Provenance
Where did the evidence come from? Who created it, when, and can its lineage be reconstructed?
Architecture
From evidence to action without losing the reasoning.
This is the proposed operational path. Expand each topic for the governing questions, limits, and current status.
01
Context
Purpose, scope, population, jurisdiction, and operating boundary.
02
Evidence
Known, claimed, inferred, contradicted, and missing.
03
Evaluation
Criteria, rigor, and versioned method.
04
Recommendation
What the evidence supports and what remains uncertain.
05
Decision
Who has authority to authorize an outcome?
06
Execution
What may happen, by whom, and within what limits?
07
Escalation
When must the system refuse, pause, or seek review?
08
Record
Can a qualified reviewer reconstruct what happened?
One path. Separate permissions.
A system may be allowed to evaluate but not decide, recommend but not execute, or prepare an action but require approval before commitment. Authority can be delegated. Accountability cannot be delegated to the machine.
Atomic EvaluationsSmall judgments with explicit boundaries
I use the Atomic Evaluation as the proposed smallest independently governable judgment. A material evaluation should define its purpose and scope; criterion, rule, or threshold; evidence requirements, provenance, and sufficiency; conflict treatment; reasoning method and rigor; machine authority and human controls; permitted outputs; validation cases; and audit requirements.
The aim is to make consequential reasoning inspectable, challengeable, versionable, and governable without pretending the reasoning itself is simple.
Evidence integrityProvenance · coverage · independence
Coverage
What should have been available? Which repositories were searched, retrieved, inaccessible, excluded, or still missing?
Independence
Is this actually corroboration? Several artifacts may derive from one upstream claim or source.
A perfectly traceable evidence package can still represent a curated reality. Traceability shows what entered the system; coverage asks what failed to enter.
Authority modelScoped · current · effective
Evaluation authority does not imply recommendation, decision, or execution authority.
- Commit-time: check current authority immediately before consequential commitment. Cached, expired, revoked, or unverifiable permission is insufficient.
- Dynamic: authority should contract when supporting evidence or controls materially weaken.
- Effective: assess the consequences reachable through tools, agents, services, credentials, delegation, and trust relationships.
Capability is not authority. Prior behavior, convenience, access, technical ability, and a user request do not by themselves establish permission.
Human judgmentApproval is not automatically review
An approval record proves that an approval event occurred. It does not prove meaningful human judgment. Review depends on relevant evidence, visible contradictions and uncertainty, reviewer competence, adequate time, and practical agency to challenge, change, reject, or escalate.
Meaningful review also does not guarantee a good decision. I do not infer a person's motive. I focus on preserving what was available, recommended, decided, and done—and on making material divergence difficult to conceal.
Boundary assuranceDeclared · configured · effective
A system instruction is not an enforceable boundary. The architecture distinguishes what the system believes it may do (declared), what infrastructure appears to permit (configured), and what consequences the complete system can actually produce (effective). The model can help identify boundary failure; it does not certify its own boundary.