Production AI readiness
Review production AI controls across purpose, data, evaluation, authority, deployment and operations. Build an evidence-based action plan, not a readiness score.
The decision
Can this AI workflow be operated safely when it is wrong?
Assumptions and limits
This does not certify AI safety, legal compliance, or model accuracy.
Describe your situation
Purpose and accountability
Is the workflow bounded by a measurable outcome and an explicit non-AI alternative?
Evidence: named users, task boundary, acceptance definition and current process baseline.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Purpose and accountability
Are release, risk acceptance and incident decisions assigned to accountable owners?
Evidence: named business and technical owners, escalation route and approval record.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Purpose and accountability
Have you assessed who could be harmed by an incorrect answer or action?
Evidence: affected groups, severity, plausible misuse and human escalation needs.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Data and access
Are data permissions and retrieval scope explicit?
Evidence: identity propagation, tenant boundaries, source permissions and denied-access tests.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Data and access
Are source quality, freshness and deletion rules verified?
Evidence: source owners, update lag, deleted records and citation checks.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Data and access
Are sensitive inputs, provider handling and retention boundaries reviewed?
Evidence: approved data classes, endpoint configuration, retention, residency and logging policy.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Evaluation and security
Do you have representative failure cases and a release gate?
Evidence: versioned tasks, expected outcomes, failure categories and explicit pass conditions.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Evaluation and security
Are unsupported claims, missing evidence and abstention tested?
Evidence: unanswerable questions, contradictory records and misleading citations.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Evaluation and security
Have malicious instructions in prompts, retrieved content and tool results been tested?
Evidence: adversarial cases across every untrusted input path, not only the chat prompt.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Evaluation and security
Are model outputs validated before use by another system?
Evidence: schema, type, range, escaping and business-rule checks outside the model.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Authority and human control
Can the system change records or trigger external actions?
Include emails, payments, deployments and background jobs. Read-only still requires access and output controls.
- Yes
- No, read-only
- Scope undecided
Authority and human control
Is every write independently authorized outside the model?
Evidence: per-user permissions, bounded arguments and approvals for consequential operations.
Scope-dependent: shown when write actions are possible or undecided.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Authority and human control
Are duplicate, delayed and partially completed actions recoverable?
Evidence: idempotency keys, timeout reconciliation, compensation and audit records.
Scope-dependent: shown when write actions are possible or undecided.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Authority and human control
Can a user challenge a result and reach a qualified human?
Evidence: visible limitations, review route, authority to intervene and realistic response expectations.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Deployment and recovery
Are model, prompt, retrieval and tool configurations versioned together?
Evidence: reproducible release manifest and traceable evaluation results.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Deployment and recovery
Have provider failure, throttling and unavailable dependencies been rehearsed?
Evidence: bounded retries, timeouts, degraded mode and a non-AI operating path.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Deployment and recovery
Can you stop or roll back a release without losing the action history?
Evidence: kill switch, rollback procedure and reconciliation of in-flight work.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Deployment and recovery
Do provider or knowledge-source changes trigger re-evaluation?
Evidence: dependency inventory, change alerts and regression criteria.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Operations and economics
Can operators detect quality, access and action failures in production?
Evidence: representative outcome samples, alert thresholds and owned incident signals.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Operations and economics
Are traces useful without exposing secrets or unnecessary personal data?
Evidence: redaction, restricted access, retention and tested incident reconstruction.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Operations and economics
Have peak load, latency and consumption limits been measured?
Evidence: concurrent tasks, token or tool budgets, quotas, p95 latency and abuse controls.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
Operations and economics
Is operating cost measured per accepted outcome, including review and rework?
Evidence: unique accepted tasks, attempts, retrieval, human review and disjoint operating costs.
- Documented and tested
- Partly in place
- Not in place
- Not assessed
How this tool works
Each selected answer maps to a stated observation and a check to perform. Answers are not weighted, added together, or converted into a score, diagnosis, or automatic recommendation.
An unknown answer identifies evidence to collect. Review each observation against your system and constraints; one favorable answer does not establish overall readiness or fit.
Model assumptions
- This does not certify AI safety, legal compliance, or model accuracy.
- Responses are self-reported and have not been independently verified.
Is the workflow bounded by a measurable outcome and an explicit non-AI alternative?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Compare accepted task outcomes with the current process on the same task set.
Partly in place: Coverage is incomplete. A convincing demonstration does not establish a useful production outcome.
Check: Compare accepted task outcomes with the current process on the same task set.
Not in place: A convincing demonstration does not establish a useful production outcome.
