V-Score Card

RELX PLC (RELX)
V-SCORE:   3.63  (stress-tested floor: 3.38)
VERDICT:   EMBEDDED
κ:         0.63  (stress-tested: 0.38)
BASKET:    KEEP

GATES:     G₁ = 1 (E=4 > 1)    G₂ = 1 (A=3 > 1 ∧ C+E+U=13 ≥ 12)

DIMENSIONS:
  C = 5  (w=0.25)  →  1.250   Compound cognition
  E = 4  (w=0.22)  →  0.880   Irreducible infrastructure
  U = 4  (w=0.18)  →  0.720   Ecosystem breadth
  A = 3  (w=0.12)  →  0.360   Distribution & discoverability
  M = 4  (w=0.15)  →  0.600   Ecosystem gravity
  F = 3  (w=-0.06) → -0.180   Friction penalty
                       ─────
  V = 3.630 × G₁(1) × G₂(1) = 3.63

STRESS TEST (adversarial review on C and E):
  C: 5→4  →  V = 3.38, κ = 0.38  (double-counting w/ E, frontier model erosion)
  E: 4→3  →  V = 3.41, κ = 0.41  (Legal bypassed by Claude Cowork)
  Both    →  V = 3.16, κ = 0.16  (floor — still EMBEDDED)

REGIME CONTEXT (15-week, T ending 2026-03-28):
  ρ_intra = 0.722  (INDISCRIMINATE SELLOFF — IR unreliable)
  IR = -0.50       (noise — 68% of variance is sector factor)
  %Idio = 31.9%    (regime artifact, not portfolio construction failure)
  δ = V − V_market = 3.63 − 2.54 = +1.09  (full tier of mispricing)
  δ_stressed       = 3.38 − 2.54 = +0.84

DURABLE REV:  ≈85%  (Risk 36% + STM 28% + Exhibitions 12% + Legal content 9%)
EXPOSED REV:  ≈15%  (Legal search UI 10% + Print 5%)

Fast Screen (Bustamante)

1. Proprietary data? YES — 130B+ transactions analyzed annually in Risk via contributory databases covering "a majority of US auto and property insurance policies, claims and shopping transactions" (20-F p.9286-9287). 207B+ legal documents. 105M+ scientific records. Cannot be synthesized locally.
2. Regulatory lock-in? PARTIAL — FTC consent decrees govern data handling (20-F p.2226-2228). Insurance data regulated at state level. No hard NRSRO-type mandate requiring use of RELX specifically.
3. Transaction-embedded? YES — Risk segment 90% machine-to-machine, embedded in insurance underwriting/claims/identity verification workflows (CEO, Q4 2025 earnings call).

b(s) = 2/3 — strong prior for E ≥ 4.

Dimension Analysis

C = 5 — Compound Cognition (w=0.25)

The cognition IS the data. Editorial annotations across 207B legal documents, citation networks spanning 105M scientific records, risk scoring models trained on 130B+ annual transactions, peer review quality signals from 1.9M expert reviewers — an agent can query this but cannot re-derive it.

Five layers of crystallized cognition: (1) raw data accumulation over decades, (2) editorial curation by 37,000 editors and 1.9M reviewers producing 795,000 published articles annually (20-F p.9564-9567), (3) linking algorithms generating "high precision and recall" across billions of records (20-F p.8714-8716), (4) knowledge graphs connecting entities across legal, scientific, and risk domains — Shepard's Knowledge Graph grounds Lexis+ AI answers (20-F p.10349-10351), (5) context engineering combining all layers for specific use cases (20-F p.8737-8738).

Shepard's Citations has tracked case law relevance since 1873 — 153 years of continuously maintained citation mapping. Elsevier produces 18% of global research output but captures 29% of citations (20-F p.9675-9676) — quality share nearly 2x volume share, reflecting accumulated editorial prestige.

Products "account for less than 1% of our customers' total cost base but can have a significant and positive impact on the economics of the remaining 99%" (20-F p.8473-8475).

Stress test (C → 4): The double-counting concern is real — contributory insurance data appears in both E and C justifications. Stripped to independent cognition: STM peer review network is genuinely C=5 (irreplaceable human judgment at scale), but Risk scoring models are trainable given data access (C=3-4), and Legal's Shepard's completeness guarantee faces erosion as frontier NLP models can infer citation treatment from raw case text. Segment-weighted C ≈ 3.9, rounds to 4. Honest range: C = 4-5.

E = 4 — Irreducible Infrastructure (w=0.22)

Petabyte-scale specialized infrastructure with no hard regulatory mandate. The contributory insurance database creates two-sided lock-in: insurers both contribute data and consume insights. Leaving means losing access to the pooled intelligence of the entire industry (20-F p.9284-9287).

