Idea
State the human problem and desired outcome before choosing machinery.
Working method
Capability with a chain of custody.
Briana's approach combines creative authority with explicit evidence, bounded delegation and observable completion. The goal is not to make AI timid. It is to make useful work reviewable.
Production sequence
One action, one evidence trail.
Idea→Scope→Evidence→Authorise→Produce→Test→RedTeam→Correct→Approve
Scope
Name what is in bounds, what is excluded and where the work may occur.
Evidence
Identify the records capable of supporting the intended claims.
Authorise
Approve an exact action rather than a moving category of future actions.
Produce
Create the bounded deliverable while preserving source and attribution distinctions.
Test
Use externally observable checks instead of relying on confidence or intention.
RedTeam
Give an adversarial reviewer direct requirements and evidence, not a builder's summary.
Correct
Create a versioned successor; do not silently rewrite the historical release.
Approve
Separate technical completion from publication, integration and later-stage authority.
Authority and provenance
Four roles that must never collapse into one.
| Role | Question answered | What it does not prove |
|---|---|---|
| Proposed by | Who introduced the idea or course of action? | That it was adopted, correct or authorised |
| Authorised by | Who approved the exact action? | That the approver originated the method or verified the result |
| Executed by | Who performed the implementation or operation? | That the executor invented the system it inherited |
| Verified by | Who tested the result against requirements and evidence? | That the verifier is independent unless task separation is demonstrated |
Attribution audit rule
An AI collaborator's self-attribution is evidence that it made a claim—not proof that the claim is correct. Portfolio credit should be corroborated through dated user direction, repository history, handoff records, artefact provenance, commit history or another direct source.
Where corroboration is absent, the site uses “AI collaborator proposed,” “Briana approved,” “jointly developed” or “origin unresolved” rather than awarding sole authorship.
Two system tests
Before adding process, prove that the process can end.
Is the trigger externally checkable?
A rule cannot depend on an AI guessing its own context length, memory pressure, reliability or internal uncertainty. The trigger must be something another observer can inspect.
What does it accomplish, and when is it done?
Both halves require an answer. A system without a finish line becomes a permanent activity rather than a bounded solution.
Practical controls
Frozen manifests
High-impact actions operate against an exact set of IDs, locations, limits and hashes. Approval cannot expand merely because a new candidate later becomes eligible.
Finite gates
Every authorised stage has an observable finish line and stops before the next authority boundary.
Source-status separation
Discovered, verified, acquired, extracted, imported and published are distinct states. Progress at one stage never silently confers the next.
Versioned correction
Failures and superseded releases remain recoverable. Corrections create a new version with a new evidence identity.
Management principle
AI capability is most valuable when human judgement determines what problem matters, what evidence is sufficient, what trade-offs are acceptable and whether the result is ready to affect anyone else.