# Core Notes claim register

Version: v0.3-draft working control layer  
Research cutoff: 2026-08-24 (Asia/Shanghai)  
Scope: 5 Parts / 12 Chapters of the English base manuscript

ID note: supplied papers display as `PAPER-01`–`PAPER-42`; platform/protocol
references display as `PLAT-01`–`PLAT-15`. Any remaining bare P-code in a
frozen control field is historical and must be resolved by context before use.

This register identifies the propositions the manuscript intends to defend. It
is not a list of accepted facts. Each empirical claim must ultimately resolve to
completed paper or official-source evidence cards; each original definition,
model, heuristic, and design principle must be labeled as such in the text.

Machine-control note: citation keys in an `EVIDENCE-ROUTED` or
`BOUNDED-IN-TEXT` evidence-route cell are required to agree with the paper-card
backlinks. Contextual citations in `METHOD-DRAFT`, `HYPOTHESIS`, and `GATE` rows
do not become supporting edges merely because they are named. The frozen
machine audit proves structural agreement and hash integrity only; all cards and
claims still require human scholarly verification and independent second review.

Status vocabulary:

- **METHOD-DRAFT** — original definition, formalization, or design principle;
  requires methods/editorial review rather than borrowed authority.
- **EVIDENCE-ROUTED** — candidate primary sources are assigned and the automated
  claim-level full-text pass is complete; human verification, exact locators,
  and independent second review remain incomplete.
- **BOUNDED-IN-TEXT** — the current manuscript uses the claim with an explicit
  design/system/version boundary; full evidence card and second review remain.
- **HYPOTHESIS** — testable mechanism or transfer proposition; not established.
- **GATE** — normative publication/release rule of this project.

## Chapter 1 — Generated Answers as an Information Interface

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C01.001 | definition | A generative search surface is modeled as a time- and state-conditioned stochastic kernel over answer, displayed attribution, and observable interaction trace. | Original formalization; IR/RAG foundations; `xiong2024when` | Every product searches the web, exposes citations, or follows one call path. | METHOD-DRAFT |
| C01.002 | conceptual synthesis | Mention, citation, absorption, fidelity, prominence, and action have different units and denominators. | `aggarwal2024geo`, `kai2026from`, `tian2026diagnosing` | Any one metric is a universal GEO score or business value. | BOUNDED-IN-TEXT |
| C01.003 | mechanism model | Visibility can fail at acquisition, retrieval, context, generation, attribution, presentation, or action; some paths can bypass displayed citation. | Original DAG/chain; Chs. 3–5 | The named stages are disclosed internals of a closed platform. | METHOD-DRAFT |
| C01.004 | empirical route | Platform, cultural, and brand visibility contrasts require versioned query and entity sampling. | `junyao2025cultural`, `kumar2026generative` | Observed mention differences prove training-corpus geography or market causality. | EVIDENCE-ROUTED |

## Chapter 2 — Claims, Evidence, Provenance, and Source Identity

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C02.001 | design principle | Source identity and claim support are separate audits. | Original editorial rule; DOI/arXiv/venue checks; W3C PROV-O | A DOI, institutional logo, or official domain validates all claims. | GATE |
| C02.002 | schema | A claim record preserves proposition, quantity, unit, population, time, locale, conditions, evidence span, provenance, conflict, and supersession. | Original CES schema; provenance foundations | Structured metadata proves truth or independence. | METHOD-DRAFT |
| C02.003 | evidence ceiling | Scientific, official, platform, vendor, practitioner, video, and local sources support different claim classes within their methods and scope. | `SOURCE_LEDGER.md`; 86-source catalog | The ceiling is a universal evidence hierarchy or source-quality score. | GATE |
| C02.004 | empirical route | Evidence preference, source-class effects, and reputation narratives must be compared by study design rather than citation clusters. | `wan2024what`, `chen2025generative`; the previously listed narrative review is context-only and is not an admitted supporting edge | Abstract language or publication type establishes persuasion mechanisms. | EVIDENCE-ROUTED |

