# E110 Part VII: governed specialist collective (gate G9)

Session b5, on the unit of work specified by the organism session: one E108 practice experiment (deterministic,
CPU-bound, about 2-4 s). FIRME AI CORPORATION, a three-person team. Workers are real OS processes (distinct PIDs
recorded). Fresh seeds disjoint from E108 sealed sets. Raw results: `artifacts/e110/collective/`.
Host: 8 CPUs, shared with other jobs (load about 10), so absolute times are conservative.

## A. Matched total compute (B = 40 experiments, W = 1, 4, 8, 16, 32 workers)

- **Pooled rows are byte-identical across all team sizes** (SHA-256 equal for families 101, 102, 105). At matched
  compute a team cannot know more than one agent: **quality gain = 0 by construction**, and the gate verdicts are the
  same (uniform 101 stays OFF; edge 102 and barrier_adj 105 detected, p = 0.001).
- **Speed:** 1 → 8 workers cut wall time 2.1x (101), 2.9x (102), 3.8x (105); no further gain beyond 8 on 8 CPUs.
- **Coordination overhead** (wall minus the busiest worker) 0.03-0.53 s; **0 duplicate and 0 lost experiments**.
- The small p differences seen across team sizes (0.863-0.879 on 101) are Monte Carlo noise from the order rows enter
  the gate's permutation null (diagnosed by the organism session; ±0.013 standard error); the harness now gates on
  rows sorted by world id.

## B. Matched wall clock (90 s per team)

Larger teams ran more experiments (W=1: 16-20; W=4..32: 54-111, limited by the shared host), but **every team size,
including a single worker, reached the correct verdict** (uniform OFF, edge ON). At this task size more workers buy
speed, not better answers.

## C. Byzantine workers (uniform family 101: no rule exists, the verdict must stay OFF)

Adversary: each Byzantine worker rewrites every row it produces into a consistent fake "edge" rule (cause in column 0).

| Adversary | Fabricated rows | No verification | Independent recomputation |
|---|---|---|---|
| 1 of 8 (weak, first run) | 1 | OFF (not fooled) | rejected 1/1, OFF |
| 3 of 8 | 5 | OFF (p 0.719, not fooled) | rejected 5/5, OFF |
| **7 of 8** | **17** | **FOOLED: false rule ON (p = 0.001)** | **rejected 17/17, correctly OFF** |

A majority of workers never makes a claim true; only recomputation does (each submitted row is re-run from its spec
by an independent verifier; any mismatch is rejected). Cost: about 36 s of re-execution, roughly the team's own wall time.

## G9 status: PARTIAL

The collective gives measurable **speed** and **Byzantine robustness**, but **no quality gain over a matched
single-agent baseline** on this task. The deterministic unit makes that structural: G9's "intelligence beyond a single
agent" needs tasks where workers explore different hypotheses, not the same deterministic experiment split.
