Deterministic Horizon · Aix-Route
Know when your AI agent should stop thinking.
After ~20 deterministic reasoning steps, even frontier LLMs fall below 50% accuracy — and they fail confidently, not gracefully. The Depth Covenant cryptographically binds your agent to a measured reasoning limit, and delegates to a tool before it crosses the horizon.
- 0
- Max deterministic horizon (steps)
- 0
- Decoherence model R²
- 0
- Cost efficiency (tool vs CoT)
Test your task
Describe a task. We estimate its reasoning depth, compare it to your model's calibrated horizon d*, and recommend native reasoning or tool delegation. Runs entirely in your browser — nothing is sent anywhere.
Calibrated model horizons
Each model's Deterministic Horizon (d*) and decoherence parameters, measured offline with Aix-Route at temperature 0. These are the exact numbers a sealed Depth Covenant is bound to. Click a column to sort.
| Model | d* | ε₀ | L_eff | Provider | Calibrated |
|---|
Served-infra numbers (Groq, Ollama) are measured on real endpoints and run lower than paper-canonical values — the phenomenon holds; the horizon just moves. Recalibrate per deployment.
How it works
Calibrate
Measure your model's accuracy across reasoning depths on real tasks. Aix-Route fits (ε₀, d*) to a super-exponential decay curve.
Bind
Seal d* and its derivation hash into the agent's covenant manifest — Ed25519-signed, re-derivable offline, with no Python at runtime.
Enforce
At runtime, the depth estimator checks each task against the horizon and delegates to a tool before the model crosses d*.
Audit
Every horizon breach and delegation is logged as a signed event to the Grand Sijil — public, append-only, court-exportable.
Seal a Depth Covenant for your agent
Cryptographically bind your agent to a measurable reasoning limit. Calibrate once, enforce forever, prove it to anyone.
Seal a Depth Covenant →