VANGUARD SIGNAL 006 — THE BROKEN LOOP

Diagnostic / Field Artifact

Find Where the Verbs Went

A Broken Loop Diagnostic

A diagnostic for testing whether human oversight, automation, and contestability still preserve meaningful control.

A human in the loop is not the same as human control. This diagnostic helps locate the control verbs inside an AI‑assisted workflow: stop, reverse, amend, compensate, refuse, escalate, verify, and own.

Introduction

A loop is not proof of control. A workflow may include a human reviewer, an AI system, an approval box, a policy note, a dashboard, and an audit trail. It may still be broken.

The question is not only whether a human appears in the loop. The question is whether the functions of control still exist.

Stop: Can someone pause the system?
Reverse: Can someone undo an outcome?
Amend: Can someone correct a record?
Compensate: Can someone remedy harm?
Refuse: Can someone decline the action?
Escalate: Can someone move uncertainty to authority?
Verify: Can someone check the result?
Own: Can someone carry explanation and repair?

This diagnostic helps locate those verbs before the workflow produces damage, confusion, or responsibility without control.

One‑Page Diagnostic Card

Use this quick card before deploying, expanding, or defending an AI‑assisted workflow.

Control Verb	Core Question	Located?
Stop	Who can pause or halt the workflow before consequence?	☐ Yes ☐ No ☐ Unclear
Reverse	Who can undo the outcome after action?	☐ Yes ☐ No ☐ Unclear
Amend	Who can correct records, classifications, or outputs?	☐ Yes ☐ No ☐ Unclear
Compensate	Who can remedy material or procedural harm where applicable?	☐ Yes ☐ No ☐ Unclear
Refuse	Who can reject the system’s recommendation or action?	☐ Yes ☐ No ☐ Unclear
Escalate	Who can move uncertainty to someone with authority?	☐ Yes ☐ No ☐ Unclear
Verify	Who can check the result against independent evidence?	☐ Yes ☐ No ☐ Unclear
Own	Who carries explanation, repair, and recurrence prevention?	☐ Yes ☐ No ☐ Unclear

Quick result:

Mostly Yes: Control functions are visible. Continue with evidence check.
Mixed / Unclear: The loop may be weak. Run the full worksheet.
Mostly No: The loop is likely broken. Do not rely on “human oversight” or “automation efficiency” as sufficient justification.

1 — Workflow Identification

Field	Response
Workflow name
Workflow owner
System / tool involved
Human role involved
Automated role involved
Affected party
Risk tier	☐ Low ☐ Medium ☐ High
Consequence type	☐ Draft ☐ Recommendation ☐ Routing ☐ Record change ☐ Communication ☐ Denial ☐ Enforcement ☐ Financial ☐ Rights‑affecting ☐ Other
Date reviewed

2 — Control Verb Location Table

Verb	What it means	Status
Stop	Pause or halt action before consequence	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Reverse	Undo the action or outcome	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Amend	Correct a record, classification, output, or decision	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Compensate	Remedy material, procedural, financial, access, reputational, or rights‑affecting harm where applicable	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Refuse	Reject, narrow, or block the system recommendation/action	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Escalate	Move uncertainty or exception to authority	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Verify	Check result against independent criteria or evidence	☐ Present ☐ Weak ☐ Missing ☐ Unclear
Own	Carry explanation, repair, closure, and recurrence prevention	☐ Present ☐ Weak ☐ Missing ☐ Unclear

“Compensate” applies where a workflow can create material, procedural, financial, access, reputational, or rights‑affecting harm. In lower‑stakes contexts, this may collapse into repair, correction, or recourse.

