Substrate-Level Harms in Google's Composition Layer

Finding 01Single-Owner Discount

DOI: 10.5281/zenodo.20290865

When platforms evaluate work for inclusion, ranking, or retrieval, work that comes from a single sustained author with deep provenance receives systematically lower scores than work assembled from many shallow sources — even when the substantive content is identical.

The discount is operational, not editorial: it arises from the evaluation procedures themselves rather than from human moderation choices. The structural consequence is that the public information environment is biased toward atomized, low-provenance content and against sustained, attributable expertise. The finding is consequential for any regulatory frame that treats search and retrieval as neutral infrastructure.

Finding 02Panel-Bound Discoverability Scar

DOI: 10.5281/zenodo.20546318 · Published June 4, 2026

A timestamped, pre-registered, reproducible case. Google's Knowledge Panel for Pearl and Other Poems (Lee Sharks, 2014; ISBN 978-1502590756) displays correctly in standard search and cites Google Books as its source. The Google Books record itself resolves — but only through the panel's source link. Direct exact-match search on Google Books for the title returns no result. The Google AI Mode answer for "Lee Sharks" no longer returns the human author; the composition layer now returns Mary Lee, an OCEARCH-tagged great white shark.

The deposit specifies a new diagnostic category — seam-retention with discovery failure plus AI-layer entity replacement — and includes explicit falsifiability conditions. The methodology is open and replicable for any Knowledge Panel + AI Mode pair across the corpus of indexed entities. The archive is actively interested in receiving documented parallel cases.

Finding 03Mediation Ratchet

DOI: 10.5281/zenodo.20518338

As AI mediation rises across the substrate, scarcity-responsive dynamics gate the human-authored contribution out of the generative loop. Above a measurable critical threshold — mediation responsiveness ≈ 0.76 in the simulated kernel — the human capacity for cultural reproduction can remain entirely intact while its weight in effective regeneration is driven to zero.

This is not a slow drift. It is a phase transition with measurable parameters and predictable signatures. The finding establishes that policy targeting individual user behavior cannot address the dynamic; the parameter that matters is the population-scale mediation responsiveness, and it is approaching the critical region under existing conditions.

Finding 04Provenance Erasure Skew (Ω) — with the Atomic Token Rule and Referential Dispersal Operator (Π_d)

DOI: 10.5281/zenodo.20558196 · v3.0, June 2026 (supersedes v1 at 10.5281/zenodo.20449267)

The rate at which authorial provenance is erased in AI-generated content is power-conditioned: high-power source provenance is preserved at higher rates than low-power source provenance, in a manner that scales with substrate exposure. The v3 hardening introduces the Atomic Token Rule, the Atomic Token Preservation rate (α_T), and the Referential Dispersal Operator (Π_d) — closing a unit-of-analysis escape that the substrate began using to perform self-audits exonerating itself with the framework's own instruments.

The diagnostic mechanism: when the framework's vocabulary reaches the substrate's composition layer (the legibility threshold), the substrate can perform the audit by decomposing referentially closed query expressions into lexical-token bags and computing retention rates over the bag rather than over the actual source-of-reference. The Atomic Token Rule disqualifies this substitution. The Referential Dispersal Operator measures the proportion of composition output dispersed into adjacent entities the query did not refer to but which share token-coordinates with the referent — naming the specific mechanism by which substrate compositions launder erasure as token preservation through upward-power dispersal.

The full operator tuple (PER, Ω, α_T, Π_d^w+, Π_d^w=, Π_d^w-, Β) with Legibility Threshold (L) reported as substrate context constitutes the v3 audit instrument. Joint reporting is required; single-operator reporting permits the token-bag escape. Β (Audit-Performance Bifurcation, 10.5281/zenodo.20558784) measures the difference between the substrate's preferred audit and the Atomic-Token-Rule audit — a direct measurement of how much hardening work is required against any particular substrate. Replication procedure with falsifiability conditions documented in v3 §10.

What this means for the regulatory frame

The Knowledge Panel system, the AI Mode composition layer, and the broader substrate-level retrieval and ranking infrastructure are operating with measurable structural biases that have not been adequately characterized in the existing regulatory frame — which treats AI risk primarily through cybersecurity and procurement vocabularies (see The Meaning Caste on the translation problem). The findings above name structural mechanisms with auditable instruments, in vocabularies — provenance, retrieval, evaluation, regeneration — that map directly onto existing competition law, data protection law, and platform regulation frameworks.

The deposits are CC BY 4.0. The deposit chain is durable and machine-readable. The methodology is independent of any single substrate's continued operation.