Freeze a baseline
Architectural responsibilities and test expectations are versioned before consequential evaluation.
RESEARCH
From derivation to permission.
The central question is whether consequential automated systems can preserve executable distinctions between formal derivation, authority, warrant, reliance and action permission.
PRIMARY QUESTION
Can an automated system remain useful without silently turning prediction or derivation into authority?
The research treats the architecture as a falsifiable hypothesis. A useful negative result is preferable to a persuasive but unsupported claim.
FormalDerivation != AuthorityToApplyRule
WARRANTED(P) != MAY_RELY_ON(P,U)
MAY_RELY_ON(P,U) != ACTION_TECHNICALLY_PERMITTED
Projection != Authority
LLM memory != authoritative system semantics
METHOD
Make assumptions executable, then try to break them.
Use governed sources
Authority, provenance and modelling conveniences are represented separately rather than silently conflated.
Run hostile tests
Negative, ambiguous and structurally novel cases are retained rather than removed from the evidence set.
Reconstruct the decision
The intended end state is a decision that can be rebuilt from governed artefacts without conversational memory.
STATUS DISCIPLINE
Claims carry a status.
DemonstratedImplemented and observable in a bounded artefact or test.
SupportedCurrent evidence supports the proposition within an explicit scope.
HypothesisA falsifiable candidate explanation or architecture requiring further evidence.
Open questionNo warranted conclusion is currently claimed.
FUNCTIONAL RESPONSIBILITIES
Responsibilities first. Technology second.
The public research description deliberately focuses on what must be done rather than prescribing one implementation technology.
Generative substrateMay propose candidate interpretations or actions.
Semantic orchestrationSelects the relevant governed domain responsibilities.
Formal/domain reasoningDetermines what follows under explicit rules and semantics.
Warrant & reliance assuranceEvaluates support, authority, applicability, scope and reliance.
Provenance & reconstructionPreserves the decision basis and versioned dependencies.
Action boundarySeparates a warranted conclusion from permission to change the world.
BOUNDARIES
What this research does not currently claim.
Universal domain independence
Elimination of AI hallucination or agent drift
Official DfT D-TRO conformance or certification
Legal advice or legal correctness
TM Forum affiliation or endorsement
Production safety for unrestricted consequential action
That generated output is automatically warranted
That a projection is independently authoritative
WORKING DEMONSTRATOR
Inspect the current public evidence.
Studio v1.1 exposes model structure, validation outcomes, generated artefacts, provenance and unresolved semantics in the browser.
Open Studio v1.1