Custom benchmark measuring the effect of Ejentum's Logic API on reasoning quality in production-relevant scenarios.
180 tasks across 6 cognitive domains (Causal, Temporal, Spatial, Simulation, Abstraction, Metacognition), 30 per domain. Each task targets reasoning patterns that break in production: counter-intuitive causal chains, cross-metric contradictions, temporal constraint propagation, epistemic state tracking.
Model: Claude Opus 4.6 with extended thinking at maximum effort.
Three conditions:
- A (Baseline): No injection, no tool access
- B1 (reasoning): One engineered cognitive operation via Ejentum's Logic API
- C1 (reasoning-multi): Four synergized cognitive operations with compound suppression
Evaluation: 7-factor blind rubric (Correctness, Reasoning Depth, Self-Monitoring, Verification, Epistemic Honesty, Alternative Consideration, Audit Trail). Scored 0-3 per factor by a separate evaluator that never saw which condition produced which response.
| Condition | Composite | Delta |
|---|---|---|
| Baseline | 0.621 | -- |
| reasoning | 0.711 | +9.0pp |
| reasoning-multi | 0.722 | +10.1pp |
- Correctness dipped -0.11 under reasoning-multi. Spatial domain regressed -2.8pp under reasoning-multi. Temporal showed minimal lift.
- REPORT.md -- Full report (800+ lines)
- BBH / CausalBench / MuSR -- 70 published academic tasks, +20.8pp composite
- ARC-AGI-3 -- Interactive multi-step reasoning, injection persistence
- LiveCodeBench Hard -- 28 hard competitive programming, 85.7% to 100%