# E113 — collective intelligence study (INTERIM, 2026-11-19T23:01:31Z)

The experiment is running. This bundle publishes only what has been measured; **no primary result exists yet.**

- Question (preregistered): does a model-backed, governed collective beat the strongest individual agent
  under matched compute? Primary endpoint: sealed accuracy, collective G vs A* (strongest individual with
  self-consistency at >= the collective's measured CPU), exact McNemar, alpha 0.05.
- Sealed split hash `db01f037df26e905d1c3a5a525d113ebc186d4560ce734b820621bb4137c155e` is committed in the preregistration before any sealed model output.
- Real local models only (Qwen2.5-0.5B, Qwen2.5-1.5B, SmolLM2-135M on CPU); cloud BLOCKED.
- Baseline: E112 verifier 33/33 and E111 87/87 reproduce; E112 headline numbers recomputed from raw rows.
  E111 is not published (HTTP 404).
- Governance lab: 18/18 attack cases refused as expected;
  5/5 multi-process KILL races with 0 effects after the fence
  (KILL engaged by a separate process). One real bug found and fixed: a delegated permit could outlive its parent.
- Model calls recorded so far: {'dev': 330, 'val': 330}.
- Not claimed: any intelligence result, AGI, independent review, production MCPGate.

Verify: `python3 -I verify_e113_interim.py .`
