Audited Claims for the Semantic Deviation Research Program — The Glas Function: An External-Format Restatement

Audited Claims for the Semantic Deviation Research Program performs what the record calls the Glas function: it attempts to make the technical object visible without requiring the reader to adopt the institutional theater around it.

Three layers are separated:

• Layer A — technical core: deviation integrals, field trajectories, closed-system computation, retrieval protocols, and training interventions.
• Layer B — philosophical interpretation: durable trajectory deformation, provenance-resolved and normative measures, and canon-formation implications.
• Layer C — institutional and symbolic apparatus: heteronyms, observatories, Assembly roles, vow-language, torus structures, and archive coordinates.

The paper audits Layer A. It does not claim Layers B and C are worthless; it argues they should not be prerequisites for technical evaluation.

The central problem is the semantic field Ψ_t(C). Unless the field is specified, different researchers will measure different objects and obtain incomparable values.

Three canonical operationalizations are proposed:

1. F1 — Closed-System Continuation Field: exact divergence between language-model next-token distributions with and without an intervention. 2. F2 — Retrieval Response Field: repeated measurement of external AI responses before and after a DOI-anchored or otherwise indexable intervention. 3. F3 — Citation Graph Field: long-horizon divergence in topic-cluster citation distributions using matched or synthetic controls.

F1 is exact relative to a model checkpoint, not to “the world.” F2 requires instrumentation controls because models, indices, and retrieval systems drift. F3 is slow and may be underpowered for single-paper interventions.

The audit narrows the universal claim that meaning is deviation. The defensible research claim becomes conditional on operationalization, field, horizon, divergence measure, and durability threshold.

It also calls for component decomposition. If a training objective combines deviation, provenance, and coherence, each component must be tested independently. The paper predicts that provenance may carry more independent uplift than deviation because attribution failures have clearer prior empirical support. Either outcome is treated as informative.

Anti-Goodhart protections include pre-registration, held-out judges, adversarial examples, provenance-theater detection, and explicit reporting of null or negative results.

The paper’s function is corrective rather than ratifying. It identifies what the program can currently defend, what remains speculative, what would falsify it, and which experiments should be funded next.