Identity before action
Every consequential action must have a known actor, a purpose, a scope and an authority boundary. Technical capability is never treated as permission.
Operating doctrine
Build systems that can act. Govern what they are allowed to do.
This is the public-safe constitutional layer behind my work across software architecture, AI control planes, automation, infrastructure and delivery. The core rule is that important systems must be identifiable, authorised, observable, recoverable, accountable and capable of proving what actually happened.
Permanent principlesAuthority Policy Human control Evidence Audit Memory Production Recovery Value
The rules that survive frameworks, models and vendors.
Default deny
Unknown, ambiguous or expired authority fails closed. Privileged, destructive, financial and production actions require stronger controls than read-only work.
Agents do not invent authority
AI may analyse, propose, draft, test and execute authorised work. It does not silently create its own permission to perform consequential actions.
Claims must be provable
"Built", "tested", "merged", "deployed", "live" and "verified" describe different states. The language used must match the evidence available.
History survives the agent
Important execution records should outlive the model, process or person that produced them and remain independently inspectable.
Memory informs; evidence proves
Long-term context is useful for reasoning and continuity, but remembered state is not silently promoted into authoritative operational truth.
Production is controlled
Changes should understand current state, target state, dependencies, recovery and verification. Emergency work is reconciled back into canonical source and evidence.
A backup must be restorable
Recovery is part of design: backup creation, independent copies, integrity, restore procedure, periodic testing and ownership.
Engineering must create outcomes
Technical quality, client value, commercial value and evidence of value belong in the same delivery system.
Definition of Done
Done means proven.
A production-bound item is not complete merely because code exists. The applicable implementation, test, review, canonical-source, deployment, operational, verification, recovery and evidence conditions must be satisfied.
- Designed — the intended solution and constraints are documented
- Built — implementation exists in a controlled source
- Tested — relevant checks have passed
- Merged — the accepted change is in its canonical branch
- Deployed — the intended environment received the change
- Live — the running service is reachable and operational
- Verified — behaviour and state were independently checked
Execution sequence
The operating loop.
Authoriseidentity, purpose, scope and policy
Executebounded tools and explicit capability
Verifytests, health, evidence and outcome
Learnfailures strengthen the next iteration
Current application
The doctrine is being turned into executable controls.
Current work applies these rules to AI execution envelopes, default-deny policy, approvals, provider routing, audit chains, metering, estate automation and evidence-led portfolio publishing.