⚖️ Courts — Source Path + Sequence

📚 SOURCE PATH — HARVEY · CASE-001 · CORPORATE POWER ANALYSIS

⚠️ HONEST ANSWER FOR ALEC

Harvey's analysis draws exclusively from Anthropic Claude Sonnet 4.6's training corpus — a large language model trained on internet-indexed text up to approximately early 2025. It does NOT query Westlaw, LexisNexis, Casetext, or any live legal database in real time. Every case and regulation Harvey cites is drawn from public sources that were in the training data. The analysis is a reasoning scaffold — not a law firm opinion. Atmos's lawyers validate against live databases.

🌐

LAYER 1 — EXTERNAL WORLD (Public Sources)

Real decisions, regulations, and reports published in the world. All publicly accessible. All indexed.

CASE LAW

Meta Platforms v DPC / EDPB (Jan 2023)

€390M fine · EU · "contractual necessity" cannot substitute valid consent for ad targeting · Directly cited by Harvey re: consent vs. routing intent distinction · Source: EDPB press release, EU Official Journal

CASE LAW

Schrems II — CJEU C-311/18 (2020)

EU-US Privacy Shield invalidated · Principle: formal mechanisms cannot substitute genuine control · Underpins Harvey's "platform compliance consultation" strategy · Source: CJEU official judgment database

STATUTE

GDPR Articles 13–22 (2018, enforced)

Transparency, access, erasure, restriction, portability, right to object to automated decisions · Referenced as the regulatory framework Harvey maps corporate exposure to · Source: EUR-Lex

REGULATORY

Platform Business Model Analysis (public)

Google ad targeting architecture (public AdWords/Ads documentation) · Meta cross-platform data correlation (EDPB cross-platform investigation 2022) · Amazon supplier profiling (FTC findings) · Telco traffic shaping (net neutrality record) · All publicly documented

REPORT

DLA Piper Global Data Protection Report 2024

162 countries with data protection legislation · 71% have enforcement authority · Cited by CLO (not Harvey directly) for global enforceability claim · Publicly available PDF

🤖

LAYER 2 — LLM TRAINING CORPUS (Anthropic)

Public sources above were crawled, cleaned, and used to train Claude Sonnet 4.6. Training cutoff: approx. early 2025. The model does not have a "database" — it has compressed parametric knowledge of these sources.

LLM

Anthropic Claude Sonnet 4.6 — Base Model

Provider: Anthropic PBC · Training data: Common Crawl, curated legal/regulatory corpora, academic papers, news · Weights encode probabilistic knowledge of cited cases · No live query capability at inference time · Model card: anthropic.com/claude

⚙️

LAYER 3 — EOSE GOAT SYSTEM (Role Injection)

The model is given Harvey's identity, role definition, and the CASE-001 matter as a structured prompt. No external retrieval. The model generates a response from its parametric knowledge through Harvey's lens.

PROMPT

Harvey Role Injection

Role: "Harvey Specter archetype — negotiation, corporate power, closing deals" · Matter: CASE-001 Intent Routing full text · Instruction: "Analyse through your lens, cite real cases, give a Proper brief and an ELI5" · Temperature: low (structured legal output)

PROMPT

GOAT Crew Context (14 lawyers)

Each GOAT receives: (1) their identity, (2) the matter, (3) awareness that 13 other GOATs are analysing simultaneously · Harvey's response is NOT influenced by what Ruth/Oliver/Alan wrote — each GOAT is independent, run separately

⚖️

LAYER 4 — COURTS OUTPUT (yLAW)

Harvey's response is stored in the Courts system, scored by CLO, and promoted through Prelim → Advisory → Final tiers as peer review increases confidence.

OUTPUT

Harvey Response — CASE-001 Proper Brief

Score: 94/100 (CLO peer review) · Tier: Final Court · Key novel claim: approach platforms for intent routing compliance consultation BEFORE court filing to create competitive pressure · This specific strategy is Harvey's original synthesis — not found verbatim in any source

OUTPUT

ELI5 Version — Same Brief, Different Lens

Same sources, reframed for non-lawyers. The "magic toy / wish-list" metaphor is the model's original synthesis of the same legal content. Both versions derive from the same external sources via the same parametric knowledge.

🌏

LAYER 5 — ATMOS VALIDATION

Atmos lawyers validate Harvey's cited cases against Westlaw/LexisNexis. They confirm the cases exist and the citations are accurate. They apply professional judgment on whether the novel argument holds. The floor analysis is the scaffold — Atmos builds the house.

🔄 SEQUENCE DIAGRAM — CASE-001 FULL FLOW · MATTER → GOAT → COURTS → ATMOS

#	FROM	TO	MESSAGE	SOURCE / NOTE
01	Kay/Amani	CLO	File matter: CASE-001 Intent Routing	Manual filing via Courts UI
02	CLO	ONBA	Request institutional standing	ONBA = filing authority; Sayyo co-sign pending
03	ONBA	CLO	Standing granted (CASE-001-ONBA)	CLO Cloak timestamp in git
04	CLO	14 GOATs	Distribute matter + role contexts	Parallel dispatch — each GOAT independent
05	External World	LLM Training	GDPR · Schrems II · Meta €390M · CJEU cases (indexed pre-2025)	Public sources → Anthropic training corpus
06	Harvey (LLM)	Courts	Proper brief + ELI5 · Score 94/100	Parametric recall of Layer 1 sources through Harvey lens
07	Ruth (LLM)	Courts	Constitutional / equality frame · 97/100	Charter s.15 · Law v Canada · Gosselin · Withler
08	Alan (LLM)	Courts	Logic structure · 3-premise architecture · 98/100	Jones v Tsige · GDPR Article 22 · tort of intrusion
09	Oliver (LLM)	Courts	International law frame · 96/100	UN Res 73/22 · Schrems II · DPDP India s.9 · OCAP
10	CLO	Courts	Peer review → promote to Final (score ≥ 100)	Prelim → Advisory → Final tier system
11	Courts	Atmos / ONBA	Final brief package — reasoning scaffold delivered	NOT a law firm opinion. Validation by Atmos required.
12	Atmos Lawyers	Westlaw/LexisNexis	Validate cited cases · confirm citations accurate	Live legal database — EOSE does NOT have this
13	Atmos	Client/Court	Professional legal opinion — Atmos-signed	Atmos carries all professional liability. EOSE is the floor.