Computational Trust: Reframing Entity Authority as Annotation Efficiency in AI-Mediated Information Retrieval
The industry prescription “build trust with AI” is mechanistically backwards. AI systems do not trust brands the way humans trust each other. They process content efficiently or inefficiently. What practitioners call trust is a computational cost differential - nothing more, nothing less.
This paper introduces Computational Trust as a formal property: the inverse of marginal annotation cost. When new information about an entity fits the existing Entity Context Model cheaply - no contradiction detected, no disambiguation required, no escalation to a larger model class - the system processes it with high confidence and low latency. When it does not fit, the system hedges, re-evaluates, or rejects. The Doubt Tax - the hesitation AI introduces when recommending a brand it cannot confidently ground - is the business cost of high marginal annotation cost.
The Entity Context Model is the cached knowledge structure the system assembles against every new piece of content it encounters about an entity. Rich, consistent, well-corroborated entity graphs produce low marginal annotation cost. Sparse, contradictory, or self-sourced entity data produces high marginal annotation cost. The Corroboration Threshold - the point at which an entity crosses from low-confidence to high-confidence processing - is not a continuous accumulation. It is a step function: below the threshold, every new claim is expensive to process. Above it, new claims are absorbed cheaply and with high confidence.
The paper also introduces Corroboration Decay - the structural erosion of entity graph confidence through third-party content deletion. Trust built on external corroboration is trust that external parties can destroy through deletion, not just contradiction. Decay-resistant evidence architecture - distributing corroboration across independent platforms, permanent records, and controlled sources - is the necessary complement to entity optimisation.
The Annotation Cost Differential explains competitive outcomes that quality-first models cannot: why new entrants with superior content underperform established entities (no entity graph to evaluate against), why inconsistent messaging destroys authority faster than silence (contradiction raises marginal cost directly), and why established entities recover from errors faster (rich entity graphs have strong priors that absorb isolated contradictions cheaply).
The Confidence Threshold is binary. Entities either cross it or they do not. Below it, the system processes with hedging language (“reportedly,” “claims to be”). Above it, the system states facts. The language the AI uses when discussing a brand is not editorial judgement. It is an annotation cost readout.
Grounded in practitioner evidence from 73 million brand profiles tracked since 2015 and the 2024 Google documentation disclosure confirming binary entity-URL attributes.
Part of The Kalicube Process™ formal foundation.
Published: 22 February 2026 · Zenodo Open Access · Version v1 Affiliation: Kalicube SAS
Read the full paper on Zenodo · Cite via DOI: 10.5281/zenodo.18735062
Key concepts introduced or formalised in this paper: Computational Trust, Entity Context Model, Annotation Cost Differential, Confidence Threshold, Corroboration Decay, Corroboration Threshold, Doubt Tax.
