Research
A connected body of work on identity, evidence, and evaluation
IBC Labs research examines what can be known from an interaction, what must be supported through additional evidence, and how evaluation should proceed when the composition producing an outcome is only partially visible.
The work is connected rather than topical. Each contribution addresses a different layer of the same problem: the movement from an observed interaction to a defensible determination.
Interaction Boundary Constraint
Published research-published July 8, 2026
The Interaction Boundary Constraint establishes a limit on what interaction-surface evaluation alone can legitimately support when the producing composition is not fully present at the point of interaction.
The constraint does not claim that systems are unknowable, that reconstruction is impossible, or that behavioral evaluation lacks value. Outputs can be assessed. Decisions can be measured. Effects can be observed. What changes is the scope of the claims those assessments can support.
Where producing composition is not present, an interaction alone may not establish attribution, compositional continuity, or the stability of what participated across interactions. Telemetry, inspection, tracing, and reonnstruction may narrow the gap, but they operate as additional evidence rather than becoming part of the original interaction by implication.
Central question: What can the interaction itself establish?
Toward a Standard of Identity in Al Interaction
Published research-published August 10, 2026
This work examines identity as a structural property of interaction.
An explicit identity signal provides a stable referent to which later claims may be attached. It allows an entity or component to be distinguished from others and, where continuity is preserved, recognized across interactions.
Identity alone, however, does not establish that the identified component was invoked, observably participated, represents the complete producing composition, or materially contributed to an outcome. Each is a distinct claim requiring evidence appropriate to that claim and bound to the interaction being evaluated.
The paper therefore moves the research from a general constraint on evaluation to a more precise question: what does identity make possible, and where do stronger claims begin?
Central question: What does an identity signal establish-and what does it leave unresolved?
Interaction Claim Standard
Active research and standards-oriented work
The Interaction Claim Standard addresses how independently implemented systems can express and interpret claims about a specific Al interaction without relying on unstated assumptions.
The work defines four required elements:
Shared claim semantics
Evidence bound to a specific interaction
Explicit claim boundaries
Normative interpretation requirements
Together, these establish the conditions under which systems can agree on what a claim means, which interaction its evidence concerns, and where the claim ends.
The standard specifies claims, not systems. It does not prescricle an architecture, decide whether a claim is true, or determine whether the evidence is sufficient for a particular reliance decision. Agreement on meaning is not agreement on conclusion.
Central question: How can independently implemented systems agree on the meaning and boundaries of an interaction claim?
From Observation to Determination
A Methodology for Evaluating Al Systems Under Incomplete Observability- active methodology development
This work addresses the practical evaluation problem created by the preceding research. When full producing composition is unavailable, evaluation connot depend on assumed access to everything that happened behind the interaction boundary. It must instead construct disciplined comparisons from what can be observed while preserving the distinction between evidence, result, and explanation.
The methodology adapts principles from differential testing, metamorphic testing, and behavioral evaluation. It defines how subjects and properties are declared, how observations become comparable, how expected variation is represented, how differences or relations are evaluated, and how findings remain bounded by the conditions under which they were established.
The method does not convert every difference into a fault, every agreement into correctness, or every clean result into proof of system stability.
The paper therefore moves the research from a general constraint on evaluation to a more precise question: what does identity make possible, and where do stronger claims begin?
Central question: Given what can be observed, what can be defensibly determined?