## Chapter 3 — Web Acquisition, Representation, and Query Planning

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C03.001 | decomposition | Discoverability, fetchability, renderability, eligibility, observed indexing, and freshness are distinct acquisition states. | Original protocol-derived decomposition; RFC 9309, Sitemaps, and dated platform documentation are candidate primary routes, but exact claim-level official-source cards remain unresolved under `U-C03.001-OFFICIAL-CARDS` | A successful fetch proves indexing, retrieval, citation, or mention. | METHOD-DRAFT |
| C03.002 | mechanism model | A natural-language request may induce a sequential, state-dependent query plan with an explicit stopping rule. | Original mechanism model informed by IR and agentic-search foundations; no supplied paper card currently admits a supporting edge | The reconstructed query tree is a closed product's internal trace. | METHOD-DRAFT |
| C03.003 | representation claim | Retrieval operates on versioned representations and passages, not author intent or document-level truth. | IR/chunking foundations; controlled examples | Co-locating a qualification universally improves production visibility. | METHOD-DRAFT |
| C03.004 | transfer boundary | Confidence- and role-augmented models can test planning hypotheses in bounded systems. | `zhao2026beyond`, `chen2026role` | Their latent variables or platform abstractions describe every engine. | EVIDENCE-ROUTED |

## Chapter 4 — Retrieval, Fusion, Reranking, and Context Allocation

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C04.001 | method | Lexical and dense retrieval scores require a declared normalization, calibration, rank-fusion, or learned-fusion rule. | Classical/neural IR foundations | Raw score addition is a neutral or universal hybrid baseline. | METHOD-DRAFT |
| C04.002 | method | Query fusion preserves recovering-query and provenance paths so repetition is not mistaken for independent support. | Original trace schema; provenance foundations | Multiple URLs or query hits establish independent confirmation. | METHOD-DRAFT |
| C04.003 | optimization model | Context allocation is a constrained evidence-coverage problem with token, redundancy, origin, and diversity trade-offs. | Original formulation; RAG/context foundations | The equation is a disclosed commercial packing objective. | METHOD-DRAFT |
| C04.004 | empirical synthesis | Content-centric, end-to-end, attribution, attention, and openness studies identify different bounded retrieval/influence questions. | `chen2025cc`, `kim2026sageo`, `patel2025maxshapley`, `alipour2026when`, `triedman2025millstone` | One benchmark ranking or attention result transfers to all production surfaces. | EVIDENCE-ROUTED |

## Chapter 5 — Generation, Citation, Attribution, and Source Influence

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C05.001 | measurement | Claim segmentation and citation-object resolution precede entailment and completeness scoring. | Citation-evaluation foundations; `tian2026diagnosing` | Response-level “has citation” measures support quality. | METHOD-DRAFT |
| C05.002 | measurement | Context selection, source use, claim absorption, and displayed attribution are non-equivalent events. | `chen2025cc`, `kai2026from`; controlled ablations | A displayed link proves source use, causation, authorship, or authority. | BOUNDED-IN-TEXT |
| C05.003 | measurement | Automated support judges require stratum-aware human calibration and error sensitivity. | Evaluation/judge foundations | A high model-judge score is ground truth across systems or languages. | METHOD-DRAFT |
| C05.004 | security route | Citation and caption interventions expose bounded failure modes that require isolated adversarial testing. | `mochizuki2025exposing`, `chen2026caption` | Every engine or citation is vulnerable at the reported rate. | EVIDENCE-ROUTED |

## Chapter 6 — Metrics, Labels, Judges, and Benchmark Design

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C06.001 | measurement | A GEO metric declares population, query/system weights, unit, denominator, missing policy, and uncertainty. | Survey sampling/statistics foundations; current protocol | A vendor composite or dashboard metric is self-interpreting. | METHOD-DRAFT |
| C06.002 | measurement | Query panels are task-stratified and split by intent clusters rather than surface paraphrases. | IR evaluation; leakage controls | Equal queries per stratum represent user demand without weights. | METHOD-DRAFT |
| C06.003 | measurement | Repeated answers are nested in query, system, and time and should not be analyzed as independent rows. | Repeated-measures foundations; `schulte2026don` | A universal repeat count or interval applies to every platform. | BOUNDED-IN-TEXT |
| C06.004 | benchmark route | GEO16-style, e-commerce, and large-scale measurement systems must be compared by units, sampling, labels, and reproducibility. | `kumar2025ai`, `bagga2025e`, `kumar2026generative` | Benchmark inclusion proves construct validity or platform generality. | EVIDENCE-ROUTED |

