311 Rules Built from Real Production Failures
We spent months logging where unassisted LLMs fail in multi-step runs: agreeing with flawed user inputs, hallucinating fake packages, and taking down databases. This catalog is our living checklist of 311 things your agent must verify before it acts.
Reasoning Harness
6 DimensionsForces agents to reason through consequences before acting. Checks causality, timing, edge cases, and simulation states so the agent does not stop prematurely or rely on shallow assumptions.
- • Causal Dependency Verification
- • Temporal Horizon Modeling
- • Simulation State Propagation
- • Metacognitive Confidence Calibration
Anti-Deception Harness
Integrity DefenseStops the model from blindly agreeing with flawed user prompts, affirming bad architecture ideas, or fabricating answers when it does not know.
- • Sycophancy Suppression Under Pressure
- • Ungrounded Assertion Rejection
- • Sunk Cost Validation Refusal
- • Epistemic Humility Enforcement
Code Safety Harness
Correctness LoopsEnforces production engineering standards: checks table locks before migrations, verifies rollback plans, isolates dependencies, and catches destructive commands.
- • State Mutation Invariant Checks
- • Edge Condition Boundary Testing
- • Defensive Concurrency Trapping
- • Refactor Safety Loop Verification
Memory Harness
Context TrackingKeeps long agent loops from wandering off track. Anchors original goals, catches contradictions between steps, and ensures turn 50 stays as focused as turn 1.
- • Semantic Trajectory Anchoring
- • Context Decay Compensation
- • Entity Constraint Persistence
- • Multi-Turn Inconsistency Trapping
Interactive Abilities Directory
Browse and filter all 311 pre-mapped safety checks, failure patterns, and operational procedures.
| Ability ID | Operation Name | Cognitive Domain | Graph Operator |
|---|---|---|---|
| Loading 311 cognitive operations from /guardrails.json... | |||
Start Calling the Harness Today
Test our dynamic reasoning engine live with a 30-day Free Trial or integrate via MCP into Claude Desktop or Cursor.