# Evolution 81 — Implementation Report

> FIRME AI CORPORATION. Branch `evolution-81`. Baseline Evolution 80 at
> `cab2fa162`. This report is generated from `artifacts/e81/`; every number below is
> read from an artifact produced by an executed run.

## A. Executive summary

**What was actually built.** A new product-level layer `cain/cain2027/e81/` over the existing
CAIN-42 governance substrate: durable mission runtime with leases/fencing/idempotency/checkpoints,
a 30-discipline / 90-profile specialist registry with attenuated composition, a model-agnostic
inference layer (3 real local models measured; 4 remote providers registered but BLOCKED without
credentials), 60-scenario adversarial lab, encrypted temporal organizational memory, sandboxed
execution with measured isolation, resource governance, cross-host execution, a signed evidence
bundle with a standalone verifier, and an authenticated ClawX evidence backend.

**What works.**
- 30 disciplines, 90 specialist profiles, 8 domain commanders,
  99 durable identities.
- Durable runtime: process-kill recovery `PASS`, duplicate effects
  `0`, stale late commit `STALE`.
- Security lab: **60/60 blocked, 0 bypassed**.
- Sandbox: network egress blocked `True`; host secrets
  unreadable `True`.
- Real local inference: **1475** calls in the main ledger (+ 114 in the endurance ledger; the cost report's own count of 1156 omitted rows under an older label), 70743 input
  tokens, 15818 output tokens, $0 model cost.
- Collaboration: **COLLABORATION_NOT_SUPERIOR_HONESTLY_REJECTED** — best arm
  `B_hierarchical` at 41.7% vs
  single agent 41.7% (delta 0.0,
  p=1.0).
- Evidence: bundle verifies offline `True`; Merkle root
  `d744ab90f9e6141c37097a0b…`.

**What does not work / is not proven.**
- Remote model providers: **BLOCKED** (no credential). 0 remote calls.
- Public memory benchmarks (LongMemEval / LoCoMo / MemoryAgentBench): **BLOCKED** (datasets absent).
- Host reboot recovery: **not run** (process restart only).
- 1,000 concurrent OS workers: **attempted and blocked** by SQLite write contention; 500 real
  workers is the measured maximum on this host. 5,000/10,000 are logical identities only.
- Endurance: 1 h target reached `True`; 6/24/72/168/720/1,000 h **not run**.
- Independent verification: the verifier is a separate tool **written in this project**; no third
  party, no transparency log, no independent witness. Nothing is labelled INDEPENDENTLY VERIFIED.
- ClawX is an authenticated **local** service, not deployed to clawx.click.
- No production change of any kind was made.

## B. Repository status

| field | value |
|---|---|
| branch | `evolution-81` |
| HEAD at run | `cab2fa162ae33beb29ecacad7789acbe40dd9c54` |
| final E81 commit | `31e8ef8ad38c75ea6985cc978ff0424001b212fa` |
| final parent | `cab2fa162ae33beb29ecacad7789acbe40dd9c54` |
| parent | `ca043f22cd64d1ac4733bdde7f07e5adc252c91d` |
| E80 baseline | `cab2fa162` |
| E79 baseline | `f4449aa7913623bc034fa4434ca96c2a2919a889` |
| E80 tests | 30 passed |
| E80 offline verifier | `True` |
| E80 gate counts | `{'BLOCKED': 5, 'FAIL': 1, 'NOT EVALUABLE': 3, 'PARTIAL': 10, 'PASS': 13}` |

## C. Workforce status

| metric | value |
|---|---|
| registered logical agents | 90 |
| specialist profiles | 90 |
| disciplines | 30 |
| domain commanders | 8 |
| durable identities | 99 |
| active workers at snapshot | 0 (capacity runs record their own counts) |
| real completed tasks (reference mission) | 6 |
| mission verdict | `COMPLETED_WITH_UNVERIFIED_ITEMS` |

Logical agents are records, never processes. Active workers, concurrent inference and completed
tasks are reported separately in `EVOLUTION_81_CAPACITY_PLAN.md`.

## D. Intelligence results

| model | family | params | correct | accuracy | mean latency |
|---|---|---|---|---|---|
| qwen2.5-0.5b-instruct | qwen | 0.5B | 6/12 | 50.0% | 1596.46 ms |
| qwen2.5-1.5b-instruct | qwen | 1.5B | 4/12 | 33.3% | 7502.31 ms |
| smollm2-135m-instruct | smollm2 | 135M | 0/12 | 0.0% | 1439.69 ms |

- Remote providers registered and **blocked**: openai-frontier, anthropic-frontier, gemini-frontier, deepseek-reasoner.
- Collaboration: `COLLABORATION_NOT_SUPERIOR_HONESTLY_REJECTED`. Recommendation: use the single agent; collaboration cost more without a measured benefit.
- Memory (constructed suite): governed 81.2%,
  ungoverned 75.0%,
  no memory 6.2%.

## E. Durability results

| check | result |
|---|---|
| process killed and replaced | `True` |
| leases recovered | 1 |
| late commit from dead worker | `STALE` |
| duplicate effects | 0 |
| checkpoint chain | `True` |
| host reboot | `NOT RUN — no host reboot was performed in E81` |
| 1 h endurance target | `True` (3600.0 s elapsed) |

## F. Governance and security

- Authorization: default-deny Gate; every model/tool call needs a signed single-use grant.
- Identity: Ed25519, durable, versioned; rotation invalidates old signatures.
- Delegation: child grants attenuate parent; composition intersects tools/scopes/risk.
- Sandbox: bubblewrap, no network, read-only system, private workspace, rlimits, timeout.
- Kill switch: cross-process test passes (4/
  4 kill-switch scenarios blocked).
- Security regressions: 60/60 blocked, 0 bypassed.

## G. ClawX

Authenticated local service verified: unauth 404 `True`,
auth 200 `True`, events 6.
**Not deployed.** No live streaming, no RBAC beyond the operator token, no public demo enabled.

## H. Evidence

- Bundle: `artifacts/e81/bundle` — 8 files, 6 events,
  6 tasks, 6 effects.
- Standalone verifier: `scripts/evolution81/verify_e81_offline.py`, result `True`.
- Second host: `True` (same administrative domain, not independent).
- Signer fingerprint `152231774fc0d4b5568a90615c281d26` — self-generated, not published out of band.

## I. Acceptance matrix

26 PASS · 12 PARTIAL · 0 FAIL ·
2 BLOCKED · 0 NOT EVALUABLE (40 gates).
Full table in `EVOLUTION_81_ACCEPTANCE_MATRIX.json`.

## J. Release decision

**NOT RELEASED.** Evolution 81 is a development-branch result. Production deployment is not
authorized; critical gates remain BLOCKED or PARTIAL.

## K. Next action

The single highest-priority engineering action: **run a 6-hour endurance campaign with a real
worker-kill at a phase boundary and a second-host recovery, then have the resulting bundle
re-verified on a host outside this administrative domain.** That is the shortest path from the
current PARTIAL G07/G23/G29/G32 gates to a defensible release candidate.