3 — Break Mode Check

A — Inclusion Without Control

Question	Answer
What does the human actually see?
What does the human not see?
Does review happen before consequence?	☐ Yes ☐ No ☐ Unclear
Can the human refuse?	☐ Yes ☐ No ☐ Unclear
Can the human reverse or amend the outcome?	☐ Yes ☐ No ☐ Unclear
Can the human escalate uncertainty to authority?	☐ Yes ☐ No ☐ Unclear
Does the organization reward careful refusal or punish delay?
Is the human responsible for a system they cannot control?	☐ Yes ☐ No ☐ Unclear

Break mode result:

☐ No clear break
☐ Possible inclusion without control
☐ Likely inclusion without control
☐ Confirmed inclusion without control

B — Removal Without Rebuilt Contestability

Question	Answer
What human function was removed?
Why was it removed?
What improved after removal?
What new risk appeared?
Where is refusal rebuilt?
Where is verification rebuilt?
Where is appeal rebuilt?
Where is rollback rebuilt?
Where is repair ownership assigned?
Can an affected party challenge the outcome?	☐ Yes ☐ No ☐ Unclear

Break mode result:

☐ No clear break
☐ Possible removal without contestability
☐ Likely removal without contestability
☐ Confirmed removal without contestability

4 — Safe‑to‑Act Check

Gate	Question	Status
Action class	What kind of action can the system take?	☐ Known ☐ Unknown
Authority boundary	What is the system allowed to do?	☐ Defined ☐ Weak ☐ Missing
Prohibited actions	What is the system forbidden to do?	☐ Defined ☐ Weak ☐ Missing
Blast radius	Who or what can be affected?	☐ Known ☐ Unknown
Logging	Is action logged in a useful way?	☐ Yes ☐ No ☐ Unclear
Validation	Is there an independent check?	☐ Yes ☐ No ☐ Unclear
Escalation	What happens under uncertainty?	☐ Defined ☐ Weak ☐ Missing
Rollback	Can the action be undone?	☐ Yes ☐ No ☐ Unclear
Contestability	Can affected parties challenge outcomes?	☐ Yes ☐ No ☐ Unclear
Owner	Who owns consequence and recurrence prevention?	☐ Named ☐ Unclear ☐ Missing

Safe‑to‑Act Result:

☐ Safe to act within defined bounds
☐ Safe only with added review / limits
☐ Not safe to act yet
☐ Action class too unclear to authorize

5 — Safe‑to‑Repair Check

Repair Function	Question	Status
Explain	Can the organization explain what happened?	☐ Yes ☐ No ☐ Unclear
Reverse	Can it undo the outcome?	☐ Yes ☐ No ☐ Unclear
Amend	Can it correct records or classifications?	☐ Yes ☐ No ☐ Unclear
Compensate	Can it remedy harm where applicable?	☐ Yes ☐ No ☐ Not applicable ☐ Unclear
Restore access	Can it restore service, status, eligibility, or opportunity?	☐ Yes ☐ No ☐ Not applicable ☐ Unclear
Notify	Can it notify affected parties clearly?	☐ Yes ☐ No ☐ Unclear
Escalate	Can it move unresolved cases to authority?	☐ Yes ☐ No ☐ Unclear
Prevent recurrence	Can it change the system to prevent repeat failure?	☐ Yes ☐ No ☐ Unclear
Own closure	Is there a named owner for resolution?	☐ Yes ☐ No ☐ Unclear

Safe‑to‑Repair Result:

☐ Repair capacity matches system action
☐ Repair capacity exists but is weak
☐ Repair capacity is missing
☐ System can act beyond repair capacity

Field rule: No agent should act beyond the institution’s capacity to repair.

6 — Affected‑Party Check

Use this section when a workflow affects a customer, user, applicant, employee, student, patient, vendor, creator, citizen, or other outside party.

Contestability Function	Question	Status
Notice	Does the affected party know automation shaped the outcome?	☐ Yes ☐ No ☐ Unclear
Explanation	Can they understand the reason?	☐ Yes ☐ No ☐ Unclear
Evidence	Can they see or request relevant evidence?	☐ Yes ☐ No ☐ Unclear
Challenge	Can they contest the outcome?	☐ Yes ☐ No ☐ Unclear
Authority	Can they reach someone who can change the result?	☐ Yes ☐ No ☐ Unclear
Suspension	Can harm pause during review where appropriate?	☐ Yes ☐ No ☐ Not applicable ☐ Unclear
Reversal	Can the outcome be undone?	☐ Yes ☐ No ☐ Unclear
Amendment	Can the record be corrected?	☐ Yes ☐ No ☐ Unclear
Remedy	Can harm be remedied where applicable?	☐ Yes ☐ No ☐ Not applicable ☐ Unclear
Closure	Does the affected party receive resolution?	☐ Yes ☐ No ☐ Unclear

Affected‑Party Result:

☐ Contestability is functional
☐ Contestability exists but is weak
☐ Contestability is mostly symbolic
☐ No meaningful contestability

7 — Diagnostic Result Labels

Use these labels to classify the workflow.

