Rule Ablation Tester for LLM compliance systems

I found no commercial product that treats the governing rule itself as the thing to perturb and fails a build when an LLM compliance verdict ignores the rule or is unstable across reruns. I also found no packaged sensitivity report mapped to the MAS AI risk guidelines for a bank's validation team to sign. The method now exists as public research, which confirms rule-blindness is a real failure, even in guards that cite the correct clause. Generic eval tools (Promptfoo, Giskard, LLMORPH) only change inputs, and model-risk platforms like ValidMind have generic stability tests but not this one. A new entrant can win by turning the research method into a turnkey harness for regtech vendors, with libraries of rule perturbations for KYC, marketing-compliance and complaints rules plus repeated runs to measure stability. The stronger moat is to become the standard independent sign-off report for Singapore banks. The MAS guidance I found asks for independent validation of high-risk uses, compensating tests for vendor AI that discloses too little, and documentation detailed enough for a reviewer to reproduce. I only saw secondary summaries, mostly of the consultation draft, so the mapping should be checked against the final MAS text. The main risk is that the test is easy to build and now public. ValidMind, Giskard or a Singapore assurance firm could ship it as a feature, so lasting value has to come from the MAS mapping, the domain rule library, acceptance by validators, and speed to market.

• What Do Compliance Detectors Read? (rule-blindness audit, arXiv 2608.16852)

  An August 2026 research paper that does almost exactly this test. It deletes, swaps and substitutes the governing rule while keeping the scenario fixed, and finds that guard models and activation probes are rule-blind: their verdicts do not change. The authors say they release the test protocol and a crossed-rule benchmark, but I could not find where these are hosted. It is academic, not a product, but it publicly proves the problem and makes the method easy to copy.

• Four-Axis Trustworthiness Benchmark for LLM-as-Judge in Principle-Based Regulation

  A research benchmark that tests LLM compliance judges under the UK FCA's Consumer Duty rules with adversarial input perturbations. It is close in spirit but changes the input rather than the rule, and it is not commercial.

• Promptfoo

  An open-source (MIT) tool for LLM evaluation and red-teaming that can gate builds in CI. It is generic and has no built-in rule-ablation tests or regulator-mapped reports, but a user could build rule ablation on top of it.

• Giskard

  An LLM testing tool that scans for robustness, bias and hallucination problems and produces HTML reports, and is pitched at regulated industries including banking. CI integration and compliance features are in the paid Hub tier. It does not test whether a verdict depends on the rule.

• LLMORPH

  An open-source research tool for metamorphic testing of LLMs: it changes inputs in controlled ways and outputs failing test pairs that a CI job could gate on. It covers generic NLP tasks, not compliance-rule ablation.

• ValidMind

  A model risk management platform for banks, built around the US Fed's SR 11-7 guidance. It has stability and sensitivity tests, LLM prompt validation and automated validation documentation. It is the incumbent best placed to add a rule-ablation test to its library.

• Project Moonshot (AI Verify Foundation / IMDA)

  Singapore's open-source LLM testing toolkit for benchmarks and red-teaming, focused on safety and security. Assurance firms such as Resaro work in the same ecosystem. I found no rule-sensitivity tests and no reports mapped to MAS rules.

Checked 2026-10-09