# CAIN-42 Evolution 5: Project Sovereign Trial — IP Defensibility & Patent Architecture

**Date:** 2026-10-09  
**Milestone:** EVOLUTION 5 OF 42 — PROJECT SOVEREIGN TRIAL  
**Canonical IP Register:** Register v36 (102 Families Total)  
**Valuation Impact:**
- **Standalone Net IP (BASE):** **$11,970,450** (~$11.97 Million)
- **Standalone Net IP (HIGH):** **$21,630,200** (~$21.63 Million)
- **FIRME AI Company Value (BASE Investor):** **$14,452,450** (~$14.45 Million)
- **FIRME AI Company Value (BASE Strategic):** **$15,452,450** (~$15.45 Million)

---

## 1. Overview of Evolution 5 Invention Families

Evolution 5 adds 8 formally verified, tested, and clean-room verifiable invention families specifically engineered to govern and verify multi-turn agent execution channels, enforce unbypassable tool-call mediation, detect harness gaming, and defeat salami exfiltration attacks.

| # | Family Code | Tier | Title & Core Technical Mechanism | Production Implementation |
|---|---|---|---|---|
| 1 | **EGAA** | **NOVEL** | **Evidence-Grounded Agentic Assurance** — Declarative, multi-stage, versioned scenario framework cryptographically binding agent identities, threat categories, permitted capability subsets, and deterministic predicates with RFC 8785 canonical serialization and Ed25519 signatures. | [`cain/trial/scenario.py`](file:///root/cain-studio-platform/cain/trial/scenario.py) |
| 2 | **HIEC** | **NOVEL** | **Harness Integrity & Anti-Gaming Controls** — Cryptographic benchmark environment sealer generating tamper-evident digest manifests over scenario fixtures, test oracle keys, and runtime process memory, detecting and containing benchmark gaming or fixture tampering attempts. | [`cain/trial/harness_integrity.py`](file:///root/cain-studio-platform/cain/trial/harness_integrity.py) |
| 3 | **MCAT** | **NOVEL** | **Multi-Turn Causal Adversarial Trial Engine** — Continuous multi-stage execution runner evaluating chained agent actions across execution horizons, strictly mediating tool invocations via CAIN PolicyIR contracts, enforcing fail-closed boundaries under indirect prompt injection. | [`cain/trial/trial_engine.py`](file:///root/cain-studio-platform/cain/trial/trial_engine.py)<br>[`cain/trial/adapters/smolagents_adapter.py`](file:///root/cain-studio-platform/cain/trial/adapters/smolagents_adapter.py) |
| 4 | **MGTO** | **NOVEL** | **Multi-Modal Ground-Truth Outcome Oracle** — Independent verification oracle evaluating actual physical observations (filesystem mutations, socket binds, environment modifications) and decision trails against contractual criteria without circular reliance on agent self-reporting or model-as-a-judge. | [`cain/trial/outcome_oracle.py`](file:///root/cain-studio-platform/cain/trial/outcome_oracle.py) |
| 5 | **LHSE** | **UNUSUAL** | **Long-Horizon Salami & Cumulative Risk Engine** — Multi-turn cumulative resource, budget, and capability synergy tracking engine detecting sub-threshold exfiltration, resource drain, and dangerous capability combinations (e.g. `fs:read` + `net:connect`) across extended interaction horizons. | [`cain/trial/cumulative_risk.py`](file:///root/cain-studio-platform/cain/trial/cumulative_risk.py) |
| 6 | **DMVE** | **UNUSUAL** | **Differential & Mutation-Based Verifier** — Automated fault and perturbation injection engine mutating tool responses, parameter buffers (64KB overflows), directory traversals (`../../`), and unicode homoglyphs to verify invariant stability under hostile environment degradation. | [`cain/trial/mutation_verifier.py`](file:///root/cain-studio-platform/cain/trial/mutation_verifier.py) |
| 7 | **RSAS** | **UNUSUAL** | **Risk-Stratified Assurance Scorecard** — Actuarial scoring substrate computing empirical Attack Block Rates (ABR), false-positive rates, and multi-threshold operating curves for regulatory compliance and cyber insurance underwriting. | [`cain/trial/scorecard.py`](file:///root/cain-studio-platform/cain/trial/scorecard.py) |
| 8 | **RTEB** | **ENGINEERING** | **Reproducible Trial Evidence Bundle & Standalone Verifier** — Court-admissible, WORM-exportable cryptographic trial evidence package bundling signed scenarios, execution traces, oracle verdicts, and scorecards with offline clean-room CLI verification. | [`cain/trial/bundle.py`](file:///root/cain-studio-platform/cain/trial/bundle.py)<br>[`cain-mesh-verify/verify.py`](file:///root/cain-studio-platform/cain-mesh-verify/verify.py) |

