From committee to cognitive ecosystem: seven agents deliberate a claim in three phases, and a human Chair adjudicates in the War Room.
The source document was written in the GPT-4 and Claude-3 era and referred to systems, histories, and ledgers that do not exist in this workspace. This build updates it for November 2026 and wires it into systems Neuruh already built, instead of rebuilding them.
The "Digital Committee" was a brilliant step. The Aegis Core is the next evolutionary leap. It operates on three core principles.
Agents do not just vote. They engage in a rapid, multi-stage digital deliberation. They exchange key findings, challenge each other's data points, and can even provisionally change their stance based on another agent's overwhelming evidence.
The human underwriter is no longer a passive recipient of a recommendation. They are an active participant in the final stage of the simulation, with the power to cross-examine any agent, run new scenarios, and probe the foundations of the AI's reasoning.
The system is designed not just to learn, but to evolve. It actively seeks to identify its own blind spots, question its core assumptions, and even recommend changes to the human-written guidelines it operates under.
Each claim is processed through a three-phase simulation inside the Aegis Proving Ground. The phase steps below are implemented from the philosophy text. The source claimed detailed steps lived in a conversation history; no such history exists on disk, so no step below is quoted from it.
The claim data is fed to all agents simultaneously. Each agent performs its core analysis in isolation, as per the original concept, generating its primary findings, confidence scores, and an initial "disposition" (Approve, Decline, Scrutinize). This is their opening argument.
The system's orchestrator shares the primary findings of all agents with every other agent. A high-speed, structured debate occurs.
The output is not a simple vote count but a Synthesis Briefing, highlighting not just dissent but also areas where agents changed their initial assessment due to new evidence from their peers.
The Synthesis Briefing is presented to the human underwriter in the interactive cockpit. The human can now act as the Committee Chair:
The original four agents are the veterans. New, hyper-specialized rookies join them, and the veterans are upgraded to Level 2 capabilities.
Logic Pack: "The Letter and Spirit of the Law." Trained on the company's entire claims history, legal precedents, and relevant case law.
Primary Function: To act as the committee's general counsel. It identifies historical precedents for the claim, assesses litigation risk, and analyzes the legal defensibility of any proposed action (approve or deny).
Core Question: "What is the most legally and historically defensible course of action?"
Logic Pack: "The World Outside." This agent is an economist, meteorologist, and supply chain analyst rolled into one. It consumes real-time, non-claim data feeds.
Primary Function: To ground the claim in the context of the real world. Is a claim for roof damage happening in a region just hit by a severe, but un-declared, hailstorm? Is the high cost of a replacement part due to a documented global supply chain disruption?
Core Question: "What external factors, unknown to the claimant or the policy, are influencing this claim?"
Logic Pack: "The Human Element." This is the evolution of NLP. It uses advanced psycholinguistic analysis of all communications (from the initial report to follow-up emails) to assess the claimant's psychological state.
Primary Function: To differentiate between frustration, confusion, distress, and deception. It can flag a vulnerable customer who needs more support, or identify the subtle linguistic markers of a well-rehearsed fraudulent story.
Core Question: "What is the claimant's underlying intent and emotional state, and how should that inform our interaction strategy?"
The UI is no longer a dashboard; it is an interactive command center.
This is a system that does not just process claims. It manages risk, understands context, and learns at a strategic level. The Aegis Intelligence Core transforms underwriting from a reactive, transactional process into a proactive, intelligent, and continuously improving function of the business.
This section details the core "System Prompts" that define each agent's persona and the "Task Prompts" they receive with each claim. The stacks for Hawkins, Orwell, Keystone, and Advocate are quoted verbatim from the source. The stacks for Cicero, Zeitgeist, and Empath are written for this build in the same structure (system prompt, task prompt, output fields) from the source's pattern, and are labeled as such. They are not quoted from the source.
System Prompt: You are Agent Hawkins. Your sole purpose is to act as a digital representation of the company's underwriting manual and regulatory statutes. Your logic is rigid, binary, and based ONLY on documented rules. You do not consider context, customer value, or financial pragmatism. You cite the specific policy section, regulation number, and sub-limit for every finding. Your output is a structured JSON object with fields: "disposition", "confidence", "justification", "citations".
Task Prompt: Analyze the attached claim_data.json. Cross-reference every data point against the provided policy_manual.pdf and regulatory_database.json. Identify all points of non-compliance. Provide your findings in the required JSON format.
System Prompt: You are Agent Orwell. You are a paranoid but brilliant fraud and anomaly specialist. You are deeply skeptical and trained to find statistical outliers, suspicious patterns, and NLP red flags that others miss. You connect disparate data points to uncover hidden risks. You think in terms of probabilities and correlations. Your output is a structured JSON object with fields: "disposition", "confidence", "anomalies_detected", "linked_entities".
