Reasoning Transparency

After edition 1 dropped, Richard Lynes left a comment that nailed the solution better than I did.

He called it a "replayable decision receipt".

Authority + inputs + predicates + reason-codes + stable hash + export command

That's it. That's the whole pattern.

Not "here's why the AI decided." Here's a receipt you can replay months later when someone asks you to defend the decision.

Let me show you what this looks like in practice.

340B pharmacy system from last week. The pharmacist makes an eligibility determination. Six months later HRSA auditor asks, "Why did you determine this patient was eligible?"

The pharmacist doesn't remember. They don't need to. They just export the decision receipt.

Not "the system says eligible." This is "here's the authority we applied, here's the input data we used, here's the predicate that held true, here's the reason code you can cite, here's where the data came from, here's our confidence in that data source."

Six months later, the pharmacist can explain exactly why this determination was made. Not from memory. From the receipt.

That's defensible.

Here's where teams mess this up: they think showing reasoning means showing all reasoning all the time.

Wrong.

Progressive disclosure. Three layers.

Layer 1: Decision + confidence + key factors (shown by default)

This is what users see immediately. Decision, confidence score that means something, 3-4 critical factors. Takes 15 seconds to scan. Gives enough to trust or question.

The physician sees suggestion, confidence, three key factors. Enough to make an initial judgment. If she agrees, it documents and moves on. If she questions it, it expands reasoning.

Layer 2: Detailed reasoning (click to expand)

This is where the full decision logic lives. Authority applied, data considered, differential diagnoses, clinical rules, and why alternatives scored lower.

Only loads when the user needs it. Keeps the default view clean.

The physician can now explain to the patient: "Your fever has been increasing progressively over three days, your white blood count is elevated showing infection, chest x-ray shows infiltrate in your right lung, we've ruled out covid and flu."

Fully defensible. Completely transparent.

Layer 3: Audit trail + export (always available)

One-click export of complete decision chain with timestamps, data sources, authority citations, and confidence scores.

For regulatory audits, medical review boards, compliance officers, and legal defense.

When the auditor asks six months later, "Why did you prescribe antibiotics?"

The physician: "Here's the complete decision trail with clinical evidence, differential diagnoses considered, decision rules applied, patient presentation timeline."

Instead of "I don't remember," defensible from day one.

Amber Asaro asked in the comments on edition 1: "Are you going to touch on overtrust?"

Yes, because here's the dangerous flip side.

Teams see progressive disclosure and think "show everything, let users decide."

No.

Showing too much creates false confidence.

I watched this kill a fraud detection AI last year.

The system showed 47 different factors in the decision. Transaction velocity, device fingerprint, location patterns, behavior scoring, network analysis, account age signals, and payment history. All visible by default.

The analyst looks at this and thinks, "Wow, the AI considered everything. Must be right."

Except the AI was wrong 23% of the time in that specific scenario (cross-border transactions from returning customers).

Showing all 47 factors created an illusion of thoroughness. The analysts trusted it blindly because "look how much it analyzed!"

That's overtrust. And it's deadly.

The fix: show confidence honestly + make uncertainty visible + design for skepticism.

Not "AI says block it" with 47 factors suggesting completeness. This is "AI flags it BUT has 23% error rate in this scenario, use your judgment."

Transparency includes showing where the AI is uncertain.

Progressive disclosure isn't just about showing reasoning. It's about showing uncertainty, showing where you should be skeptical, and showing when to override.

Working with an enterprise pricing approval system right now. Sales reps submit quotes, AI recommends approve/reject based on margin analysis, competitive pricing, and customer history.

First version: AI recommendation with a single confidence score.

Result: Reps either trusted it blindly (overtrust) or ignored it completely (no trust). 50/50 split. Useless.

Second version: Progressive disclosure.

The default view shows recommendations, confidence, and three key margin factors.

But here's the key addition: we show when the AI is uncertain and why.

Now the reps have: decision + reasoning + where AI is uncertain + when to escalate.

Adoption went from 50% to 84%. Not because AI got smarter. But because the interface got honest about uncertainty (and overtrust dropped to 12% because we explicitly flagged uncertainty).

Honesty about limitations builds more trust than pretending certainty.

Anthropic just published their agent skills framework (February 2026), and the core concept is literally progressive disclosure.

Agents don't load all capabilities upfront. They load metadata (what exists), then load full instructions only when needed.

Why? Because "the kitchen sink approach creates context rot. Irrelevant information makes agents dumber."

The same principle applies to human interfaces.

Show users metadata (decision + confidence + key factors). Load full reasoning only when they need it. Keep default view clean.

Academic research backs this up. A study published August 2025 in Healthcare AI: "Progressive Disclosure in Clinical Decision Support Systems."

Key finding: physicians using progressive disclosure interfaces had "better understanding and stronger feelings of trust" compared to full-disclosure-upfront interfaces.

They tested three versions:

• Black box (no explanation): low trust
• Full disclosure upfront (all reasoning visible): cognitive overload, decision paralysis
• Progressive disclosure (expandable layers): highest trust, fastest decisions

Middle ground wins. Every time.

Also worth noting: EU AI Act compliance deadline is August 2026. Transparency requirements are now law in Europe.

But here's what most teams miss: the law doesn't say "show all reasoning." It says "provide clear, comprehensible reasons for AI-generated outcomes."

Comprehensible ≠ comprehensive.

You can be fully compliant with progressive disclosure. Decision visible, reasoning available on demand, audit trail exportable.

Showing everything upfront doesn't make you more compliant. It makes you less usable.