# E86 Research Review — What Changes an AGI Infrastructure Design

**Scope.** This is a design review, not a claim to have reproduced the cited papers. The URLs were
supplied with the E86 mission; their detailed results were not independently replicated here. Where
a statement is an architectural inference it is labelled as such.

## Sources reviewed

1. **UltraHorizon (ICML 2026)** — long-horizon agent evaluation.
   https://proceedings.mlr.press/v306/luo26ai.html
2. **AMA-Bench (ICML 2026)** — memory/agent benchmark.
   https://proceedings.mlr.press/v306/zhao26bs.html
3. **AgentWorld (2026)** — arXiv:2609.31590.
4. **Locating Hidden Failures Makes Long-Horizon Agents More Reliable (2026)** — arXiv:2609.17930.
5. **Continual Learning Exam (2026)** — arXiv:2610.33956.
6. **NIST, Software and AI Agent Identity and Authorization (2026)**.
7. **NIST, Design Tenets for Agent Identity Management Systems (2026)**.

## What the problems are

| Problem | What the literature points at | E86 implication |
|---|---|---|
| Long-horizon reliability | Success at one step does not compose into success over hundreds; hidden failures accumulate | Mission state must be durable and inspectable; completion needs evidence, not a final message |
| Memory | More context is not memory; retrieval must be provenance-aware and temporally valid | Typed memory with source, confidence, validity, supersession and revocation |
| Generalization | Performance on familiar task families does not establish transfer | Held-out and unseen-domain tasks; report per-domain, never aggregate-only |
| Continual learning | Learning can improve a measured slice and still regress elsewhere | Promotion requires held-out comparison, independent approval and rollback |
| Agent identity | A model or process is not an identity; authority must be delegated and attenuated | Durable Ed25519 identity, signed grants, expiry, revocation, fencing |
| Verification | An agent's self-report is not verification | Independent verifier, effect-bound receipt, offline check |
| Failure localization | Detecting *where* a long run failed is more useful than a final score | Per-step evidence and causal replay, not only terminal outcome |

## What has been demonstrated (architectural reading)

- Long-horizon benchmarks expose that small models and short-context agents degrade over long tasks.
- Memory benchmarks distinguish retrieval from reasoning; storing text is not learning.
- Agent identity work increasingly treats agents as non-human principals with delegated, scoped,
  revocable authority.
- Verifiable execution needs an enforcement point and a receipt bound to the exact action.

## What remains unsolved

- General cross-domain transfer at small model scale.
- Reliable self-improvement without external approval.
- Independent verification of a vendor's evidence at low cost.
- Immediate stop of in-flight effects across arbitrary resources.
- Ground-truth evaluation for open-ended professional work.
- Hardware-backed attestation on commodity hosts.

## E86 design consequences

1. **Plane separation** — runtime, cognition, memory, workforce, verticals and governance are
   separate interfaces. A model may change; authority and evidence do not.
2. **Evidence-gated completion** — a mission cannot be `COMPLETED_VERIFIED` unless every declared
   criterion has an evidence record and an independent check.
3. **Typed memory** — episodic, procedural, semantic, causal, mission, organizational, agent and
   evidence memory are distinct; none is authority.
4. **Governed learning** — experience → candidate → validation → comparison → external approval →
   promotion → monitoring → rollback. A proposer cannot approve its own change.
5. **Capability routing, not model worship** — route on measured per-domain capability, cost,
   latency and privacy; report BLOCKED when a provider is unavailable.
6. **Workforce honesty** — count registered identities, OS processes, active workers, model
   requests, completed tasks, verified outcomes and external effects separately.
7. **Stop semantics** — distinguish stop-new-work from stop-in-flight-effects. A kill switch that
   permits hundreds of post-engagement commits is not an immediate global stop.
8. **Interoperability** — an external agent must be able to register, request authority, submit a
   mission, call a governed tool, receive a decision, execute only if allowed, submit evidence,
   verify a receipt and be revoked.

## What E86 does not claim

No AGI. No independent witness. No third-party audit. No remote frontier result. No production
deployment. No claim that a paper's result transfers to CAIN without replication.