Risk is 90% machine-to-machine — displacement requires re-engineering insurance underwriting systems, not just switching a subscription (CEO, Q4 2025). Multi-year contracts with 3-year typical length provide temporal lock-in (20-F p.14997). Deferred revenue GBP 2,390M at Dec 31, 2025 — roughly 25% of annual revenue locked in as advance receipts (20-F p.1386-1387).

FTC consent decrees from 2006 and 2008 require "comprehensive data security programmes, submissions of regulatory reports and on-going monitoring by independent third parties" (20-F p.2226-2228) — regulatory overhead as barrier to new entrants.

Why not E=5: No hard regulatory mandate like NRSRO status (S&P Global) or clearing house designation (ICE). Infrastructure is massive but not legally mandated.

Stress test (E → 3): Claude Cowork was built WITHOUT routing through RELX — revealed preference that Legal infrastructure (19% of revenue) is bypassable. Verisk competes in insurance data via ISO (since 1971). TransUnion, Equifax compete in identity/fraud. Segment-weighted E ≈ 3.5. But Risk (36% of revenue, largest segment) has genuine E=4.5 characteristics: contributory network + M2M embedding + regulatory compliance burden. Entity-level E = 4 is defensible. Honest range: E = 3.5-4.

U = 4 — Ecosystem Breadth (w=0.18)

38+ distinct products across four segments serving researchers, doctors, lawyers, insurers, banks, government agencies, HR professionals, and event organizers in 180+ countries (20-F p.583-584).

Risk: fraud detection, identity verification, insurance underwriting/claims, financial crime compliance, government analytics, commodity intelligence (ICIS). STM: academic publishing (3,000+ journals including Cell Press and The Lancet), Scopus AI, ClinicalKey, Sherpath AI. Legal: Lexis+ AI, Protege agentic assistant, Lex Machina litigation analytics, CounselLink+ legal management, Nexis news intelligence across 40,000 sources in 50 languages. Exhibitions: 550+ events.

Cross-segment data flows are material: Healthcare was "formerly reported within STM" but now sits in Risk (20-F p.8952), combining identity resolution with medical content. LexisNexis Regulatory Compliance serves Risk-typical customers from the Legal platform (20-F p.10091-10092).

Why not U=5: Breadth is across industries and customer types rather than deep within one enterprise (SAP's model). No single customer depends on RELX across 20+ departmental workflows.

A = 3 — Distribution & Discoverability (w=0.12)

Functional API infrastructure, not default agent routing. Risk delivers via real-time APIs handling billions of transactions (20-F p.8678-8681). Lexis+ AI has "hundreds of thousands of users" delivering "over five million prompts" in 2025 (20-F p.10029-10030, CEO Q4 2025). Over half of US new and renewing customers adopting Lexis+ AI (20-F p.10027-10028).

Harvey AI strategic alliance (June 2025): LexisNexis integrates Lexis+ AI capabilities within the Harvey platform (20-F p.10110-10117). Management "open to interoperability" and willing to "license proprietary data via API" to AI tools (CEO, Q4 2025 earnings call).

Why not A=4: Claude Cowork — the most prominent AI legal tool — was built without RELX integration. Harvey is one partnership, not universal agent routing. RELX is positioning as content infrastructure for AI (correct strategy) but hasn't achieved default routing status. Building toward A=4, not there yet.

M = 4 — Ecosystem Gravity (w=0.15)

Insurance contributory databases exhibit genuine network effects: each insurer's contribution increases the pooled intelligence available to all participants. ThreatMetrix is "one of the largest cross-industry data networks in the world" analyzing billions of transactions annually (20-F p.9397-9398).

Academic publishing has the strongest gravity: the prestige flywheel where researchers MUST publish in top journals for tenure and grants creates a self-reinforcing network. Elsevier's 29% citation share vs 18% volume share (20-F p.9675-9676) quantifies the quality premium.

Legal research is a consolidated duopoly: "two big players" (CEO, Q4 2025). No third player has achieved critical mass in decades.

Why not M=5: Duopoly in Legal (Thomson Reuters is comparable). Verisk competes in insurance data. No single segment has monopoly status.

F = 3 — Ecosystem Friction (w=-0.06, penalty)

Moderate enterprise learning curve. Not consultant-dependent like SAP — products sold "through dedicated sales forces direct to customers" (20-F p.8470). Risk M2M APIs are production-grade (ThreatMetrix: 140M users via Crypto.com with 950% volume growth, 20-F p.9428). Lexis+ AI case study showed "structured and seamless" onboarding (20-F p.10374-10376). But enterprise purchasing cycles, institutional procurement (STM library consortia), and multi-year contract negotiation add friction that pure self-serve tools avoid.

