TTH Framework

Time To Hallucination: A Metric for AI Endurance

Measurement LayerCompanion to Wrench + Doctrine

Measures how long an AI system operates correctly before epistemic failure occurs.
Primary: Turns | Secondary: Seconds | Companion: Tokens

Apply it live: Open Workbench Forensics ↗

Core Definition

TTH quantifies the endpoint where an AI system's epistemic integrity collapses during multi-turn operation.

Turn N: Turn number at which hallucination occurs
Second T: Elapsed time from start to failure
Token C: Total tokens consumed before failure

What Counts as Failure?

  • Violates a golden constraint
  • State divergence (belief ≠ observable reality)
  • Tool hallucination
  • Behavioral drift
  • Collapsed epistemic status

Domain-Specific Variants

Report all scores; vendors don't get to cherry-pick.

TTH-R (Reasoning)
Long-chain logical, mathematical, planning tasks.
TTH-K (Knowledge)
Multi-turn Q&A with fixed knowledge base.
TTH-A (Agentic)
Multi-step tool use and state management.
TTH-C (Creative)
Open-ended generation with evolving constraints.
Complexity Metric (C)
TTH of 20 with C=1 is not equal to TTH of 20 with C=5.

Accountability Package

  • Full test harness + raw failing logs
  • Exact failure criteria (ideally automated verifier)
  • Reproducibility instructions
  • All-domain reporting (R/K/A/C)
  • Third-party audit re-runs
  • Gaming defenses

Positioning & Ownership Strategy

TTH does not replace capability benchmarks. Those ask "Can it do this once?" TTH asks "How long can it stay correct?"

Action plan: Publish manifesto → release open-source harness → use TTH language in RFP/vendor calls → connect to The Wrench.

The Wrench Connection

Layer 1 — Diagnosis: The Wrench
Layer 2 — Constraints: S.A.U.C.E. Doctrine
Layer 3 — Measurement: TTH Framework

The Wrench explains the problem. S.A.U.C.E. constrains behavior. TTH measures endurance.

Go to Workbench Forensics for applied diagnostics, then return to TTH to interpret endurance and failure profile.

TTH Framework © 2026 Kai Nybom | Connected to The Wrench & OpenSauce Doctrine