Built to Confess

Opening Scene

There are two ways a case ends up with a clear account of what happened. Some suspects are simply forthcoming from the start, offering a coherent account without needing to be pressed. Others require the full investigative toolkit — SHAP, LIME, sensitivity testing — to reconstruct after the fact. Article 3 introduced glass-box models as the first kind. This article steps back and names the deliberate choice: build for transparency from the start, or invest in reconstructing it afterward.

In Plain English

Ante-hoc (or intrinsic) interpretability comes from choosing a model architecture that’s transparent by design, as covered in Article 3. Post-hoc explanation applies techniques like SHAP, LIME, or PDPs to a model after it’s already trained, regardless of its internal complexity. Neither approach is universally better — they represent a genuine design tradeoff between accuracy, transparency, and the effort required to bridge the gap between them.

The Old Way

Before this distinction was formalized in machine learning, the same choice appeared in other domains:

Building a legal contract in plain, transparent language from the start, versus writing dense legalese and hiring a lawyer later to explain it.
Designing a piece of software with clear, self-documenting code, versus writing dense, optimized code and adding documentation afterward.
A company built from the start around transparent decision-making processes, versus one that later hires consultants to explain how decisions actually get made.

In each domain, building for transparency from the start and explaining after the fact were both viable but genuinely different strategies with different costs.

What’s Changing (and Why AI Is the Reason)

As post-hoc explanation techniques like SHAP have matured and become more reliable, the practical cost of choosing a complex, high-accuracy model and explaining it afterward has dropped significantly, making that path more viable than it once was.
At the same time, research into inherently interpretable model architectures — designed to approach black-box accuracy while staying naturally transparent — has narrowed the accuracy gap Article 3 described, making ante-hoc interpretability a more competitive choice too.
Regulatory and organizational pressure, covered in Article 19, increasingly favors documented evidence of interpretability by design in the highest-stakes domains, even when a purely accuracy-driven choice might favor a black-box model with strong post-hoc explanations.**

The Metaphor, Fully Extended

The Investigation	Ante-Hoc vs. Post-Hoc Concept
A suspect who’s forthcoming from the very start	A model that’s interpretable by design
A suspect requiring the full investigative toolkit afterward	A complex model requiring post-hoc explanation techniques
Choosing which kind of suspect to build a case strategy around	Choosing between glass-box design and black-box-plus-explanation
A case built entirely on reconstructed evidence, versus one built on direct testimony	Post-hoc explanation versus intrinsic interpretability

For Beginners: What to Actually Do

When starting a new modeling project, explicitly ask whether ante-hoc or post-hoc interpretability better fits the actual stakes involved.
Don’t assume post-hoc explanation is automatically “good enough” — revisit Article 11’s faithfulness concerns before relying on it for high-stakes decisions.
Learn at least one model type from each category well enough to compare their real tradeoffs directly.

For Practitioners and Leaders: The Deeper Layer

Make the ante-hoc-versus-post-hoc decision explicitly and early in a project, documenting the reasoning for later audits.
In the highest-stakes domains, favor ante-hoc interpretability where the accuracy cost is genuinely small, given growing regulatory expectations.
Track how the accuracy gap between interpretable and black-box models evolves for your specific use case — it’s a moving target as both fields keep improving.

Quick Recap

Ante-hoc interpretability builds transparency into a model’s design from the start; post-hoc explanation reconstructs it after training.
Both are viable, genuinely different strategies with real tradeoffs in accuracy, cost, and effort.
Post-hoc techniques have matured considerably, while ante-hoc methods have also narrowed the accuracy gap.
Regulatory pressure increasingly favors documented, by-design interpretability in the highest-stakes domains.

Where This Fits in the Series

Article 13 named the deliberate ante-hoc-versus-post-hoc choice. Article 14 looks at a specific, modern case of post-hoc-adjacent interpretability: attention weights inside deep neural networks.