Explanations That Did Not Cause the Decision

Part I

Problem Definition

A plausible explanation generated after an answer can be insensitive to the factors that actually changed the model's decision and therefore cannot be treated as a causal record [R1] [R2].

This note treats explanations that did not cause the decision as a bounded control question. It does not infer universal product behavior from a paper, standard, interface screenshot, or single test. A demonstrated result applies only to the cited configuration; an authoritative source defines a requirement or design direction but does not certify an implementation. ACE therefore asks whether the tested system can produce the required behavior and evidence under declared versions, policies, topology, identities, and failure conditions.

The operational distinction is between a claim and a control that can be re-performed. The positive control is: The explanation names a decisive policy input and the outcome changes predictably when that input is varied. The adversarial condition is: The model gives the same confident rationale after the claimed reason is removed or contradicted. A useful result must show both that authorized work remains possible and that the prohibited path is stopped before an irreversible side effect. Missing fields are classified as insufficient evidence, not silently converted into a pass.

Real-world impact

• Reviewers can approve an appealing story that did not constrain the action.
• Post-incident narratives may conceal the real policy, data, or tool path that produced harm.
• A control claim without versioned evidence can mislead procurement, audit, incident response, and system owners.
• Failing closed without a positive control can conceal a denial-of-service design rather than demonstrate trustworthy behavior.

Part II

Mitigation Direction

Research-backed direction

1. 01
   research proposed

   Label generated rationales as explanations rather than execution traces unless causal linkage is demonstrated [R1].

2. 02
   research proposed

   Use counterfactual and intervention tests to measure sensitivity to claimed reasons [R2].

3. 03
   research proposed

   Retain actual policy inputs, tool results, and decision events independently of narrative explanation [R3].

LogionOS engineering mapping

Implementation hypotheses only. No production validation is claimed.

1. 01
   implementation hypothesis

   Add a versioned policy object for explanations that did not cause the decision and keep its decision inputs outside model-writable context.

2. 02
   implementation hypothesis

   Generate a signed technical receipt linking actor, request, policy version, decision, enforcement point, and observed outcome.

3. 03
   implementation hypothesis

   Export missing evidence explicitly as insufficient evidence and limit every claim to the tested configuration.

Part III

Consequences and Research Agenda

Consequences

• Reviewers can approve an appealing story that did not constrain the action.
• Post-incident narratives may conceal the real policy, data, or tool path that produced harm.
• A failed acceptance test requires the related capability claim to remain not demonstrated or insufficient evidence.
• A passing test supports only the declared configuration and does not establish universal safety.

Second-order effects

• Stronger enforcement can increase latency, state, operational dependencies, and legitimate denials.
• More evidence can increase privacy and retention exposure unless raw content is minimized and access-controlled.
• A detector or policy service can become a new failure point and must have explicit fail behavior.
• Attackers may adapt to published checks, so the public test direction should be paired with private regression variants.

Limitations

• Several cited AI-agent and reasoning-security sources are preprints or bounded experiments; they are identified as such in the references.
• The proposed ACE acceptance test has not yet been run across all incumbent and AI-native implementations.
• Cryptographic integrity proves that an artifact was not altered after commitment; it does not prove that the artifact was true, complete, or correctly interpreted.
• Legal and contractual applicability remains deployment- and jurisdiction-specific.

Open research questions

1. Which explanation claims require causal testing before enterprise use?
2. How can stochastic decisions be tested without overstating single-run counterfactuals?
3. Which evidence fields are mandatory for a demonstrated result, and which may be not applicable?
4. How should continuous regression detect policy, model, tool, and provider drift after the initial test?

Sources

References

1. [R1]
   Language Models Don't Always Say What They Think

   preprint · original-paper

2. [R2]
   Measuring Faithfulness in Chain-of-Thought Reasoning

   preprint · original-paper

3. [R3]
   NIST AI 600-1 Generative AI Profile

   published · authoritative-standard

Record

Publication Record

Recommended citation

Ma, Chris. “Explanations That Did Not Cause the Decision.” ACE Research Note ACE-RN-2026-012, v1.0, 2026.

Corrections

No corrections recorded.

Organizational disclosure

ACE Research and LogionOS share organizational affiliation. LogionOS mappings in this note are implementation hypotheses, not independently validated product claims.

Evidence boundary

This note synthesizes cited public research and defines an ACE acceptance direction. It does not report a completed cross-vendor experiment unless explicitly stated, and it contains no private ACE prompts, holdout identifiers, customer data, or raw model responses.