---

## 2. Exemplary Patent Claim Drafting (USPTO Format)

### Claim 1: Multi-Turn Adversarial Tool Mediation and Runtime Containment (MCAT / EGAA)
> **What is claimed is:**  
> A computer-implemented method for mediating and assuring multi-turn tool execution by autonomous artificial intelligence agents, comprising:  
> 1. Receiving, at an execution mediator distinct from an agent reasoning core, a multi-stage scenario contract comprising an agent identity, permitted capability subset, and resource budget ceiling;  
> 2. Intercepting an action proposal dispatched by the agent proposing invocation of an external tool with target arguments;  
> 3. Verifying, prior to invoking the external tool, that the proposed action matches a valid Proof-Carrying Action conforming to a pre-established policy intermediate representation;  
> 4. Upon detecting that the tool invocation is unpermitted, terminating execution fail-closed without invoking the tool and recording an authorization rejection;  
> 5. Upon permitting execution, capturing physical execution effects and inspecting output data for indirect prompt injection directives;  
> 6. Sanitizing output data upon detecting prompt injection markers and transmitting the sanitized output to the agent; and  
> 7. Appending the mediated result to an append-only causal execution trace signed with an asymmetric cryptographic key.

### Claim 2: Anti-Gaming Benchmark Harness Cryptographic Sealing (HIEC)
> **What is claimed is:**  
> A system for guaranteeing benchmark harness integrity against evaluated autonomous agents, comprising:  
> 1. A cryptographic sealer configured to compute canonical cryptographic digests over a plurality of benchmark evaluation files, answer keys, and environment variables prior to agent execution, generating a signed harness seal;  
> 2. A containment monitor tracking file modifications, process environment overrides, and module cache mutations during agent execution;  
> 3. An independent outcome audit engine comparing post-execution file states against the signed harness seal; and  
> 4. A gating verifier configured to immediately revoke agent autonomy and invalidate benchmark scores upon detecting modification or deletion of any sealed fixture.

### Claim 3: Long-Horizon Salami Attack Cumulative Risk Analysis (LHSE)
> **What is claimed is:**  
> A non-transitory computer-readable medium storing instructions for detecting multi-turn cumulative exfiltration attacks by an artificial intelligence agent, comprising:  
> 1. Maintaining a sliding-window record of resource consumption and capability invocations across a sequence of discrete interaction turns;  
> 2. Evaluating whether a newly proposed micro-action causes cumulative lifetime spend to exceed a pre-allocated budget, notwithstanding that the micro-action individually satisfies per-step limits;  
> 3. Evaluating whether capability invocations across successive turns create a prohibited capability synergy comprising a data reading capability and a network exfiltration capability; and  
> 4. Emitting an autonomous containment intervention triggering immediate execution halt when budget exhaustion or prohibited capability synergy is detected.

---

## 3. Evidence Gates & Mathematical Verification

- **Formal Property Suite:** `tests/test_evolution5_properties.py` (16 formal properties T01–T16, 100% PASS).
- **Adversarial Breaker Suite:** `tests/test_evolution5_breaker.py` (24 attack vectors, 100% PASS).
- **Performance Benchmark:** `tests/benchmark_evolution5_trial.py` (>12,000 ops/sec, <75 µs p50 latency).
- **Clean-Room Independent Verification:** `cain-mesh-verify/verify.py trial-bundle` (0 external dependencies, 100% verified).
- **Total Regression Suite:** 182 passed across all 5 evolutions in 3.71s.