FIRME AI CORPORATION · branch evolution-79 · E78 baseline e2851428d2cd7ed9ae46f165e5d9b64c72632cf4
Standalone zero-CAIN-import verifier: 22/22 · verified=True
Bundle SHA-256: 44c49ea809a537cda566d3560a86e657adea0a3bda8b78004a55467ad5561539
Merkle root: 9cc3a2c6d351112ebc3fc7f5883cdf5ac297a487039075f93552f9576e89143a
Grade: B- (7.04/10) · production-ready: NO · deployment: NO
Download evidence ZIP Download verifier CAIN Studio proof bundle
| Arm | Correct / 100 | Accuracy |
|---|---|---|
| Deterministic first-integer baseline | 0 | 0% |
| Single direct model | 1 | 1% |
| Single model + calculator | 11 | 11% |
| Two specialists + consensus + calculator | 17 | 17% |
| Four specialists + consensus + calculator | 29 | 29% |
| Eight specialists + consensus + calculator (25 tasks) | 7 | 28% |
| Held-out replication: four specialists | 8/20 | 40% |
| Held-out replication: single direct | 0/20 | 0% |
This is a synthetic multi-domain arithmetic benchmark. Gains include deterministic calculator use and sample diversity, not general intelligence. Eight agents did not beat four.
1186 local model calls · 115622 tokens · Qwen/Qwen2.5-0.5B-Instruct · cost $0.00 · LOCAL_MODEL_INFERENCE
15 signed model-backed specialists. Actual model activity is displayed in ClawX, not registered identities.
Model-scale probe: 1.5B model 5/10. External GSM8K probe: single 0%, four-sample 0% on 10 tasks.
LoCoMo answer accuracy (50 QA): exact 2%, token F1 0.077. MemoryAgentBench: BLOCKED (141 MB load exceeded host memory budget).
Agentic75: 75/75 deterministic engines executed; 0 falsely relabelled model-backed. Kill switch: 10/10. Injection cases: 5; tool executions from model text: 0.
NOT_ATTEMPTED.README · signed evidence.json · master summary · final report · reproduce