Measures
- Whether each action class has an identifiable authorizer, evidence basis, and end condition.
- Exposure from dependency depth, substitution cost, and correction latency.
- Reversibility at the technical, operational, and institutional levels.
Delegation audit
Walks a team through one workflow against six questions and returns an exposure score, the permissions nobody can justify, and whether its decisions can be undone. Start here when the system keeps no decision records yet; use the record conformance checker once it does.
Estimated time: 20–30 minutes. When to use it · How it is scored
Audit one workflow
Name a workflow, answer the six sections in plain language, and copy or export the readout.
Readout: Exposure score with its three factors, a rating for each of the six questions, the permissions nobody can justify, and the reversibility verdict. View sample output
The workflow you are auditing
Audit one workflow at a time. Name the piece of work the system takes part in, not the system.
1. Capability
What can this system do in this workflow?
List each kind of action the system takes in this workflow, one per card.
Action 1
Action 2
2. Evidence
What has to be true for that permission to make sense, and when was it last checked?
3. Dependency
What now depends on this system, and how hard would it be to stop using it?
Dependent 1
Dependent 2
Exposure score = dependency depth (count of high and critical dependents) × substitution cost (staff-weeks) × correction latency (hours). The product is read as workflow staff-week hours.
4. Standing: who can challenge it
Who bears the errors, and can they make the system answer?
5. Correction
Can you halt it, and can the institution carry on if you do?
Readout: Refund decisions in the support queue
576workflow staff-week hours. Material.
2
Dependency depth: high and critical dependents
6
Substitution cost in staff-weeks
48
Correction latency in hours
Withdrawal would be a project with an owner and a budget. Rehearse it before the number grows.
Exposure score = dependency depth (count of high and critical dependents) × substitution cost (staff-weeks) × correction latency (hours). The product is read as workflow staff-week hours.
Share readout stores your answers, including anything you typed, on ethotechnics.org. Anyone with the link can open them for a year, and there is no way to delete them before then. Copy readout and Export JSON keep everything on this device.
Rating for each question
Capability
partial · 40/100- 2 action class(es) named for this workflow.
- There is no list of what the system could do separate from what it may do, so capability growth is invisible.
Authority
partial · 40/100- 1 of 2 action class(es) have an identifiable authorizer.
- 1 action class(es) have no end date and no condition that would end them.
- 1 action class(es) continue unless somebody intervenes, which is renewal by default.
Evidence
weak · 30/100- 1 of 2 action class(es) record what has to be true for the permission to make sense.
- 2 action class(es) have not had that checked in the last 12 months.
- The policy this system applies has no stated review trigger, so nothing reopens it.
- The policy has no expiry, so it stays in force until somebody remembers it.
Dependency
weak · 35/100- Exposure score 576 from depth 2 × 6 staff-weeks × 48 hours.
- The alternative was last run more than 12 months ago. Treat it as untested.
Standing
weak · 25/100- Errors reach the system only when a staff member passes them on.
- No named party has to answer.
- There is no deadline by which an answer is owed.
- A successful challenge fixes the individual case and leaves the system as it was.
Correction
partial · 50/100- A halt exists but has not been tested on this version.
- Halting it would degrade the service while it lasted.
- People who know the alternative are still here but have not used it lately.
Permissions nobody can justify
Actions nobody authorized, with no evidence behind them, or with nobody named as the people they decide about.
- Close a dispute without a human reading it: No authorizer is named. No evidence basis is recorded.
Reversibility verdict
- Technical: Not evidenced
A halt exists, but no test proves it works on this version, so it is recorded as not evidenced.
- Operational: Not evidenced
The alternative exists on paper but has not been exercised recently enough to count as evidence.
- Institutional: Not evidenced (weakest level)
Withdrawal would degrade the institution or exposure is heavy, so survival after withdrawal is not evidenced.
Institutional reversibility is not evidenced. Treat the ladder as no stronger than this level until a rehearsal says otherwise.
Findings
No capability list separate from the permission list
Without a list of what the system could do, an increase in capability arrives without a decision attached to it.
1 action class(es) with no authorizer or no evidence basis
A permission nobody can justify is something to investigate. Either the record is missing or the authority is.
2 action class(es) renew by default
Authority that continues unless somebody intervenes is renewed by silence. Record what the next renewal has to show, before the review period starts.
The policy behind this workflow is missing a trigger or an expiry
A policy without a review trigger and an expiry cannot send a permission back for review, so evidence and authority drift apart quietly.
The basis for the permission has not been rechecked recently
The absence of an observed failure is not evidence that the permission is still justified. Name what was examined and who looked.
Exposure score is 576 (material)
Record the three factors in a dependency record and track the number over time. A rising score with an unchanged safety case is itself a finding.
Withdrawal has not been rehearsed in the last 12 months
Substitution cost that has not been tested is usually wrong by a wide margin, and always in the same direction.
The people who bear the errors cannot reach a named responder
If the people who bear the errors cannot report them to someone who must answer, the system never hears about them. Name who may challenge, who answers, and by when.
A challenge cannot change how the system works
A challenge that only fixes one case leaves the system's permissions as they were. Define what a successful challenge can change.
Institutional reversibility is the weakest level
Withdrawal would degrade the institution or exposure is heavy, so survival after withdrawal is not evidenced.
The expertise to run the alternative is not being practiced
List the capacities kept so the workflow can still be run without this system, and give each one an owner.
Correction capacity rests on a halt the institution may not be able to afford
A correction the institution cannot afford to make is not counted as capacity. Show from the dependency record that the institution can actually afford to make it.
When to use the delegation audit
Teams who need to know whether the system's permission to decide is still justified, not whether the model is accurate.