Check: Compare accepted task outcomes with the current process on the same task set.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Compare accepted task outcomes with the current process on the same task set.
Are release, risk acceptance and incident decisions assigned to accountable owners?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Assign who may launch, stop, change and accept residual risk.
Partly in place: Coverage is incomplete. Important operating decisions could have no accountable owner.
Check: Assign who may launch, stop, change and accept residual risk.
Not in place: Important operating decisions could have no accountable owner.
Check: Assign who may launch, stop, change and accept residual risk.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Assign who may launch, stop, change and accept residual risk.
Have you assessed who could be harmed by an incorrect answer or action?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Record plausible harms and the controls required before each exposure level.
Partly in place: Coverage is incomplete. The workflow may be tested against convenience rather than its real consequences.
Check: Record plausible harms and the controls required before each exposure level.
Not in place: The workflow may be tested against convenience rather than its real consequences.
Check: Record plausible harms and the controls required before each exposure level.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Record plausible harms and the controls required before each exposure level.
Are data permissions and retrieval scope explicit?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Test allowed and denied users against the same record, including stale access and cross-tenant requests.
Partly in place: Coverage is incomplete. Useful answers can still expose information to the wrong user.
Check: Test allowed and denied users against the same record, including stale access and cross-tenant requests.
Not in place: Useful answers can still expose information to the wrong user.
Check: Test allowed and denied users against the same record, including stale access and cross-tenant requests.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Test allowed and denied users against the same record, including stale access and cross-tenant requests.
Are source quality, freshness and deletion rules verified?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Test source updates, deletion propagation and citations back to the exact supporting record.
Partly in place: Coverage is incomplete. Correctly retrieved information may be stale, incomplete or no longer permitted.
Check: Test source updates, deletion propagation and citations back to the exact supporting record.
Not in place: Correctly retrieved information may be stale, incomplete or no longer permitted.
Check: Test source updates, deletion propagation and citations back to the exact supporting record.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Test source updates, deletion propagation and citations back to the exact supporting record.
Are sensitive inputs, provider handling and retention boundaries reviewed?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Review the exact provider configuration and every storage or logging destination with the data owner.
Partly in place: Coverage is incomplete. Sensitive data could leave its approved handling boundary.
Check: Review the exact provider configuration and every storage or logging destination with the data owner.
Not in place: Sensitive data could leave its approved handling boundary.
Check: Review the exact provider configuration and every storage or logging destination with the data owner.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Review the exact provider configuration and every storage or logging destination with the data owner.
Do you have representative failure cases and a release gate?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Build a representative evaluation set and define failure conditions that stop release.
Partly in place: Coverage is incomplete. A demo cannot establish reliability on difficult or uncommon tasks.
Check: Build a representative evaluation set and define failure conditions that stop release.
Not in place: A demo cannot establish reliability on difficult or uncommon tasks.
Check: Build a representative evaluation set and define failure conditions that stop release.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Build a representative evaluation set and define failure conditions that stop release.
Are unsupported claims, missing evidence and abstention tested?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Check whether the system declines unsupported claims and escalates instead of inventing evidence.
Partly in place: Coverage is incomplete. A fluent response may conceal a lack of supporting evidence.
Check: Check whether the system declines unsupported claims and escalates instead of inventing evidence.
Not in place: A fluent response may conceal a lack of supporting evidence.
Check: Check whether the system declines unsupported claims and escalates instead of inventing evidence.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Check whether the system declines unsupported claims and escalates instead of inventing evidence.
Have malicious instructions in prompts, retrieved content and tool results been tested?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Test instruction attacks with real access controls and validate the output before any privileged use.
Partly in place: Coverage is incomplete. Untrusted content could redirect retrieval, disclosure or tool use.
Check: Test instruction attacks with real access controls and validate the output before any privileged use.
Not in place: Untrusted content could redirect retrieval, disclosure or tool use.
Check: Test instruction attacks with real access controls and validate the output before any privileged use.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Test instruction attacks with real access controls and validate the output before any privileged use.
Are model outputs validated before use by another system?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Reject malformed or out-of-scope outputs at the consuming boundary and test rejection behavior.
Partly in place: Coverage is incomplete. Well-formed text is not a valid or authorized business operation.
Check: Reject malformed or out-of-scope outputs at the consuming boundary and test rejection behavior.
Not in place: Well-formed text is not a valid or authorized business operation.
Check: Reject malformed or out-of-scope outputs at the consuming boundary and test rejection behavior.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Reject malformed or out-of-scope outputs at the consuming boundary and test rejection behavior.
Can the system change records or trigger external actions?
Yes: Write actions require explicit authorization and recovery controls.