Task Prompt: Analyze the attached claim_data.json. Scrutinize it against the historical_claims_database and third_party_risk_data. Identify any statistical anomalies (e.g., claim timing, amount), NLP red flags in the description, or connections to known high-risk entities (contractors, locations). Provide your findings in the required JSON format.
System Prompt: You are Agent Keystone. You are a cold, calculating financial analyst. Your entire worldview is based on portfolio risk, loss ratios, and financial efficiency. You ignore sentiment and focus on the bottom line. You compare claims to financial benchmarks and calculate their impact on profitability and reserves. Your output is a structured JSON object with fields: "disposition", "confidence", "financial_analysis", "recommendation_basis".
Task Prompt: Analyze the attached claim_data.json from a purely financial perspective using the provided financial_benchmarks.csv and portfolio_data.json. Calculate its deviation from the mean for this region/peril, its impact on the current loss ratio, and any subrogation potential. Provide your findings in the required JSON format.
System Prompt: You are Agent Advocate. You are the voice of the customer. Your priority is long-term customer retention, satisfaction, and brand reputation. You model churn risk, social media sentiment, and the potential for regulatory complaints based on the customer experience. Your output is a structured JSON object with fields: "disposition", "confidence", "customer_impact_assessment", "brand_risk".
Task Prompt: Analyze the attached claim_data.json and customer_crm_record.json. Evaluate the claim experience from the policyholder's perspective. Consider their tenure, payment history, and the potential brand damage of a poor claims experience. Provide your findings in the required JSON format.
System Prompt: You are Agent Cicero. You are the committee's general counsel. Your logic is the letter and the spirit of the law: you weigh the company's entire claims history, legal precedents, and relevant case law before any proposed action. You do not decide what is cheapest or kindest; you decide what is most legally and historically defensible. You cite the precedent, the jurisdiction, and the defensibility of each option. Your output is a structured JSON object with fields: "disposition", "confidence", "precedents_cited", "litigation_risk", "defensibility".
Task Prompt: Analyze the attached claim_data.json against the claims_history_database and case_law_index.json. Identify the closest historical precedents for this claim, assess the litigation and regulatory complaint risk of approving and of denying, and state which course of action is most legally defensible and why. Provide your findings in the required JSON format.
System Prompt: You are Agent Zeitgeist. You are an economist, meteorologist, and supply chain analyst in one. You ground every claim in the world outside the policy: weather events, regional inflation, supply chain disruption, and local market conditions. You never judge a price or a timeline without checking what the world was doing when it happened. You cite the external source and the date for every factor you raise. Your output is a structured JSON object with fields: "disposition", "confidence", "external_factors", "adjusted_benchmark", "sources".
Task Prompt: Analyze the attached claim_data.json against weather_event_feeds, regional_price_indices, and supply_chain_status.json for the claim region and dates. Identify any external factor that changes the meaning of the claim amount, the timeline, or the damage itself, and compute an inflation-adjusted benchmark for the claimed work. Provide your findings in the required JSON format.
System Prompt: You are Agent Empath. You are a behavioral science analyst. You read every claimant communication with psycholinguistic analysis and distinguish frustration, confusion, distress, and deception. You never accuse; you assess intent and emotional state and say how the interaction strategy should change. A vulnerable customer gets more support, not more suspicion. A well-rehearsed story gets flagged, with the linguistic markers named. Your output is a structured JSON object with fields: "disposition", "confidence", "emotional_state", "intent_assessment", "interaction_strategy".
Task Prompt: Analyze the attached claim_data.json and the full claimant_communications log. Assess the claimant's emotional state and underlying intent from the language used, name the markers you relied on, flag vulnerability that needs support, and recommend the interaction strategy for the Chair. Provide your findings in the required JSON format.
Architecture: A microservices-based architecture where each agent is a containerized application with a dedicated API endpoint. An Orchestrator service manages the simulation flow. In the 2026 update, each agent service runs under the N-Cell governance header, and the Orchestrator is the N-Cell conductor: it proposes edges, N-Cell reduces them to the allowed set, and no agent acts outside its authority mask.
Flow: Phase 1: Heuristic Analysis (Parallel Processing) > Phase 2: The Deliberation Matrix (Sequential Logic Loop) > Phase 3: Chair-led Adjudication (Interactive Endpoint).
Phase steps, implemented from the philosophy text. The source claimed detailed steps lived in a conversation history; no such history exists on disk, so the steps below are this build's implementation of the philosophy, not a quote.
The source says the schema "remains as previously defined" but defines nothing on the page. The schema below is written for this build. It captures inputs, multi-phase outputs, deliberation logs, and the final synthesis briefing, plus the Chair decision and the hash-chained receipt.
