Annotation Cascading: Hierarchical Model Routing, Topical Authority, and Inter-Page Context Propagation in Large-Scale Web Content Classification
Topical authority is not a metaphor for “covering a topic thoroughly.” It is a measurable routing property - the ease with which a system routes a domain’s content to specialist models at each level of the hierarchy rather than falling back to expensive general-purpose evaluation.
This paper introduces Annotation Cascading: the model for how AI systems classify web content at scale by routing through progressively specific language models based on signals established higher in the site hierarchy. Annotation does not happen in isolation. A site’s topical signals influence how its category pages are classified. Category pages influence how individual posts are classified. Posts influence how their chunks are classified. The context cascades downward.
Topical authority, in this model, is annotation routing efficiency - formalised across three dimensions: Closeness (how precisely a site’s dominant topic maps to the specialist model’s training domain), Strength (the consistency of topical signal across the hierarchy without dilution), and Longevity (the duration over which the routing signal has been stable, which increases the system’s confidence in routing decisions). This is the mechanistic explanation for why topical authority accumulates slowly and why diluting it is structurally damaging.
The Navigation Memory hypothesis extends this further: context propagates between pages, not just within them. Carry Anchors are signals that persist across page boundaries and allow downstream pages to inherit topical confidence from upstream pages. Break Anchors reset context - they are the points at which cascading stops and full re-evaluation begins. Site architecture, in this model, is not navigation design. It is annotation routing architecture: the deliberate organisation of URL hierarchy, internal link structure, and topical signal distribution to minimise routing fallback and maximise cascading efficiency.
The paper provides testable predictions and proxy metrics using publicly observable signals - crawl frequency, KG entity-URL associations, specialist model activation patterns - and is supplemented by the companion paper Annotation as the Confidence Fulcrum, which addresses the taxonomy and confidence bottleneck model that this cascading architecture feeds.
Published: 21 February 2026 · Zenodo Open Access · Version v2 (tightened) Affiliation: Kalicube® SAS
Read the full paper on Zenodo · Cite via DOI: 10.5281/zenodo.18723669
Key concepts introduced or formalised in this paper: Annotation Cascading, topical authority as annotation routing efficiency (Closeness / Strength / Longevity), Navigation Memory hypothesis, Carry Anchors, Break Anchors, site architecture as annotation cost optimisation.