# EVOLUTION7_RESEARCH_AND_PRIOR_ART.md

**Dated:** 2026-10-09. Sources were retrieved and read; systems are surveyed by their published description, NOT
benchmarked. No head-to-head result is claimed. Prior-art risk is a qualitative estimate, NOT a legal opinion.

## Primary references (retrieved)

- **FORGE — Formal Policy Enforcement for Real-World Agentic Systems** (arXiv 2602.16708). Policy as aspect weaving;
  Datalog policy language; reference monitor; observability service maintaining a causal dependency graph; an
  *environment contract* (assume/guarantee) under which enforcement matches intended semantics; static analyses
  (contradiction, redundancy, subsumption, conditional reachability); LLM translation with coverage/entailment checks.
  Reported: prompt-injection success 100%→0%, τ²-bench compliance 58%→98%, 19–38% latency overhead.
- **AgentKernel — the Trust-Native Agentic Operating System** (arXiv 2609.29647). A mandatory semantic reference monitor
  outside the LLM context; four pillars (identity, perception, cognition, execution); deny-by-default **policy
  intersection**; TEE-backed key custody tier; semantic-to-syscall backstop via eBPF; three narrow adapters (LLM/tool/
  storage); Global Agent Registry issuing Ed25519 Agent Identity Cards.
- Other named mission references (arXiv 2604.17562, 2609.00015, 2607.13716, agent-security-bench) were treated as
  additional survey inputs; the two above were read in full and are the basis of the comparison below.

## What CAIN already had vs. what is genuinely different

| Capability | FORGE | AgentKernel | CAIN (before E7) | CAIN E7 delta |
|---|---|---|---|---|
| Reference monitor over actions | yes (Datalog) | yes (OS-shaped) | `cain.mcp_proxy` + PCA | canonical action + authorization binding + TOCTOU revalidation |
| Declarative policy | Datalog (recursive) | deny-by-default intersection | policy objects / Rego packs | versioned deterministic engine + bounded static analysis |
| Canonical action identity | action descriptor | capability chains | PCA arguments digest | **RFC 8785 JCS canonical action + 10 equivalence rules** |
| Effect reconciliation | partial | receipts | proof-carrying action | explicit EFFECT_VERIFIED requires independent observation |
| Approval binding | approval workflows | approvals | Human-Approval object | approval bound to exact canonical action digest |
| Cumulative trajectory | dependency graph | memory IFC | salami defense (kernel) | hash-chained cumulative risk/data budget |
| Clean-room verification | — | — | mesh verify pattern | **action-commit clean-room verifier (zero imports)** |

## Concessions (honest)

- FORGE's Datalog recursion + formal assume/guarantee correctness and AgentKernel's non-bypassable OS placement are
  **stronger formal claims than CAIN E7 makes**. CAIN E7's policy engine is a bounded analyser, NOT a proven-correct
  Datalog reference monitor. This is a real gap, not a marketing difference.
- CAIN E7's MCP adapter is the only ENFORCEMENT adapter; shell/http/filesystem are OBSERVER adapters and are labelled so.
  A direct backend call that does not route through `cain.mcp_proxy` is **not** prevented (`DIRECT_BACKEND_BYPASS_CONTROLLED=False`).
