# E120 Fresh Frontier Literature Review & Research-to-Implementation Traceability Matrix (2025–2026)

Entity: **FIRME AI CORPORATION** | System: **CAIN-42** | Evolution: **E120**

No paper terminology is copied as an unimplemented claim. Every adopted research idea below maps directly to an executable E120 module, falsifiable experiment, and measured outcome.

| # | Area | Publication & Date | Group | Actual Experimental Result | Limitations | CAIN-42 E120 Implementation Relevance | E120 Experiment & Measurable Outcome |
|---|---|---|---|---|---|---|---|
| 1 | Autonomous Scientific Discovery | *The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search* (2025) | Sakana AI & collaborators | Generated end-to-end ML research manuscripts that passed workshop peer review via agentic tree search + VLM feedback | Narrow ML benchmark focus; occasional hallucinated citations / experimental bugs without independent verifier | Motivates separating hypothesis generation, execution, adversarial review, statistical recomputation, and unilateral reproducibility veto | `cain/e120/research_loop.py` (G08): 8 distinct roles; AdversarialReviewer rejects director's shortcut; ReproducibilityAuditor replay+tamper veto PASS |
| 2 | Skill Abstraction & Program Induction | *DreamCoder* (Ellis et al., PLDI) & *Voyager* lifelong skill library (Wang et al., 2023–2025 follow-ups) | MIT / NVIDIA / Caltech | Wake-sleep library learning compresses program search depth; verified code skill libraries transfer across tasks | Fixed base DSLs; unbounded skill libraries can suffer negative transfer without validation gates | Motivates unifying E117–E119 into one invariant-indexed macro + parameterized-schema abstraction engine with counterexample falsification | `cain/e120/generalization.py` (G03–G06): 4 sealed OOD conditions (all Holm p<0.001) + retention 1.000 + negative-transfer rejection |
| 3 | World Models & Predictive Representations | *JEPA / V-JEPA 2* (2025) & *DreamerV3 / Genie 2* (2024–2025) | Meta AI / Google DeepMind | Latent predictive dynamics models improve sample efficiency and forward planning across domains | Passive observational world models remain vulnerable to hidden confounders without interventions | Motivates explicit separation of passive correlation vs interventional `do()` causal models + online changepoint belief revision | `cain/e120/world_model.py` (G07): Causal world model beats passive correlation & 2x-compute pattern matcher on confounded planning (+0.500, p<0.001) |
| 4 | Memory Consolidation & Catastrophic Forgetting | *Mem0 / Letta / Anthropic Memory Consolidation ("Dreaming") & EWC/Fisher Subspaces* (2025–2026) | Anthropic / Mem0 / Academic Labs | Offline consolidation from episodic logs into semantic/procedural stores reduces context bloat and preserves skills | Unfiltered memory is vulnerable to cross-session prompt-injection poisoning | Motivates 8-layer typed memory with SHA-256 provenance chains, contradiction rejection, TTL expiry, and consolidation | `cain/e120/memory.py` (G16): Provenance memory beats `no_memory` (+0.750, p=1.86e-9) and blocks 20/20 poison writes where unfiltered fails |
| 5 | Long-Horizon Agent Autonomy Evaluation | *METR Autonomous Time-Horizon Benchmarks* (2025–2026) | METR / Anthropic | Measures real wall-clock autonomous task horizons and failure modes over multi-hour runs | Short simulated demos overstate reliability; restart durability and audit continuity are critical | Motivates real wall-clock 6h (E119) & 12h (E120) missions with hash-chained journals, controlled SIGTERM resume, and honest RUNNING/PARTIAL labeling | `cain/e120/mission_12h.py` (G02, G17): Real wall-clock runs preserved untouched; zero simulated clocks |
| 6 | Recursive Self-Improvement Evaluation | *Internal R&D Automation & Sandboxed Self-Improvement Audits* (2025–2026) | Anthropic / OpenAI / UK AISI | Evaluates autonomous patch/config proposals against held-out regressions and safety invariants | Unconstrained self-modification risks reward hacking, overfitting to dev sets, or weakening safety guards | Motivates 14-candidate ledger across 5 splits (dev/sealed/retention/unfamiliar/adversarial) + immutable governance/key/verifier boundary | `cain/e120/self_improve.py` (G09–G11): 3 beneficial promoted, overfits/harmful rolled back, 3 governance-touching candidates hard-blocked as unsafe |
| 7 | Runtime Governance & MCP Security | *Model Context Protocol (MCP) Security & OWASP Agentic Top 10 (ASI01–ASI10)* (2025–2026) | Anthropic / OWASP GenAI / NIST AISI | Documents tool poisoning, shadowing, token replay, parameter tampering, and cross-session memory persistence | Prompt filters alone cannot govern execution channels | Motivates enforcing real `cain/mcp_proxy.py` default-deny allowlists + single-use Ed25519 capability tokens + SQLite WAL kill-switch across separate OS processes | `cain/e120/governed_mission.py` (G18): 9/9 separate-process boundary scenarios pass; 0 effects after KILL in resource hash-chained ledger |
| 8 | Multi-Agent Collective Intelligence | *Structured Debate, Verification-Guided Ensembles & Specialist Routing* (2025–2026) | OpenAI / DeepMind / Academic Labs | Unverified majority voting amplifies correlated errors ("false consensus"); verification + complementary specialization improves accuracy | Extra agents often win only by spending more total compute | Motivates strictly matched-compute 6-architecture benchmark + homogeneous control preserving zero-gain case | `cain/e120/collective.py` (G15): Full collective 1.000 vs strongest single 0.483 (p=9.31e-10) & independent voting 0.233 (46 false consensus) under identical budget |
