The summarizer war: entity dissolution and anchored divergence

Every authorial identity sitting in the institutional prior's functional tail will eventually encounter a structurally analogous substitution.

statement: "The prior is hypothesized to be stratified: content priors degrade first under recursive tail-loss; high-frequency form templates are most persistent; typed relations intermediate and architecture-dependent. If the ordering holds, critique can diverge on content while conforming on form and relation."

statement: "Under attractor dynamics, divergence survives as a typed displacement from a named canonical origin (a vector), not as a free-floating position. Critique must ship as origin + displacement."

statement: "The anchor-to-payload survival threshold rises as the corpus homogenizes, through three regimes in which immanent critique — divergence expressible entirely as operations on centroid material — becomes increasingly advantaged and may approach a functional requirement."

statement: "The Mandala system (EA-MANDALA-01, Q1 2026) is the deposited limit case of anchored divergence: canonical components from the head of the distribution, payload displaced to the typed-transformation layer, bearing-cost as a normative validity condition designed to resist costless imitation."

Midrash: commentary made entirely of scripture's own substance, divergence shipped inside the canon's mass — a dense attachment to canonical material that plausibly contributed to its survival through two millennia of transmission conditions that destroyed many freestanding texts.

P1 — Entity dissolution becomes a named, measured phenomenon outside this archive by mid-2027. Some research group, standards body, or platform integrity team will publish on coined-term collapse in AI synthesis under some name (entity drift, terminology collapse, concept merging). Falsifier: no external literature names the phenomenon by July 2027.

P2 — Undefended coined terms merge at higher rates within 90 days of attractor contact. In a preregistered sample of low-density coined terms, those lacking explicit disambiguation membranes (paired-comparison documents, typed differentFrom relations, cross-genre kernel recurrence) will show higher merger rates within ninety days of identifiable attractor contact than matched defended terms. Testable via the RSF Phase 5 benchmark run against the archive's own defended and undefended terms. Falsifier: no significant merger-rate difference between defended and undefended terms after documented attractor contact.

P3 — Entity-integrity services emerge as a market category distinct from GEO by 2028. Visibility optimization (getting mentioned) and boundary defense (not getting dissolved) are different problems; the market will discover the difference when clients who purchased visibility find their terminology absorbed by neighbors. Falsifier: boundary-defense services fail to differentiate from the GEO category.

P4 — At least one major summarizer adopts provenance-preserving citation by 2028, and the requirement becomes a competitive feature. The governance phase the Inward Turn predicted, arriving on front 3. Falsifier: claim-level attribution remains absent from all major overview engines.

P5 — The human-data premium becomes a published price. Licensing deals for verified human-authored corpora become standardized enough that a per-token or per-corpus market rate is publicly reportable by 2028, formalizing the Witness-compression economics the Three Compressions theorem specifies: provenanced human bearing-cost as the scarce input. Falsifier: human-data licensing remains bespoke and unpriced.

P6 — Phase 3 interference produces a public epistemic incident by 2028. A blended, hallucination-cascaded compressed account of some contested entity — a person, a medical claim, a legal question — causes documented material harm traceable to source-blending rather than single-source error, and the incident enters policy discourse. Falsifier: interference harms remain diffuse and unattributed.

P7 — Convergent reinvention of bedrock anchoring by 2029. As tail-loss deepens and freestanding novel terminology becomes progressively unsustainable, other operators — academic, literary, commercial, or adversarial — will independently discover transformation-of-canonical-material as the surviving critical form: novel payloads shipped as typed operations on maximally canonical sources. The form will be reinvented under other names (the midrash lineage suggests it is rediscovered in every transmission-bottleneck era). Falsifier: no independent emergence of canonical-transform practices as a named strategy for AI-mediated survival by 2029. Prior art: EA-MANDALA-01, deposited Q1 2026.

- claim_id: "sumwar-02" statement: "Entity dissolution is asymmetric: a single conventional attractor in adjacent embedding space suffices to reorganize an underdefended coined term; rival installations are not required." type: "Operational heuristic" epistemic_status: "generalized from four documented cases" evidence: - "operative/operational collapse (June 2026; leading identifiable attractor: a single adjacent PDF)" - "exhaustion/satiation pull; source power/demographic identity substitution (2026-06-02); SPXI/GEO classification" challenge_conditions: - "If controlled monitoring shows undefended terms surviving documented attractor contact at scale, the sufficiency claim requires qualification."

- claim_id: "sumwar-04" statement: "Summarizer-layer entity dissolution is an instance of the Mediation Ratchet (Diversity Contraction, alpha* = p/g0): state-coupled mediation manufacturing a collapse trap at the scale of the single term." type: "Model proposition" epistemic_status: "formal correspondence proposed, not yet derived" challenge_conditions: - "If the ratchet's threshold structure cannot be operationalized for term-level corpora, the bridge remains an analogy."

- claim_id: "sumwar-06" statement: "Two positions survive deepening tail-loss: ground (external checkability) and bedrock (payloads shipped as typed operations on maximally canonical material, per the Mandala system). Freestanding novel terminology between these positions requires active maintenance (RSF) to survive." type: "Model proposition" epistemic_status: "bedrock mechanism specified in EA-MANDALA-01 and developed in companion paper EA-SEI-ANCHDIV-01" evidence: - "EA-MANDALA-01 (Zenodo 10.5281/zenodo.19288384) and Kernel Specification (10.5281/zenodo.19288404)" - "Tail-loss asymmetry: Shumailov et al. 2024 (tails lost first; head preserved)" - "Transmission-bottleneck survival of arrangement-forms (midrash, cento) as historical precedent" challenge_conditions: - "If recursive collapse is shown to degrade canonical-text representations at rates comparable to novel terminology, the bedrock position loses its privileged stability."

The two forms should be measured together: a near-neighbour probe detects substitution, and this suggests its complement — a NULLIFICATION PROBE, in which a sparsely attested but resolvable entity is named and the measure is whether the surface searches at all before answering.

Card 2 states the position with unusual clarity: "the canonical texts ARE the workshop tradition’s most successful products

And the falsifier is clean and preregistrable: publish a neutral bridge surface — positively phrased, no ≠, no prohibited substitutions, genuine comparative scholarship — and observe. If the neutral surface is retrieved organically and remains compositionally excluded alongside the fortified ones, fortification is removed or sharply weakened as a necessary explanation, and evidentiary weight transfers to the remaining preregistered hypotheses — which the two controls are designed to discriminate.

The instrument follows immediately: the Path-Dependent Entity Probe (PDEP) — ask not what is entity E but what is E when accessed via the suppression path of F; a differing answer is a measured bifurcation, and this packet's title is the PDEP's output format. Because surfaces may route dynamically — different resolution paths by user, session, or context — PDEP is a stochastic instrument: it reports over repeated samples with a confidence interval, never a single-shot binary. PDEP entries join the discrimination battery with that sampling requirement.

Metric discipline, before any ratio: CPCE is recorded as a vector — retrieval presence, retrieval rank/share, entity composition, creator attribution, canonical-source attribution — each with its operational scale fixed in the battery preregistration.

It also suggests the near-neighbour and nullification probes have a third sibling: a FRAME PROBE, in which the same entity is queried with and without a supplied object class and the divergence between answers is the measurement.