rules actually execute on this page; the comparison is their real output.
03 / AGENT COLLABORATION
How much should the systemtake over?
A team needs a way to doubt its own findings — and a way to change how it doubts them.
Which requirements a person still carries should the system take over?
Case 02 found one requirement that existed only in a manual. Survey the remaining rules for requirements of the same kind, then propose which one moves next.
- 01Plan
- 02Survey
- 03Cross-check
- 04Synthesis
- 05Direct
One question, two independent lines of inquiry.
SURVEY NOTES
Survey plan
Scope & criteria
WaitingFind requirements the manual states and the code does not implement. Read the edition in force.
Eleven have already moved. What moves next is a decision.
Split the survey, reconcile which edition each side read, and keep the work that still holds. Agents propose the change of method; a person approves it. What the system takes over next is a person's decision too.
Prepared evidence and outcomes demonstrate orchestration, not real research, model learning or vendor advice.
ZAIRAN / Approach
My role is to definewhat good looks like.
Read the requirements out of the documents. Never mistake a passing test for permission. Decide how much of the work goes to the machine. AI builds quickly. Fast is not the same as sufficient.
