# Evolution 82 Final Audit

## A. Inherited from E81

E81 head `43420d3be6776a926956b85433b2b070f46c83fd`, parent `b40f656fdbad66fd6ae646aabdee01d5991e3995`, 64/64 tests, 18/18 offline
checks, preserved artifacts and negative findings. E80 publish commit `a8c02bd73` is not an
ancestor; it is a publish script only and was not merged blindly.

## B. Implemented in E82

Accounting and audit-chain integrity, write-once registration, provider architecture, a fixed
benchmark with hidden holdout, ten collaboration arms, ten-discipline specialist matrix, routing,
backup/restore/RPO/RTO, cross-host failover with signed single-use authorisation, real host-reboot
recovery, governed scoped memory with encryption and poisoning defence, staged model-worker
capacity, real endurance, a read-only operator view, artifact-derived gates and a signed offline
evidence bundle.

## C. Defects fixed

F-01 call undercount; F-02 forked/tail-only audit validation; F-03 benchmark call accounting;
F-04 post-hoc registration; F-05 refuted at baseline and replaced with a real backup/restore test.
E81's hard-coded gate statuses were replaced with artifact-derived gates.

## D. New capabilities demonstrated

- Cross-host mission recovery: **COMPLETED**, 0 duplicate effects, stale
  commit refused, event and checkpoint prefixes preserved, RPO 0.025 s,
  RTO 2.557 s.
- Real reboot recovery with an active mission: boot id changed
  `4a5b981b-e662-4c08-ac56-d1513ea74821` -> `eb40c04f-3a7a-4ace-9833-f898e0591c23`, {'DONE': 20},
  durable identity survived, 0 duplicate effects.
- 10 real simultaneous local model inferences.
- 86/86 adversarial scenarios blocked.

## E. Still experimental

The 1.5B/0.5B/135M models are small CPU models. Collaboration showed no significant advantage.
Specialist prompts did not improve overall accuracy. 25 process-local workers are memory-blocked;
50/100 workers share one serialised model. Long endurance beyond the measured campaign is not
attempted. Memory contradiction detection is lexical/key-based. No third-party audit.

## F. Tested against real models

Local only: Qwen2.5-0.5B-Instruct, Qwen2.5-1.5B-Instruct, SmolLM2-135M-Instruct. **0 remote calls.**

## G. Tested using local models

Benchmark (126 calls), collaboration (367 calls),
specialists (200 rows / 300 real calls), capacity (10/50/100 workers), endurance
(165 calls).

## H. Mocks or simulations

The HTTP adapter self-test uses an in-process stub labelled `LOCAL_HTTP_TEST`; it is not counted
as a remote call. The lost-acknowledgement scenario explicitly models a lost acknowledgement
against the real SQLite store. Cross-host and reboot missions use synthetic deterministic jobs.
The 50/100-worker stages use one real resident model with a serialised request queue. No mock
response is counted as a real inference.

## I. Actual intelligent worker concurrency

10 simultaneous local model inferences; 50/100 concurrent workers against a serialised shared
model (simultaneous inference 1). 25 process-local workers blocked by memory.

## J. Distributed recovery

Cross-host PASS; real host reboot PASS; lost acknowledgement deduplicated.

## K. Endurance

COMPLETED, 3600.0 s wall clock, 165 calls.

## L. Model calls and cost

Native E82 governed-inference ledgers: **861 terminal attempts**,
861 successful, 0 failed, 1 incomplete;
by class {'REAL_LOCAL_MODEL': 861}. Capacity worker processes made a further
**321 real model calls** recorded in their own JSON evidence.
Total real model calls: **1182** (of which
6 are the exploratory smoke run, excluded from confirmatory
analysis). E81 is reconciled separately at
1589 unique invocations. **Billed cost $0.00**; host time is not priced.

## M. Governance and security

86/86 blocked. Denied/expired/revoked/replayed grants never
executed. Worker grants refused before model load. Memory never grants authority.

## N. Evidence

Bundle `artifacts/e82/bundle`, Merkle root `63868034b46e901f2cec392cd25c22ab947516e594543e02155857a9c31c5ac9`, 240 files,
offline verifier **55/55**, tamper tests
**10/10 rejected**. Key class:
`SELF_GENERATED_OPERATOR_KEY`; no independent witness.

## O. Cross-host verification

Host B `45.76.169.191` in the same administrative domain. This is not third-party
verification. Clean-clone reproduction: commit `6b22cb5acd6d2e4229295c296980697f67d419a8`, **36 tests passed**,
offline verifier **55/55** (same host and Python environment; not a
third-party reproduction).

## P. Acceptance matrix

{"BLOCKED": 2, "PASS": 42} across 44 gates. Full matrix in
`E82_ACCEPTANCE_MATRIX.json`.

## Q. Release decision

**CONDITIONALLY RELEASED**

## R. Remaining blockers

- `G11` **BLOCKED**: Remote provider inference demonstrated or BLOCKED (reason: No provider credential was injected into the isolated runtime; 0 remote calls were made.)
- `G33` **BLOCKED**: ClawX public integration status accurately classified (reason: BLOCKED: no production DNS, credential or owner authorization was used; clawx.click is not wired to this service)

## S. Recommended next Evolution

1. Inject a real provider credential and run the pre-registered remote benchmark under a hard
   dollar ceiling; compare frontier vs local on the same rows.
2. Raise process-local worker memory efficiency (quantisation / shared weights) so the 25-worker
   stage can run, then measure 50 simultaneous inferences.
3. Run 6h and 24h real endurance campaigns (then 72h) while sampling memory and storage.
4. Replace lexical memory-contradiction detection with typed entailment and add off-host memory
   replication.
5. Obtain a third-party review of the evidence bundle and a witness cosignature for the Merkle root.
6. Repair and re-qualify the quarantined PBFT cluster before using it for any customer evidence.

## T. Evidence provenance and cross-session coordination

- E82 code/evidence commit: `28f44a1cc5644328f0a49e12a65f61330a9cf14c`.
- The evidence bundle's `SOURCE.json` records `2d9c6f3` and `dirty: true` because the final gate
  and evidence files were being written at build time. The bundle's own file hashes, Merkle root
  and Ed25519 signature are what the offline verifier checks; the final handoff commits add only
  documentation and the replication record.
- The E81 branch advanced to `f2673a6ad` (owner-authorized static evidence publication) after the
  E82 baseline. E82 recorded the divergence and merged `f2673a6ad` with `--no-ff` after review;
  no E81 history was rewritten.
- E84 session: `/root/cain-e84-worktree`, branch `evolution-84`, based on E83 `215f2b74f`. E82 is
  **not** in that ancestry and E84 is **not** in E82's. They are complementary lineages; do not
  merge blindly. E82 wrote only `EVOLUTION84_COORDINATION.md` and a read-only peer review of
  `cain/receipt/core.py` into the E84 worktree, with findings F-E84-01 (signed checkpoint
  overwritten at the same size), F-E84-02 (consistency proof recomputes the old root), F-E84-03
  (no contiguous-index/previous-leaf check), and F-E84-04/05 (no witness, plaintext payloads).
  The E84 Merkle inclusion/consistency code passed an exhaustive 1..64 test.
