# E114 architecture

```
                  Cognitive Research Director (proposes; cannot accept its own conclusions)
                     │ objectives → questions → hypotheses → ranked backlog
                     ▼
   DecisionPoint (CAIN decides, authority key) ──grant──► Enforcer (MCPGate role, pins authority + kill keys)
                     ▲                                       │ admits / refuses, signed hash-chained receipts
       human approver key (high-risk, promotion)             ▼
                                                    Experiment executes (world model, active learning,
   KillSwitch (independent kill key, same lock) ───►  continual learning, skills, self-improvement, mission)
                                                             │ raw rows
                                                             ▼
                          Independent Verifier (own key) recomputes → signed verdict
                                                             │
                                                             ▼
               Evidence-Native Memory: immutable sources → staged → validated (validator ≠ author) → versioned heads
                                                             │
                                                             ▼
                        Knowledge graph + backlog → next research cycle → bundle → offline verifier (no CAIN imports)
```

Contracts:
- **Grant** `cain.e114.grant.v1`: agent, action, resource, params_hash, data_label, nonce, iat, exp, signed by the authority key.
- **Verdict** `cain.e114.verdict.v1`: claim, CONFIRMED/REJECTED, checks, recomputed metrics, raw sha256, signed by the verifier key.
- **Memory record**: (rid, version) immutable rows; the visible state is a movable head; every change is journaled in a hash chain; `grants_authority: false`.
- **Training record** `cain.e114.training_record.v1`: start checkpoint, dataset hashes + provenance, config, evaluation suite, compute, model version, regression.

Subsystem status: implemented and exercised — director, planner/backlog, knowledge graph, memory, world model, continual learning, skills,
experimentation loop, model registry, verifier, governance, mission engine. Delegated: collective intelligence (E113).
Partial: scientific discovery (no automated literature ingestion), ClawX registry (static bundle).