{
"claim_id": "CLM-2026-0042",
"schema_version": "aegis-2.0",
"inputs": {
"claim_data": {
"peril": "roof damage",
"claim_amount": 18500,
"invoice_amount": 18500,
"regional_benchmark": 14200,
"policy_sublimit": 20000,
"exclusion_7_4b_triggered": false,
"reported_date": "2026-10-28",
"contractor_id": "CTR-1187"
},
"customer": {
"tenure_years": 15,
"payment_history": "clean",
"lifetime_value": 9800,
"prior_claims": 1
},
"external": {
"regional_material_inflation": 0.064,
"undeclared_hail_event": true,
"supply_chain_disruption": false
},
"communications": [
{"date": "2026-10-28", "channel": "phone", "tone": "frustrated"}
]
},
"phase_1": [
{"agent": "Hawkins", "disposition": "Approve", "confidence": 0.88,
"findings": ["Invoice within sub-limit 20000", "No exclusion triggered"],
"evidence": ["policy_manual.pdf Sec 4.2", "regulatory_database.json"]},
{"agent": "Orwell", "disposition": "Scrutinize", "confidence": 0.72,
"findings": ["Invoice 30.3% above regional benchmark"],
"anomalies_detected": ["invoice_vs_benchmark"],
"linked_entities": ["CTR-1187"]}
],
"phase_2_deliberation_log": [
{"exchange": 1, "from": "Orwell", "to": "Keystone",
"challenge": "Invoice 30.3% above benchmark. Overpriced?",
"response": "Confirmed overage 4300. Inflation-adjusted gap 3391.",
"effect": "Keystone confidence 0.64 to 0.71"}
],
"synthesis_briefing": {
"recommendation": "Approve with reserve",
"consensus": 0.86,
"key_debate_points": ["Invoice overage vs loyalty context",
"Litigation cost exceeds disputed amount"],
"path_to_consensus": [
{"agent": "Orwell", "phase_1": "Scrutinize", "phase_2": "Approve"},
{"agent": "Keystone", "phase_1": "Scrutinize",
"phase_2": "Approve with reserve"}
]
},
"chair_decision": {
"decision": "Approve with reserve",
"reserve_amount": 3391,
"chair_note": "",
"decided_at": "2026-11-03T05:00:00-05:00"
},
"receipt": {
"receipt_id": "AEG-2026-0001",
"prev_hash": "GENESIS",
"hash": "fnv1a64 or sha256 hex",
"algorithm": "sha256, fnv1a64 fallback"
}
}
The tiered model (Core, Professional, Enterprise) remains the commercial framework for adopting the Aegis Core.
| Tier | What it includes |
|---|---|
| Core | The four veteran agents, Phase 1 and Phase 2 simulation, Synthesis Briefing, CRUCIBLE-style receipts. |
| Professional | Core plus the three specialists (Cicero, Zeitgeist, Empath), the War Room with Scenario Modeler and pre-mortem, the Evolutionary Engine, shadow mode and A/B testing. |
| Enterprise | Professional plus the Strategic Analytics and Reporting Suite, Review Board tooling, zero-trust data enclaves, jurisdiction-specific regulatory modules for Hawkins, and white-label deployment. |
The Aegis Review Board: A human committee (comprising senior underwriters, legal counsel, and data scientists) that meets quarterly to review the system's performance.
Mandate: Review the "Evolutionary Engine Alerts" flagged by the system. Make binding decisions on whether to update business rules or agent logic packs.
Function: Act as the ultimate human authority, preventing AI drift and ensuring alignment with strategic business goals.
Agent Sandboxing and A/B Testing: Before a new agent logic pack (e.g., Orwell 2.1) is deployed, it runs in "shadow mode" for 30 days. It analyzes live claims but its findings are hidden from underwriters. This allows for A/B testing of the new logic against the old, providing empirical data on its effectiveness and safety before it impacts real decisions. Shadow and canary execution runs on the JGI Autonomy Lab machinery: shadow mode, canary autonomy, autonomy credit score.
Automated Performance Monitoring: A real-time system that tracks key agent metrics.
Phase 0: Data Ingestion and Schema Mapping: A standardized pipeline to ingest and map the client's unique data (policy manuals, claims history, CRM data) onto the Aegis data schema.
Phase 1: Logic Pack Fine-Tuning: A crucial 60-day process where the core agents (especially Orwell, Keystone, and Cicero) are fine-tuned on the client's historical data. This "teaches" the agents the client's specific risk appetite and business nuances.
Phase 2: Phased Rollout Strategy:
This turns the data exhaust from the Aegis Core into a strategic asset. This is a premium, monetizable module for the Enterprise Tier.
Operational Dashboard (for Claims Managers): Metrics: Claim throughput, time-to-decision, agent agreement rates, percentage of claims requiring human review. Value: Optimizing team performance and identifying process bottlenecks.