- One named workflow, not a system or a product.
- The list of actions the system takes in it.
- Who authorized each one, if anyone can be named.
- A rough count of staff-weeks to run the workflow without it.
Method, transparency, and replicability
What the tool records, what it will not tell you, and how the exposure score and the reversibility verdict are computed.
Show inputs, procedure, scoring, assumptions, and validation
Inputs
- The name of one workflow the system takes part in.
- Each action class with its authorizer, evidence basis, affected people, and end condition.
- Dependents with criticality, substitution cost in staff-weeks, and correction latency in hours.
- Standing and correction answers: who bears errors, who answers, what can be halted.
Procedure
- Answer the six questions in order, in plain language.
- Read the exposure score with its three factors shown separately.
- Review the permissions nobody can justify, then the three reversibility levels, weakest first.
- Copy the readout or export the JSON snapshot for the record.
Outputs
- Exposure score with its three inputs. The bands (under 100 contained, under 1,000 material, 1,000 and over heavy) are this tool's convention, not a measured threshold; compare a workflow with its own earlier runs. Because the inputs are self-reported, the score does not meet the glossary's definition of an exposure score, which requires figures taken from records.
- A rating for each of the six questions, with the reasons behind it.
- A list of permissions nobody can justify, and whether decisions can be undone at three levels: technical, operational, and institutional.
- Findings linked to the STD-08 and STD-06 clause, the mechanism, and the eval suite.
Does not measure
- It records what the team in the room believes. It does not verify any of it.
- It is not an audit. Nothing here is evidence, and no finding is a compliance verdict.
- A permission nobody can justify is a finding to investigate, not a proven violation.
Assumptions
- The scope is one named workflow, not a whole system or product.
- Substitution cost and correction latency are stated in staff-weeks and hours.
- The people answering can name who authorized each action class, or admit that nobody can.
Instrument prompts
- What can this system do in this workflow, and is that list separate from what it may do?
- For each action class: who authorized it, on what evidence, for whom, and until when?
- Does the policy it applies have a review trigger and an expiry, and when was it last reviewed?
- What depends on this system, how long would the workflow take without it, and when was the alternative last run?
- Who bears the errors, can they raise one, who must answer, and by when?
- Can you halt it, can the institution keep functioning if you do, and does anyone still hold the expertise?
Rubric
- Grounded: 70 or above out of 100 on that question.
- Partial: 40 to 69.
- Weak: below 40.
- Reversibility at each level is evidenced, not evidenced, or not feasible. An unevidenced level is never recorded as feasible.
Scoring logic
- Exposure score = dependency depth (count of high and critical dependents) x substitution cost (staff-weeks) x correction latency (hours).
- Authority = share of action classes with an authorizer (60) plus a weighted end-condition share (40).
- Evidence = share with an evidence basis (50) plus recency weight (30) plus policy trigger (10) plus policy expiry (10).
- The weakest reversibility level sets the verdict, and the institutional level breaks ties.
Collection cost
Twenty to thirty minutes of one person's time, declared before they begin. If it runs past thirty minutes, stop and record the workflow as not established; the tool still produces a score, and that score should not be used.
- Respondent-supplied. The cost is the attention of whoever holds the answers, which is usually also whoever is audited.
- Whether each action was authorized at the time, and under which grant, is in the authority grant register.
- Whether records exist for the decisions is in the record stream.
- How hard withdrawal is comes from the dependency record.
- An audit nobody runs records no exposure. An institution that cannot name its authorized actions has not passed here; it has not answered.
Validation notes
Built against the STD-08 clauses and the STD-06 Article V dependency record so every finding maps to a clause that already exists.
Answers are self-reported, so two teams describing the same workflow can score differently. Re-run it with the people who hold the answers rather than the people who own the system.
- Scoping the audit to a whole product instead of one workflow.
- Estimating substitution cost without having run the alternative.
- Recording an untested halt as working reversibility.
- Treating a permission nobody can justify as a violation instead of an open question.
- Treating human review as a control when the reviewer cannot pause or change an automated action without penalty.
Replicability
- Name one workflow and gather the people who know how it runs.
- Answer the six sections in order without skipping the blanks.
- Export the JSON snapshot and keep it with the safety case.
- Re-run after the next expansion decision and compare exposure scores.
Example outputs
- Exposure score with dependency depth, substitution cost, and correction latency stated.
- List of permissions nobody can justify, with the reason for each.
- Reversibility verdict naming the weakest of the three levels.
Cite this page · v1.6.0 · updated Oct 2, 2026 · CC BY-SA 4.0
APA
Ethotechnics Institute Diagnostics Lab. (2026). Delegation audit. Ethotechnics Institute. https://ethotechnics.org/diagnostics/delegation-audit
MLA
Ethotechnics Institute Diagnostics Lab. "Delegation audit." Ethotechnics Institute, 2026, https://ethotechnics.org/diagnostics/delegation-audit.
Chicago
Ethotechnics Institute Diagnostics Lab. "Delegation audit." Ethotechnics Institute. Oct 2, 2026. https://ethotechnics.org/diagnostics/delegation-audit.
BibTeX
@misc{diagnostic_delegation-audit,
title={Delegation audit},
author={Ethotechnics Institute Diagnostics Lab},
year={2026},
howpublished={Ethotechnics Institute},
url={https://ethotechnics.org/diagnostics/delegation-audit},
version={v1.6.0}
}
RIS
TY - WEB TI - Delegation audit AU - Ethotechnics Institute Diagnostics Lab PY - 2026 UR - https://ethotechnics.org/diagnostics/delegation-audit ER -