# Evolution 80 — Architecture Gap Analysis (E79 → E80)

Written before implementation. It lists what Evolution 79 could not do, what Evolution 80 builds,
and what stays out of reach on this host.

| # | Capability asked of E80 | State in E79 (verified) | Gap | E80 plan | Expected class |
|---|---|---|---|---|---|
| 1 | Choose the reasoning architecture per task | Every task took the same arm; a calculator was forced onto label questions and broke them (14/104) | No routing, no applicability check | Capability-aware router: families from question text, per-family policy fitted on train, frozen before test | Measurable |
| 2 | Know when collaboration helps or hurts | Four multi-agent arms, none significantly better than one agent | No analysis of why; no controller | Ten architectures on the same tasks; adaptive graph that adds a second path only when it can change the answer | Measurable |
| 3 | Heterogeneous models | One model (0.5B) in the benchmark; a 1.5B probe | No registry with measured capability | Three local models measured; registry with license, weight hash, latency, format compliance, per-family accuracy | Partial — three small CPU models only |
| 4 | Skill acquisition | A self-improvement candidate, not promoted; no training | No learning from data | Hand-written LoRA on a 135M model; untouched baseline; held-out, transfer, regression; versioning and rollback | Measurable, narrow |
| 5 | Memory | 10-question constructed corpus, 0.10 → 0.70 | Small; no encryption at rest; no forgetting | Typed store with provenance, validity, supersession, quarantine, tenant-bound encryption, tombstoned forgetting; five-ability evaluation | Measurable, constructed |
| 6 | Autonomous research | One fixed question; replication on an overlapping set | Replication not independent | Pre-registered hypotheses, frozen policy hash, replication in a separate process on a disjoint split | Measurable |
| 7 | Self-improvement | Candidate awaiting approval | No failure clustering | Failure clusters from raw rows → candidate → held-out comparison → approval withheld | Partial |
| 8 | Agentic 75 | 75 importable, 0 integrated in E79 | No dependency graph, no composition | Audit of all 75, capability graph, composition inside a mission | Partial |
| 9 | Durable execution | 4 worker kills, 0 duplicates | No coordinator kill, stale-writer, storage-failure or interrupted-inference test | Drill with all of those, real OS processes | Measurable on one host |
| 10 | Security | 57 scenarios | New surfaces (router, skills, expression execution) untested | New lab against the E80 gate, sandbox, memory, evidence | Measurable, same-project |
| 11 | Mission control | Grounded synthesis mission | No dependency graph across organizations | Mission graph with dependencies, budgets, replanning and evidence-based completion | Partial |
| 12 | ClawX | Local state file | Not authenticated, not deployed | Backend state for the 20 views, token-gated local endpoint | Partial — not deployed |
| 13 | MCPGate | Tool gate per agent | No single-use grants on model calls | Every model and tool call needs a signed single-use grant; bypass tests | Partial — reference gate, not the production proxy |
| 14 | Evidence | 144-check verifier | — | New bundle and verifier recomputing E80 numbers from raw rows; second-host check | Measurable, self-signed |

## Gaps E80 cannot close on this host

| Gap | Why | Class |
|---|---|---|
| Remote model providers | No credential, no funding; purchase prohibited | BLOCKED |
| Models ≥ 3B at benchmark volume | CPU-only shared production host | NOT EVALUATED |
| Public benchmarks (MultiAgentBench, AgentBench, LongMemEval, LoCoMo, MemoryAgentBench, AMA-Bench) | Context length and compute | NOT EVALUATED |
| gVisor / Kata / Firecracker isolation | Not installed | NOT EVALUATED |
| Cross-host model execution, network partition | Second host has 2 vCPU / 3.4 GB | NOT EVALUATED in E80 |
| 6 h – 168 h endurance | Session and host budget | BLOCKED |
| Independent witness, transparency log | Publication not authorized | BLOCKED |
| Third-party audit | None commissioned | BLOCKED |

## Design rule carried through every component

Intelligence proposes; it never authorizes. The router returns a path name. Memory returns text
with `grants_authority: false`. A skill is a file with a hash. Only the gate's signed, single-use,
expiring grant lets anything run, and only a human can promote a skill or a policy.