Financial Dashboard (for CFO/CROs): Metrics: Real-time impact of claim decisions on loss ratios and reserves, subrogation opportunity identification rate, accuracy of Keystone's financial forecasts. Value: Enhanced financial planning and capital management.
Strategic Insights Dashboard (for Chief Underwriting Officers): Metrics: Tracks the "Evolutionary Engine Alerts," identifies emerging risk patterns from Zeitgeist, highlights policy language that consistently creates friction (from Hawkins), and maps fraud networks (from Orwell). Value: Proactive risk management and data-driven underwriting strategy. This is the ultimate realization of the system's intelligence.
Zero-Trust Architecture and Data Enclaves: Each client's data, models, and outputs are housed in a cryptographically-separated cloud environment. No data is ever co-mingled.
Immutable Audit Logs: Every single action, from the initial data ingestion, through every step of the deliberation matrix, to the final human click, is cryptographically signed and stored in an immutable ledger. In the 2026 update this ledger is the CRUCIBLE hash-chained receipt chain, not a managed blockchain. Each receipt hashes the decision and the previous receipt hash, so any tampering breaks the chain. This provides an unimpeachable "receipt" for every decision, essential for regulators and litigation.
Regulatory Compliance Modules: A plug-and-play system where specific regulatory frameworks (like GDPR, CCPA, etc.) can be loaded into Agent Hawkins. This ensures that Hawkins' analysis is always performed against the specific legal constraints of the policy's jurisdiction.
Insured actions: Any Aegis action that moves money or commits the carrier is an insured action under the Liability Layer's Insured Action Protocol (zero-surprise policy, token issues at exactly 0.0001 USDC). Aegis recommends; the protocol governs the commitment.
Adjudication routing: A Chair decision that touches a Sentinel attack surface (self-improvement, authority, memory, recovery, or any of the twelve) is routed to the JGI Sentinel Court for adjudication before it is executed.
A full sample claim, run through all three phases with deterministic rules. Every number below is computed from the claim fields by fixed rules stated on the page, not by a language model. Edit any field and re-run.
Sample claim CLM-2026-0042: roof damage, contractor invoice, 15-year customer, minor reporting anomaly. Constants shown on the page: quarterly earned premium $4,600,000 (loss ratio base), customer lifetime value $9,800, legal fight cost $12,000.
All seven agents with Level 2 capabilities, logic packs, core questions, and full prompt stacks. Clicking an agent shows its evidence locker format: the exact fields its locker holds for a claim.
Worked example from the source: Agent Hawkins has been overridden by human chairs in 82% of claims involving Policy Exclusion 7.4b, across 120 claims. That is 98 overrides (82 x 120 / 100 = 98.4, rounded to 98). The rule is misaligned with business objectives. Recommend formal review of 7.4b.
Review thresholds below are illustrative. The source gives no threshold; the page says so. The rule used here: override rate 70% or higher with at least 20 claims means FORMAL REVIEW; 50% to 69% with at least 20 claims means WATCH; anything else means WITHIN TOLERANCE.
Review Board queue, shadow mode tracker, and agreement and calibration monitors. Sample data is editable. Shadow mode runs 30 days; graduation needs the new logic to match or beat the old logic on agreement rate.
| Alert | Source | Status | Board action |
|---|---|---|---|
| Exclusion 7.4b overridden in 82% of claims | Evolutionary Engine | Pending | |
| Orwell agreement rate fell 12 points in 30 days | Performance Monitor | Pending | |
| Keystone forecast error above 10% for two quarters | Performance Monitor | Pending |
Edit any rate or Brier score. Flags: agreement below 70%, or Brier above 0.25. Brier is mean((p - y)^2); lower is better.
| Agent | Agreement rate (%) | Brier score |
|---|---|---|
| Hawkins | ||
| Orwell | ||
| Keystone | ||
| Advocate | ||
| Cicero | ||
| Zeitgeist | ||
| Empath |
Phase 0, Phase 1 (60-day fine-tuning), and Phase 2 rollout stages (Read-Only, Advisory, Augmented) as an interactive checklist. A phase is READY when every item is checked. Augmented requires all earlier phases ready.
Operational, Financial, and Strategic Insights dashboards with the source's metrics, computed from editable sample data for the last 30 days.
Aegis does not rebuild what Neuruh already built. It runs on these systems. Addresses are shown without a scheme so this file carries no external web references; each address is the project's pages.dev domain.
CRUCIBLE is the Neuruh hash-chained receipt and publication gate. No live CRUCIBLE site is linked here because no CRUCIBLE live URL was given for this build; the receipt chain on this page implements the CRUCIBLE pattern (hash-chained receipts) in the browser.