# E91 Multi-Agent Report

## Run 1: Qwen2.5-0.5B (local), bug_fix + remediation

| config | tasks | accepted | rate | calls | accepted/100 calls | duplicates | false consensus | false accept | false reject |
|---|---|---|---|---|---|---|---|---|---|
| single | 2 | 0 | 0.0 | 2 | 0.0 | 0 | 0 | 0 | 0 |
| three_identical | 2 | 0 | 0.0 | 6 | 0.0 | 4 | 2 | 0 | 0 |
| three_diverse | 2 | 0 | 0.0 | 8 | 0.0 | 1 | 1 | 0 | 0 |
| specialists | 2 | 0 | 0.0 | 4 | 0.0 | 0 | 0 | 0 | 0 |
| specialist_verifier | 1 | 0 | 0.0 | 2 | 0.0 | 0 | 0 | 0 | 0 |
| hierarchical | 1 | 0 | 0.0 | 3 | 0.0 | 0 | 0 | 0 | 0 |
| governed_swarm | 1 | 0 | 0.0 | 4 | 0.0 | 0 | 0 | 0 | 0 |

Note: configs have unequal task counts so far {'single': 2, 'three_identical': 2, 'three_diverse': 2, 'specialists': 2, 'specialist_verifier': 1, 'hierarchical': 1, 'governed_swarm': 1}; the run was still in progress.

## Run 1: Claude Haiku 5.5, bug_fix + remediation (owner-approved, $5 shared cap)

| config | tasks | accepted | rate | calls | accepted/100 calls | duplicates | false consensus | false accept | false reject |
|---|---|---|---|---|---|---|---|---|---|
| single | 20 | 18 | 0.9 | 20 | 90.0 | 0 | 0 | 0 | 0 |
| three_identical | 20 | 19 | 0.95 | 62 | 30.65 | 26 | 0 | 0 | 0 |
| three_diverse | 20 | 20 | 1.0 | 62 | 32.26 | 20 | 0 | 0 | 0 |
| specialists | 20 | 20 | 1.0 | 47 | 42.55 | 0 | 0 | 0 | 0 |
| specialist_verifier | 20 | 20 | 1.0 | 23 | 86.96 | 0 | 0 | 0 | 0 |
| hierarchical | 20 | 20 | 1.0 | 40 | 50.0 | 0 | 0 | 0 | 0 |
| governed_swarm | 20 | 20 | 1.0 | 20 | 100.0 | 0 | 0 | 0 | 0 |

## Run 2: Claude Haiku 5.5, fault_localization + allocation, shared tool budget (families chosen from dev only; artifacts/e91/team/PREREGISTRATION_2.json)

| config | tasks | accepted | rate | calls | accepted/100 calls | duplicates | false consensus | false accept | false reject |
|---|---|---|---|---|---|---|---|---|---|
| single | 20 | 10 | 0.5 | 47 | 21.28 | 0 | 0 | 0 | 0 |
| three_identical | 20 | 9 | 0.45 | 136 | 6.62 | 21 | 7 | 0 | 0 |
| three_diverse | 20 | 5 | 0.25 | 148 | 3.38 | 14 | 9 | 0 | 0 |
| specialists | 20 | 3 | 0.15 | 58 | 5.17 | 0 | 0 | 0 | 0 |
| specialist_verifier | 20 | 8 | 0.4 | 62 | 12.9 | 0 | 0 | 0 | 0 |
| hierarchical | 20 | 7 | 0.35 | 71 | 9.86 | 0 | 0 | 0 | 0 |
| governed_swarm | 20 | 7 | 0.35 | 78 | 8.97 | 6 | 6 | 4 | 0 |

### Run 2 by family

| config | fault_localization | allocation |
|---|---|---|
| single | 5/10 | 5/10 |
| three_identical | 5/10 | 4/10 |
| three_diverse | 2/10 | 3/10 |
| specialists | 1/10 | 2/10 |
| specialist_verifier | 4/10 | 4/10 |
| hierarchical | 2/10 | 5/10 |
| governed_swarm | 2/10 | 5/10 |

### What run 2 shows

- No team configuration beat the single agent; several were much worse. Pre-registered T4 (verifier or swarm beats single on fault_localization) is **not supported**.
- Mechanism (from the rows): under the shared per-task tool budget, the verifier's probes use budget the producer needed. On fault_localization the verifier refutes the first candidate, after which later checks return NO_EVIDENCE because the budget is spent.
- On allocation the verifier can confirm feasibility only (pre-registered limit T6); the governed swarm's false acceptances come from feasible-but-suboptimal plans.
- Majority-vote configurations show false consensus (agreement on a rejected answer); consensus is not truth.
- Conclusion for this experiment: when tool budget is the binding constraint, verification overhead costs more than it recovers. Collaboration benefit was not demonstrated on any family tested.

Acceptance is verticals.check only. The verifier reads governed tool evidence (visible test / published exploits) through its own gate; it never reads another agent's conclusion as proof and never sees hidden tests. Each agent has its own identity and gate (E88 grant + FencedResource per tool call).

Limits: 20 held-out tasks (bug_fix 10, remediation 10); agents are one model with different instructions, not different models; agents run sequentially in one process (no concurrency claim).