## Chapter 7 — Experiments, Causal Inference, and Temporal Uncertainty

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C07.001 | causal model | A treatment effect is relative to versioned source states, assignment, comparison, estimand, system state, and assumptions. | Potential-outcomes/experimental-design foundations | A before/after difference is causal by default. | METHOD-DRAFT |
| C07.002 | inference | Uncertainty follows the sampling and assignment unit; repeated runs estimate within-cluster variability. | Cluster inference, bootstrap, randomization inference; bounded diagnostic route through `vishwakarma2026what` | Hundreds of response rows imply hundreds of independent units. | METHOD-DRAFT |
| C07.003 | temporal design | Difference-in-differences, switchbacks, and interrupted time series require uptake, carryover, trend, and concurrent-shock diagnostics. | Causal/time-series foundations | Adding a design formula removes platform drift. | METHOD-DRAFT |
| C07.004 | interference | Competitive and adaptive GEO can violate no-interference assumptions; displacement and context composition must be reported. | `vishwakarma2026what`, `wu2026from` | A target's gain is ecosystem welfare or a stable isolated effect. | EVIDENCE-ROUTED |

## Chapter 8 — Content and Evidence Engineering

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C08.001 | design principle | A governed fact ledger precedes expression optimization and records scope, time, evidence, ownership, approval, and correction. | Original system design; provenance/accessibility standards | A ledger or schema makes claims true or current automatically. | GATE |
| C08.002 | hypothesis | Co-locating claims, units, conditions, methods, and limits may reduce passage-level evidence loss. | Controlled reconstruction; `yu2026structural` | Page structure universally causes higher closed-platform citation. | HYPOTHESIS |
| C08.003 | taxonomy | Expression-preserving, evidence-surfacing, evidence-expanding, ecosystem, and prohibited edits require different gates. | Original editorial taxonomy; intervention literature | All content changes are equivalent “GEO tactics.” | METHOD-DRAFT |
| C08.004 | empirical synthesis | Content interventions have stage-, model-, domain-, and candidate-state-specific effects, including negative results. | `wu2025what`, `tian2026diagnosing`, `ye2026ecogeo`; `yu2026structural` is retained only as a quarantined reporting example | A tactic, structure, or learned optimizer has one common mechanism or universal effect. | EVIDENCE-ROUTED |

## Chapter 9 — Optimization Algorithms and Agentic GEO

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C09.001 | optimization model | GEO is a constrained vector optimization over permitted source states with non-compensable fidelity and risk gates. | Original formulation; optimization foundations | One scalar score legitimately offsets severe unsupported-claim risk. | METHOD-DRAFT |
| C09.002 | evaluation | Development, validation, and test panels are separated by intent cluster; reused tests become development data. | ML/IR evaluation foundations | Paraphrase-only splits prevent leakage. | GATE |
| C09.003 | empirical synthesis | Multi-query fusion, output control, rewriting, and feature-level optimization use different objectives and threat assumptions. | `zhou2026if`, `jin2026controlling`, `luttgenau2025beyond`, `liu2026think` | Their gains are directly comparable or benign by definition. | EVIDENCE-ROUTED |
| C09.004 | governance | Adaptive systems may diagnose, propose, test, and retain bounded memory; accountable roles approve, release, monitor, and roll back. | `tian2026diagnosing`, `yuan2026agenticgeo`; `wu2026from` as an exploratory/high-reporting-risk case; project release model | Benchmark gains establish autonomy, factual reliability, scalability, or permission to publish. | GATE |

