HR Control Workbench — Hongmin Stewart

01 / The problem

Real payroll contains different kinds of work.

A simple automation model assumes:

Input → Calculation → Payment

Rules: What should happen under known conditions?
Systems: How should authorized transactions be processed?
Exceptions: Where do expected and actual results disagree?
Investigation: Why did they disagree?
Judgment: What does the evidence actually support?
Authority: Who is allowed to change the result?

Processing work and decision authority are not the same thing.

02 / Redesigning the work

Allocate work by capability—and by authority.

I separated the workflow according to what each participant is actually good at—and what authority it should hold.

RulesDeterministic control

Calculate
Compare
Validate required conditions
Apply thresholds
Flag defined exceptions

Same authorized inputs should produce the same result.

SystemsControlled execution

Store authorized data
Execute configured calculations
Maintain workflow state
Record transactions and changes
Enforce permissions

The system executes authority. It should not invent it.

AIInvestigation capacity

Compare records across sources
Surface anomalies
Identify patterns
Generate possible explanations
Expose contradictory evidence
Prepare information for review

AI can make investigation faster. Its explanation is a hypothesis—not authorization.

HumanJudgment & authority

Evaluate evidence
Resolve material ambiguity
Approve corrections
Interpret policy where necessary
Determine escalation
Authorize overrides
Own the consequence

Human attention should be concentrated where judgment actually matters.

The Human role should not be moving information between systems simply because the systems cannot communicate.

Human attention should be concentrated where judgment actually matters.

03 / The decision boundary

Detection ≠ explanation ≠ authorization

Detection: Something appears inconsistent.Rules, systems, or AI may identify it.
Explanation: Why might it have happened?AI may investigate and propose possibilities.
Authorization: What should actually change?Where employee or financial consequences are material, the appropriate Human authority decides.

A detected anomaly is not automatically an error.

A plausible explanation is not automatically true.

And neither grants authority to alter payroll.

04 / Normal work vs. exception work

Two paths. Different control needs.

Normal path

Authorized data
Deterministic validation
System processing
Reconciliation
Release

When evidence agrees and the transaction remains within defined authority, adding Human relay can create work without adding control.

Exception path

Exception detected
Evidence assembled
AI-assisted investigation where useful
Human judgment
Authorized action / escalation / no action
Evidence retained

AI belongs inside the investigation path. It does not hold release authority.

05 / Where the model can fail

Control the failure, not just the process.

Correct calculation / wrong configuration

A system faithfully executes an incorrect rule.

Control: Reconciliation and independent expected-result checks.

Plausible explanation / insufficient evidence

AI proposes a convincing explanation for a discrepancy.

Control: Separate hypothesis from verified evidence.

Correct detection / unauthorized correction

An anomaly is correctly identified, but an automated correction exceeds its authority.

Control: Material changes require defined approval authority.

Human present / no meaningful oversight

A Human clicks “approve” after AI has already framed the issue, selected the evidence, and recommended the action.

Control: The reviewer must have sufficient evidence, understandable reasoning, and real ability to reject or escalate.

Human-in-the-loop does not automatically mean meaningful Human oversight.

06 / Evidence gate

End in an evidence state—not an AI confidence score.

Exceptions should end in an evidence state—not simply an AI confidence score.

Pass: Evidence supports the proposed conclusion or action within defined authority.
Fail: Evidence contradicts the conclusion or a required control was not satisfied.
Unknown: Evidence is insufficient to justify a conclusion.

Sometimes UNKNOWN is the correct controlled result.

The appropriate response may be additional evidence, Human review, or escalation.

07 / What changed

Reduce Human relay. Preserve Human judgment.

A common Human-in-the-loop model:

Automation does the work → Human checks it

That can leave people performing routine review, transferring information, or rubber-stamping machine recommendations.

The lab produced a more deliberate allocation:

Rules: Handle deterministic logic.
Systems: Execute authorized transactions.
AI: Expand investigation capacity.
Evidence: Constrain conclusions.
Humans: Own material judgment, escalation, and authority.

Reduce unnecessary Human relay.Preserve meaningful Human judgment.

08 / Current status

Released Baseline v0.2.
GO · FROZEN.

The work progressed from a control model to a working private application, then through regression, adversarial bypass testing, and the server Release Gate.

21/21: Adversarial bypass tests passed
S1–S10: Server Release Gate passed
0: True authority bypasses
0: Release blockers

Golden Case
Regression
Adversarial Bypass Testing
Server Release Gate
Frozen Release Evidence

Released means the validated build was frozen as the baseline inside the owner’s private workspace. It does not mean public SaaS, commercial use, or employer deployment.

Material input changes invalidate prior confirmation. Re-analysis and re-confirmation are required before authority can be restored. Same-value rewrites act as no-ops. Under the tested contract, audit history is append-only and Human decision writes do not overwrite AI recommendation content.

Any later product change invalidates this release baseline and requires the Release Gate to be rerun.

AI can analyze and recommend. It cannot grant itself authority.

Known non-blocking post-release items

P1 Old analysis output does not yet carry per-item superseded or stale markers. Case-level re-analysis requirements, stage rollback, escalation invalidation, and server-side gates prevent stale analysis from retaining authority.

S5 Multiple Human decisions currently use append semantics. Later decisions create independent records; earlier decisions remain unchanged. Until the semantics change, the latest entry is treated as current.

Test harness note: The first RG-B confirmation failure resulted from sending saveanswer and analyzecase concurrently. The sequential rerun passed; this was a harness race, not a product defect.

Integrity boundary: Hongmin owns the framing, authority model, controls, governance standards, acceptance criteria, adversarial review standard, GO decision and final accountability. AI agents assisted with implementation, test execution, scenario expansion, debugging and documentation drafting.

09 / What this case demonstrates

Technology changes work. Decision rights must change deliberately.

This case started with payroll, but the underlying problem is broader:

When technology becomes more capable, how should work and decision rights change around it?

The answer is not to automate everything possible and add a Human approval button at the end.

Calculation from judgment
Detection from explanation
Explanation from authorization
Human presence from meaningful oversight
Technical capability from decision rights

Automate the deterministic.Accelerate the investigative.Preserve Human authority where consequences become material.

AI can change who—or what—does the work.

It does not eliminate the need to decide who owns the consequence.

Selected receipts