# Evolution 81 — Research Bibliography

## How this list is used

Each item is a primary source for a design choice, followed by what was **actually adopted**, what
CAIN already had, and what remains missing. No source is represented as existing CAIN functionality.
No external dataset was downloaded in this evolution.

Current identifiers were confirmed by targeted web search on 2026-11-16 (queries and URLs recorded
in `artifacts/e81/research_log.json`): MCP current revision **2025-11-25**; LongMemEval
**arXiv 2410.10813** (ICLR 2025); LoCoMo **arXiv 2402.17753** (ACL 2024); MemoryAgentBench
**arXiv 2507.05257**. Search results confirmed identifiers only; they are not evidence that CAIN
implements any cited system.

| # | Source | Relevance | Adopted here | Still missing |
|---|---|---|---|---|
| 1 | Castro & Liskov, *Practical Byzantine Fault Tolerance* (OSDI 1999) | quorum reasoning behind CAIN-42 | not re-implemented; CAIN-42 quorum is external to E81 | an E81 mission anchored by an independent validator quorum |
| 2 | RFC 6962 (Certificate Transparency) Merkle tree | tamper-evident inclusion | domain-separated Merkle root in the evidence manifest | an inclusion/consistency proof API and a public log |
| 3 | RFC 8785 (JSON Canonicalization Scheme) | deterministic signing input | canonical JSON for every signature | full JCS number canonicalization review |
| 4 | SQLite WAL documentation | durable single-file state | WAL + `synchronous=FULL` + `BEGIN IMMEDIATE` | multi-writer/PostgreSQL durability at scale |
| 5 | Bubblewrap documentation | unprivileged sandboxing | namespaces, read-only binds, cap drop, rlimits | user-namespace policy on arbitrary hosts |
| 6 | gVisor / Kata Containers / Firecracker documentation | stronger isolation | probed and reported absent | installation and escape testing |
| 7 | MCP specification, revision 2025-11-25 | tool boundary | tool calls are Gate-gated in the reference path | production MCPGate proxy integration for E81 |
| 8 | A2A protocol specification | agent-to-agent delegation | delegation attenuation is implemented locally | an A2A transport and cross-organization test |
| 9 | OWASP Agentic Security Initiative (ASI01–ASI10) and AST10 | threat categories | 60-scenario lab covers injection, tool abuse, identity, memory, evidence | a formal mapping of every scenario to every ASI item |
| 10 | LongMemEval, arXiv 2410.10813 (ICLR 2025) | long-term memory evaluation | adapter contract + honest BLOCKED | dataset acquisition and a real score |
| 11 | LoCoMo, arXiv 2402.17753 (ACL 2024) | long-conversation memory | adapter contract | dataset acquisition and a real score |
| 12 | MemoryAgentBench, arXiv 2507.05257 (2025) | agent memory abilities | adapter contract | dataset acquisition and a real score |
| 13 | Sigstore / Rekor | transparency | bundle signing pattern; no Rekor entry | a public transparency-log entry |
| 14 | in-toto / SLSA | provenance | manifest schema inspired by attestations | a published SLSA provenance |
| 15 | OpenTelemetry GenAI semantic conventions | model/tool telemetry | model call fields (model, provider, tokens, latency) | OTLP export |
| 16 | Hugging Face model cards (Qwen2.5, SmolLM2) | local model selection | three cached CPU models measured | larger/frontier models |
| 17 | *Lost in the Middle* (arXiv 2307.03172) | context placement | short, bounded context in memory prompts | long-context retrieval study |
| 18 | *Chain-of-Thought* / self-consistency (arXiv 2201.11903, 2203.11171) | multi-sample voting | arms C/M implement voting | a model large enough for CoT to help |
| 19 | *Debate* / multi-agent collaboration literature | collective intelligence | arms B–O implement and measure; H1 rejected | a positive, significant collaboration result |
| 20 | Kubernetes Agent Sandbox (SIG) | agent isolation on K8s | architecture noted | no cluster, not tested |

## Research honesty

- All model claims come from the measured rows in `artifacts/e81/model_evaluation.json` and
  `collaboration_rows.jsonl`, not from vendor marketing.
- The negative collaboration result is reported as measured. It was not tuned away.
- The bibliography lists what was **not** obtained (no Rekor, no gVisor, no public memory datasets,
  no remote provider, no third-party audit).