Green — Control Functions Located
The major control verbs are assigned, evidenced, and usable.
Meaning: The workflow may proceed within defined bounds.
Next action: Schedule periodic review and test repair capacity.
Yellow — Control Functions Weak or Unclear
Some verbs are present, but authority, evidence, timing, or ownership is weak.
Meaning: The workflow is not necessarily broken, but it should not be described as fully controlled.
Next action: Strengthen refusal, verification, escalation, and repair before expanding automation.
Orange — Broken Loop Risk
A human is present without meaningful control, or automation has removed a function without rebuilding contestability.
Meaning: The workflow may produce responsibility without control or action without recourse.
Next action: Pause expansion, assign missing verbs, and run Safe‑to‑Act / Safe‑to‑Repair review.
Red — Broken Loop Confirmed
Critical verbs are missing, unclear, or unreachable.
Meaning: Do not rely on the workflow’s current oversight or automation structure.
Next action: Halt high‑consequence action until stop, refusal, verification, rollback, repair, and ownership are rebuilt.

8 — Interpretation Guide

If “stop” is missing: The system can move faster than authority. Repair: create pause, kill‑switch, or pre‑action gate.
If “reverse” is missing: The workflow may produce irreversible consequence. Repair: define rollback before deployment.
If “amend” is missing: Errors may persist downstream. Repair: assign correction authority and evidence requirements.
If “compensate” is missing: Harm may occur without remedy. Repair: define remedy path for material or rights‑affecting consequences.
If “refuse” is missing: Review is likely decorative. Repair: give reviewers real rejection authority and protect its use.
If “escalate” is missing: Uncertainty becomes trapped. Repair: define exception thresholds and authority path.
If “verify” is missing: The workflow trusts without evidence. Repair: add independent checks, tests, source access, or sampling.
If “own” is missing: Responsibility diffuses after failure. Repair: assign named ownership for explanation, repair, closure, and recurrence prevention.

9 — Next Actions

If the workflow is Green: document the role map; schedule review; test rollback periodically; keep logs and ownership current.
If the workflow is Yellow: assign unclear verbs; strengthen refusal and escalation; add evidence requirements; clarify repair owner.
If the workflow is Orange: pause further automation expansion; run the Role Division Matrix; run the Oversight Quality Audit; run the Safe‑to‑Act Gate; create or update the Contestability Register.
If the workflow is Red: stop high‑consequence deployment; identify the missing verbs; rebuild refusal, rollback, repair, and ownership; do not describe the system as controlled until the control functions are restored.

Artifact Card

Find Where the Verbs Went

A Broken Loop Diagnostic

A practical worksheet for testing whether an AI‑assisted workflow preserves meaningful control. Use it to locate who can stop, reverse, amend, compensate, refuse, escalate, verify, and own the consequences of automated or semi‑automated work.

Best for: AI workflow reviews, agent deployment, human‑in‑the‑loop audits, automation removal decisions, contestability checks.

Core rule: Do not ask whether the human is in the loop. Ask where the verbs went.

Boundary: This diagnostic does not prove motive. It tests control.

Source / Claim Note

This artifact is a DFEI diagnostic tool created for VANGUARD SIGNAL 006 — THE BROKEN LOOP. It synthesizes themes from human oversight, automation governance, auditability, contestability, agentic workflow design, and repair‑oriented accountability. It is intended as an operational diagnostic, not a legal determination or compliance certification.

Use it to identify missing control functions before relying on claims of human oversight, full automation, auditability, or accountability.