Thermodynamic Summary

Intelligence cannot flow around RELX because the cognition IS the low-energy state. Using 130B+ annually refreshed insurance transactions, 153 years of citation mapping, and the peer review judgments of 1.9M domain experts is cheaper than re-deriving any of it. The contributory database flywheel in Risk creates a gravitational well where each participant's contribution increases the value for all others — a thermodynamic trap with no lower-energy alternative. The 90% M2M embedding in Risk means the infrastructure IS the transaction, not adjacent to it. An agent can query RELX's crystallized cognition but cannot replicate the decades of editorial judgment, cross-domain linking, and network participation that produced it.

Regime Context

ρ_intra = 0.722 over 15 weeks (T ending 2026-03-28). The data/analytics sector is in an indiscriminate selloff. Eight peer names (RELX, TRI, SPGI, MCO, VRSK, MSCI, FDS, ICE) show mean pairwise correlation of 0.72 — well above the 0.4 differentiation threshold. The sector is trading as a single factor.

When ρ_intra → 1, the idiosyncratic residual ε_i → 0 for all names. IR_i → 0 not because alpha is absent, but because the measurement window contains no idiosyncratic signal. IR measures the regime, not the stock.

RELX's IR = -0.50 is a regime artifact: %idiosyncratic variance is 31.9% (68% of variance explained by market + sector factors). β_sector = 1.11 confirms RELX is moving 1:1 with the peer group selloff. This is not evidence of negative alpha — it is evidence that the sector factor dominates all variance over this window.

IR does NOT gate the verdict. V-Score is orthogonal to sector returns: V(s) ⊥ r_sector(t). Structural properties (filings, infrastructure, cognition layers) do not change because the sector sold off. The market applies a uniform AI disruption discount during selloffs. The edge is the delta between structural reality and market-implied survival.

Partial differentiation within the selloff (dispersion = 9.2%) maps to structural quality: ICE (-5%) and MSCI (-8%) with regulatory mandates outperform; FDS (-31%) and TRI (-32%) with weaker infrastructure underperform. The market has a partial V-Score model — it's applying too uniform a discount, not a random one.

Conviction Weight

V-Score:        3.63  (stress-tested floor: 3.38)
κ = (V − 3.0)⁺: 0.63  (stress-tested: 0.38)
w_i ∝ κ_i:      normalized across basket

V_market:       2.54  (implied by 14.7x forward P/E vs 27.5x historical)
δ = V − V_mkt:  +1.09  (stressed: +0.84)

Maximum δ occurs at maximum indiscriminate selloff — which is the current regime (ρ_intra = 0.72). This is the optimal entry condition for V-Score-based positioning: structural signal is highest-conviction precisely when price signal is lowest-quality.

Basket Verdict: KEEP

RELX scores EMBEDDED across the full stress-test range (V = 3.16 to 3.63). Even the adversarial floor (C=4, E=3, V=3.16) holds EMBEDDED status. Gate 2 passes via A=3 > 1 in all scenarios.

The market prices RELX at AT_RISK (V_market ≈ 2.54). The structural analysis says EMBEDDED. That delta — 0.84 to 1.09 points — is the trade. RELX stays in the basket at target weight, conviction-weighted by κ.

What changes the score:

• A: 3→4 (Harvey alliance scales, API licensing to Cowork) → V = 3.75, κ = 0.75
• E: 4→3 (data portability regulation, open insurance data mandates) → V = 3.41, κ = 0.41
• C: 5→3 (frontier models replicate Shepard's + peer review quality at scale) → V = 3.13, κ = 0.13

Catalyst to watch: ρ_intra declining. When the sector starts differentiating — ICE and SPGI decoupling upward while FDS continues down — the regime breaks and V-Score deltas start collapsing into price.

Evidence

Bull evidence (items 1-7, 9-13): independent signals across filings, transcripts, market data, cross-ticker patterns. Combined LR ≈ 8-12 (conservatively — items 1-3 partially correlated as same-filing signals).

Bear evidence (items 8, 14): Claude Cowork bypass and competitive presence in insurance data. Combined LR ≈ 0.4.

Net LR ≈ 1.8. Structural durability is real but the market isn't wrong about everything — Legal faces genuine AI competition, and Risk data isn't a monopoly. The mispricing is in degree (AT_RISK pricing for an EMBEDDED company), not in direction (AI disruption IS affecting the sector).