The Difference Between "What Would Change the Answer" and "What You Can Actually Do"

Most AI explanations tell you a lie of omission. They say "if your income were $20K higher, you'd get the loan" — technically true, practically useless. The frontier of explainable AI isn't finding smaller changes; it's finding changes a human can actually make. PACE draws that line explicitly by pairing a neural predictor with a symbolic rulebook of what's feasible.

The Core Problem: Plausibility ≠ Feasibility

A counterfactual explanation identifies the minimal input change that flips a model's decision. The math is easy. The realism is not.

Standard methods optimize for distance — the smallest edit that changes the outcome. But smallest isn't achievable. They'll suggest lowering your age, reversing your gender, or acquiring a PhD by Tuesday. The recommendation is mathematically valid and humanly absurd.

The missing ingredient is domain knowledge: which attributes are immutable, which changes are directional (education can rise, rarely fall), and which combinations are coherent.

PACE's Framework: Split the Job in Two

PACE's insight is architectural separation of concerns:

• Neural layer — a predictive model does what neural nets do best: learn patterns and classify.
• Symbolic layer — Answer Set Programming (ASP) encodes human-readable rules that constrain the search to feasible interventions only.

The neural net answers "what does the model predict?" The symbolic layer answers "what am I allowed to change, and how?" Neither could do the other's job well alone.

Mental model — The Two-Key Lock: A good counterfactual needs both keys turned at once. Key one (neural) confirms the change flips the decision. Key two (symbolic) confirms the change is allowable in the real world. One key opens nothing.

Why Neuro-Symbolic, and Why Now

Pure deep learning is fluent but ungovernable — it can't natively represent "you cannot decrease your years of education." Pure symbolic AI is governable but brittle — it can't learn from messy data.

Neuro-symbolic AI fuses them: data-driven perception, rule-driven constraint. PACE is model-agnostic, so the symbolic layer can wrap around any classifier. The rules live outside the network, editable by domain experts who never touch the model weights.

In the Adult Income case study, ASP rules permit changes to education, occupation, and working hours while freezing immutable attributes like age and race. The result: recommendations that are interpretable, actionable, and consistent with how reality works.

The Actionable Takeaway (Beyond ML)

This is a design principle for any recommendation system — human or machine.

The Feasibility Filter. Before offering advice, run it through three gates:

1. Immutable check — Does this require changing something that cannot change? Discard it.
2. Directional check — Does the change move in a plausible direction and magnitude? A 40-hour work week can become 50; it rarely becomes negative.
3. Coherence check — Do the suggested changes hold together as a realistic path, not a set of independent miracles?

Advice that skips these gates is optimization theater. It looks rigorous and helps no one.

The Deeper Lesson

The best explanations aren't the most accurate — they're the most usable. A model that tells you the truth about what would work, but not about what you can do, has answered the wrong question.

Separate knowing from acting. Let your predictive engine model reality. Let an explicit, editable rule layer govern what counts as a legitimate move. When the two layers agree, you get explanations people can trust — and act on.

---

Sources & Further Reading

• PACE: A Neuro-Symbolic Framework for Plausible and Actionable Counterfactual Explanations — https://arxiv.org/abs/2607.01306