CT-SAFR: Safe and Interpretable Chain-of-Thought Reasoning for Autonomous Robots

A multi-layered verification framework for trustworthy AI-driven robotic decision making.

C. Temel, “CT-SAFR: Safe and Interpretable Chain-of-Thought Reasoning for Autonomous Robots: A Multi-Layered Verification Framework for Trustworthy AI-Driven Robotic Decision Making,” in 2026 IEEE Conference on Artificial Intelligence (CAI), IEEE, May 2026, pp. 598–603, doi: 10.1109/CAI68641.2026.11536646.

IEEE Xplore DOI Conference program

BibTeX

@inproceedings{Temel2026CTSAFR,
  title     = {CT-SAFR: Safe and Interpretable Chain-of-Thought Reasoning for
               Autonomous Robots: A Multi-Layered Verification Framework for
               Trustworthy AI-Driven Robotic Decision Making},
  author    = {Temel, Cagri},
  booktitle = {2026 IEEE Conference on Artificial Intelligence (CAI)},
  pages     = {598--603},
  year      = {2026},
  month     = may,
  publisher = {IEEE},
  doi       = {10.1109/CAI68641.2026.11536646}
}

The problem

Chain-of-Thought prompting lets a language model plan a robot task in explicit steps, and those steps look like an explanation of what the model is doing. Recent work shows they often are not. Reasoning models verbalize their actual decision process only 25 to 39% of the time, and faithfulness degrades by a further 44% on harder tasks, exactly the tasks where a robot most needs oversight. A robot that acts on an unfaithful reasoning chain can damage equipment, the environment, or a person, and current verification methods do not catch this failure mode.

The approach

CT-SAFR treats a reasoning chain as an artifact to be checked rather than an explanation to be trusted. It assumes no single technique can guarantee safety, so four independent layers run in sequence and any one of them can send the robot to a fallback mode: conservative operation, reduced autonomy, a safe-state transition, or human escalation.

Layer	What it checks	Measured cost
1. Structural verification	Logical coherence independent of meaning: circular dependencies, unsupported assertions, broken entailment, malformed reasoning	12 ms median, catches ~15% of problematic chains
2. Physical constraint validation	Joint limits, velocities, accelerations, collision boundaries, force thresholds, operational envelopes, checked outside the language model so a persuasive but wrong chain cannot cross it	45 ms at p95, 99.8% detection accuracy
3. Semantic consistency	Self-consistency decoding over k = 5 sampled reasoning paths; agreement above 80% executes, 50 to 80% triggers conservative behavior, below 50% escalates to a human	part of the sub-500 ms budget
4. Interpretability interface	Decision visualization, confidence reporting, and an audit log for the human operator	real-time monitoring and override

The design point worth noting is layer 2. It is deliberately independent of the reasoning process, which means physical safety does not depend on the reasoning chain being faithful at all. That is how the framework addresses the faithfulness problem without needing to solve it.

Results

Measure	Result
Unsafe reasoning detection (n = 500)	94.2% (95% CI 91.8–96.6%)
False positive rate	3.1%
Verification latency	under 500 ms
Reduction in unsafe reasoning outputs, warehouse robot case study	87% (p < 0.001)
Task completion rate, across 1,000 operation hours	+3.5%

An ablation study quantifies each layer separately and shows that every layer contributes safety benefits the others do not, which is the empirical argument for the defense-in-depth structure rather than a single stronger checker.

Related work of mine

CT-SAFR adapts verification techniques from CogniTest, my earlier multi-layered verification framework for LLM-assisted software testing, and applies them under the physical grounding and real-time constraints of robotics. The auditability side of the same problem is treated in TRACE.

← All publications TRACE →