# E86 Baseline — E85 Forensic Reproduction

**E86 branch:** `evolution-86` at `/root/cain-e86-worktree`, created from E85 `06da643ff812dc315caade173b871bd5c6e16121`.
**E84 baseline:** `2c419a042cfbfbb3b136d6fdf530a1fde5f566d3` (ancestor).
**E85 branch preserved:** `/root/cain-e85-worktree`, not modified by E86.
**E82 is a parallel lineage**, not an ancestor of E85; it is not merged.

## Reproduced before any E86 change

| Check | Result |
|---|---|
| E85 tests (`tests/test_e85_*.py`) + E84 receipt test | **40 passed** in 61.04 s |
| E85 offline verifier | **43/43 checks**, exit 0 |
| E85 bundle Merkle root | `f5098de95587af63e5c69e6d4de704a3bf0d34570a8805270b8af4812b40a7e4` |
| E85 acceptance | **20 PASS / 15 PARTIAL / 2 FAIL / 1 BLOCKED / 1 NOT EVALUABLE** |
| E85 decision | **NOT RELEASED** |
| E85 HEAD | `06da643ff` |

The E85 worktree had three untracked frontier-evaluation files from a concurrent session
(`artifacts/e85/frontier/`, `artifacts/e85/frontier-smoke/`, `scripts/evolution85/run_frontier_eval.py`).
E86 did not commit, delete or modify them.

## What E85 actually measured (from its artifacts and verifier)

- 100 held-out tasks, 10 families, 9 arms; 2,122 real local model calls; **0 remote**.
- Single 0.5B: 24/100 (24.0%); 1.5B: 40/100 (40.0%); 135M: 10/100 (10.0%).
- Learned strategies: 25/100 (25.0%), delta +0.010, exact McNemar p=1.0 — **not significant**.
- Best collaboration `C_hierarchical`: 26/100 at 4.12× tokens — **not significant**.
- Unseen-domain transfer: 2/10 single, 2/10 transferred — no gain.
- 10,000 registered identities; up to **500 real worker processes**; 3,600.0 s endurance; 14,400 effects.
- Kill switch: 10 scopes, 33/33 adversarial scenarios refused. At 500 workers, **328 commits landed after the engage timestamp** and `effects_match_done` was false. The hard kill stops new claims; in-flight effects still land. Sub-second stop is met at ≤100 workers, not 500.
- Three vertical prototypes: evidence pack precision/recall 1.0/1.0; fabrication audit 1.0/0.7; MCP tool review 0.8889/0.8.
- 0 independent witnesses, no third-party audit, no hosted MCP receipt integration, no production deployment.

## Inconsistencies found

1. **Public homepage E85 strip vs E82 strip.** Both exist on the live homepages. E85 is the newest
   evolution; E82 is a parallel intelligence/durability lineage. The E86 evidence registry treats
   **E85 as the current baseline** and links E82 as a separate lineage. Neither is rewritten.
2. **Frontier evaluation files are untracked** in the E85 worktree. They are not part of E85's
   signed bundle and must not be counted as E85 evidence.
3. **The 500-worker result is PARTIAL**, not a kill-switch pass. E86 must not repeat the stronger
   claim.

## E86 starting point

E86 inherits E85's measured components and must add the missing infrastructure planes:

- a provider-neutral runtime with explicit capability profiles and budgets;
- a durable mission contract whose completion requires evidence;
- typed memory with provenance and a governed learning pipeline;
- a workforce dispatcher that separates registered identities from running processes;
- three vertical missions with objective verification;
- an external-agent interoperability flow;
- an artifact-derived acceptance matrix and an offline verifier.

E86 does not inherit a right to claim AGI. It inherits an architecture and a set of honest gaps.
