Kevin Astuhuaman

Opinionated interface patterns for consequential AI

Enterprise AI needs an interface for uncertainty.

One reference workflow packages evidence, confidence, authority, approval, observable traces, failure, recovery, and absence into a product system that operators can actually reason about.

Open the reference product Inspect seven contracts
01Provenance chainWhat evidence produced this output?
02Calibrated confidenceHow much uncertainty matters for this decision?
03Permission boundaryWhat can this person and agent do now?
04State-bound approvalWho owns the consequential decision?
05Observable run traceWhat happened without exposing private reasoning?
06Scoped failureWhat stopped and what remains valid?
07Honest empty stateWhat does the system know when evidence is absent?

Reference implementation · synthetic data

Account onboarding review

Ready for operator review

Northstar Labs

Enterprise account · SYN-440 · Observed 2026-07-12 09:40 PT

Evidence ready
Agent recommendation

Prepare standard onboarding

Required sources agree on the legal entity, plan, and onboarding owner.

Pattern 01

Provenance

4 required sources
  1. CRM accountSystem record · 2026-07-12 09:40 PT
    VerifiedEntity, owner, and region match
  2. Signed orderDocument extraction · 2026-07-12 09:28 PT
    VerifiedPlan and start date extracted
  3. Billing profileAPI record · 2026-07-12 09:37 PT
    VerifiedBilling entity confirmed
  4. Security reviewPolicy record · 2026-07-11 09:42 PT
    CurrentStandard controls accepted
Pattern 05

Observable trace

Events, not hidden reasoning
  1. System
    Collected four required sourcesNo writes performed
  2. Policy
    Applied onboarding policy v3.4Draft boundary selected
  3. Model
    Prepared recommendation R-2048Evidence references attached

Reusable product contracts

A kit for decisions, not decoration.

Each pattern names the user question, the minimum product contract, and the failure it is designed to prevent.

PatternProduct questionMinimum contractFailure prevented
01 · Provenance chainWhat evidence produced this output?

Source, observed time, transformation, and current availability

Untraceable recommendations and stale-source ambiguity

02 · Calibrated confidenceHow much uncertainty matters for this decision?

Calibrated band, evidence coverage, threshold, and intended action

False precision and confidence-as-permission

03 · Permission boundaryWhat can this person and agent do now?

Actor, allowed action, prohibited action, and scope

Capability being mistaken for authority

04 · State-bound approvalWho owns the consequential decision?

Owner, exact state binding, expiration, and invalidation rule

Stale approval surviving an input change

05 · Observable run traceWhat happened without exposing private reasoning?

Timestamp, actor, observable event, input or output reference

Opaque execution and chain-of-thought leakage

06 · Scoped failureWhat stopped and what remains valid?

Failed boundary, preserved work, retry scope, and owner

All-or-nothing recovery and duplicate side effects

07 · Honest empty stateWhat does the system know when evidence is absent?

Absence reason, impact on decision, and bounded next action

Fabricated certainty and silent no-data states

Operating principles

Trust comes from inspectable boundaries.

Evidence before explanation

Show what supports the output before asking users to trust the summary.

Authority before action

Make the allowed actor and exact approval boundary visible at the decision point.

Recovery before retry

Preserve completed work, isolate the failed boundary, and recompute stale decisions.

Absence before invention

An honest empty state is more useful than a fluent answer without evidence.