## Chapter 10 — Multimodal, Platform, Brand, and Evidence Ecosystems

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C10.001 | representation | Pixels, captions, alt text, metadata, and surrounding pages are distinct representations and treatments whose effects and failure modes differ. | `du2026multimodal`, `chen2026caption`; `zhang2026generative` as a bounded industrial architecture case | A bundled change identifies one modality's effect or a multimodal benchmark implies a live-platform result. | EVIDENCE-ROUTED |
| C10.002 | measurement | Platform comparisons require surface, version, query, locale, geography, account, time, repeat, and tool/search state; a dynamic alias is not necessarily a version lock. | Original observation-cell protocol; `perplexity2026presets` first-party interface boundary | A provider/model label or copied request freezes all backend state. | GATE |
| C10.003 | measurement | Brand mention, correctness, citation, absorption, prominence, action, and risk remain separate. | Chs. 1 and 6; brand studies | Mention rate establishes brand value, sentiment, recommendation, or causality. | METHOD-DRAFT |
| C10.004 | ecosystem model | Source independence depends on provenance origins and derivation, not URL count. | W3C PROV-O; `ye2026ecogeo` | Effective-origin count is complete, causal, or a quality score. | METHOD-DRAFT |

## Chapter 11 — Manipulation, Detection, Defense, and Mechanism Design

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C11.001 | threat model | GEO creates attack surfaces across acquisition, retrieval, context, generation, attribution, multimodal inputs, and measurement. | `nestaas2024adversarial`, `hu2025dynamics`, `mochizuki2025exposing`, `du2026multimodal`, `wen2025position`, `sun2026how` | All current production systems share every vulnerability. | EVIDENCE-ROUTED |
| C11.002 | definition | Evidence-preserving edits retain support, material qualifications, provenance, and disclosure; improved visibility is not part of the definition. | Original definition | Truthful isolated sentences guarantee a non-manipulative implication. | METHOD-DRAFT |
| C11.003 | evaluation | Defense reports answer utility, attribution, honest-source utility, attack risk, and false positives by source class. | Security evaluation foundations; defense papers | Block rate alone demonstrates a safe or fair defense. | METHOD-DRAFT |
| C11.004 | empirical synthesis | SCI-Defense, SafeGEO, GPE, GEO-Flag, and mechanism-design work address distinct verification, detection, and incentive tasks, and each exposes material failures or reporting ceilings. Adversarial dynamics remain routed only through C11.001. | `yu2026sci`, `wen2026safegeo`, `wang2026gpe`, `chu2026geo`, `xu2026mechanism` | Their metrics belong on one leaderboard, an average defense score proves coverage, or a detector proves malicious intent. | EVIDENCE-ROUTED |

## Chapter 12 — Governance, Standards, Regulation, and Incident Response

| Claim ID | Type | Draft proposition | Evidence route | Forbidden generalization | Status |
|---|---|---|---|---|---|
| C12.001 | governance model | Governance maps a concrete release to identity, changed claims, evidence, risk, controls, approvals, monitoring, and rollback. | Original release-evidence model; NIST/ISO governance sources | A policy statement or checklist proves control effectiveness. | GATE |
| C12.002 | status rule | Institutional sources enter through verified issuer, identifier, edition, status, effective date, scope, and official URL. | Appendix C; NIST/ISO/SAC/CAC official identity records | A draft, project, framework, standard, law, regulatory measure, and platform policy are interchangeable. | GATE |
| C12.003 | control rule | A crosswalk records applicability, owner, implementation artifact, test, threshold, exception, and refresh date. | `ISO/IEC 42005:2025`; `GB 45438—2025`; official-source register; qualified review | A crosswalk is compliance, certification, or legal advice. | GATE |
| C12.004 | incident protocol | Incident response preserves exact system/source evidence, contains harm, propagates correction, and verifies closure. | NIST risk/incident concepts; original protocol | Correcting one public page removes all dependent copies or model effects. | METHOD-DRAFT |

## Release accounting

The v0.3-draft register contains **48 claim IDs**: four per chapter. It is complete as
an architectural inventory, not as scientific validation. Before v1.0:

1. every `EVIDENCE-ROUTED` or `BOUNDED-IN-TEXT` item must name completed evidence
   cards, exact version, evidence span, method, result, limitations, and reviewer;
2. every `METHOD-DRAFT` item must be labeled in the manuscript and pass methods
   review for definition, domain, assumptions, units, and counterexamples;
3. every `HYPOTHESIS` must remain explicitly testable and cannot be rewritten as
   a result without a registered experiment;
4. every `GATE` must have a validator, review checklist, or release artifact;
5. superseded claims remain in a changelog rather than disappearing silently.