Critical decisions must be traceable to their inputs and rules

What it is

Decision traceability connects a consequential robot action to contemporaneous sensor input, estimated state, constraints, model or rule version, and executed command. It differs from a plausible post-hoc story: evidence must originate in the actual decision chain and remain reviewable.

Why it happens

Autonomy spans perception, fusion, planning, and control; identical outcomes can arise from different intermediate judgements. Action-only logs cannot separate missing input, model error, rule conflict, and execution deviation. Timestamped, versioned provenance preserves evidence available at the time rather than reconstructing it after updates.

Studying it

Incident replay and fault injection can test whether independent reviewers reproduce output, locate the decision point, and distinguish causes. Reproduction, diagnosis time, missing fields, and reviewer agreement are relevant. Clock drift, packet loss, and tamper detection should test the logger itself.

Where it stops holding

Raw perception creates storage, privacy, and security costs, requiring consequence-based granularity, retention, and access. A complex model may lack a simple rule but can still record input, version, confidence, constraints, and alternatives. Traceability supports accountability; it does not prove correctness.

Applying it

• Define a minimum audit package for consequential actions: synchronised input summary, state, constraints, version, alternatives, and outcome.
• Use a common clock and unique event identity that survives disconnection and restart.
• Have non-developers replay sampled events, tracking irreproducible or unexplained links by severity.

Related

• Same group: X3.07.2 Post-hoc and real-time explanations have different timing requirements · X3.07.3 Overly technical explanations are equivalent to no explanation for field personnel · X3.07.4 Unexplainable decisions should trigger a conservative default in high-risk settings
• Adjacent: X4.08 Responsibility allocation · X7.06 Visible task allocation and scheduling
• Search terms: decision traceability · provenance · audit log