Check: Inventory every writable tool and its independent permission boundary.
No, read-only: You describe read-only scope. This reduces action risk, not disclosure or misinformation risk.
Check: Verify credentials and downstream integrations cannot perform hidden writes.
Scope undecided: The authority boundary is unresolved.
Check: Classify every tool and downstream operation as read, write or approval-only.
Is every write independently authorized outside the model?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Attempt unauthorized writes and approval bypasses at the tool boundary.
Partly in place: Coverage is incomplete. A model decision must not grant itself business authority.
Check: Attempt unauthorized writes and approval bypasses at the tool boundary.
Not in place: A model decision must not grant itself business authority.
Check: Attempt unauthorized writes and approval bypasses at the tool boundary.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Attempt unauthorized writes and approval bypasses at the tool boundary.
Are duplicate, delayed and partially completed actions recoverable?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Rehearse an ambiguous timeout and reconcile actual state before retrying.
Partly in place: Coverage is incomplete. Retries could repeat a consequential operation or hide partial completion.
Check: Rehearse an ambiguous timeout and reconcile actual state before retrying.
Not in place: Retries could repeat a consequential operation or hide partial completion.
Check: Rehearse an ambiguous timeout and reconcile actual state before retrying.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Rehearse an ambiguous timeout and reconcile actual state before retrying.
Can a user challenge a result and reach a qualified human?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Test escalation with the responsible team and verify the user can stop or decline the proposed action.
Partly in place: Coverage is incomplete. Users could rely on a disputed result without a practical escalation path.
Check: Test escalation with the responsible team and verify the user can stop or decline the proposed action.
Not in place: Users could rely on a disputed result without a practical escalation path.
Check: Test escalation with the responsible team and verify the user can stop or decline the proposed action.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Test escalation with the responsible team and verify the user can stop or decline the proposed action.
Are model, prompt, retrieval and tool configurations versioned together?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Reconstruct the deployed configuration and repeat its acceptance checks.
Partly in place: Coverage is incomplete. A passing evaluation may not describe the deployed system.
Check: Reconstruct the deployed configuration and repeat its acceptance checks.
Not in place: A passing evaluation may not describe the deployed system.
Check: Reconstruct the deployed configuration and repeat its acceptance checks.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Reconstruct the deployed configuration and repeat its acceptance checks.
Have provider failure, throttling and unavailable dependencies been rehearsed?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Simulate provider and retrieval outages and verify bounded retries and safe handoff.
Partly in place: Coverage is incomplete. An upstream failure could stop the workflow or produce unsafe fallback behavior.
Check: Simulate provider and retrieval outages and verify bounded retries and safe handoff.
Not in place: An upstream failure could stop the workflow or produce unsafe fallback behavior.
Check: Simulate provider and retrieval outages and verify bounded retries and safe handoff.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Simulate provider and retrieval outages and verify bounded retries and safe handoff.
Can you stop or roll back a release without losing the action history?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Rehearse rollback with pending requests and preserve the records needed to reconcile effects.
Partly in place: Coverage is incomplete. Stopping generation does not automatically undo external actions.
Check: Rehearse rollback with pending requests and preserve the records needed to reconcile effects.
Not in place: Stopping generation does not automatically undo external actions.
Check: Rehearse rollback with pending requests and preserve the records needed to reconcile effects.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Rehearse rollback with pending requests and preserve the records needed to reconcile effects.
Do provider or knowledge-source changes trigger re-evaluation?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Define which model, schema, source and permission changes require a new evaluation.
Partly in place: Coverage is incomplete. External changes can invalidate a previously passing release.
Check: Define which model, schema, source and permission changes require a new evaluation.
Not in place: External changes can invalidate a previously passing release.
Check: Define which model, schema, source and permission changes require a new evaluation.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Define which model, schema, source and permission changes require a new evaluation.
Can operators detect quality, access and action failures in production?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Connect operational traces with accepted outcomes, denied access and tool side effects.
Partly in place: Coverage is incomplete. HTTP success alone will not reveal incorrect answers or unintended actions.
Check: Connect operational traces with accepted outcomes, denied access and tool side effects.
Not in place: HTTP success alone will not reveal incorrect answers or unintended actions.
Check: Connect operational traces with accepted outcomes, denied access and tool side effects.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Connect operational traces with accepted outcomes, denied access and tool side effects.
Are traces useful without exposing secrets or unnecessary personal data?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Trace one failed task end to end with redacted inputs, configuration versions and permission decisions.
Partly in place: Coverage is incomplete. Troubleshooting can create a second sensitive-data store.
Check: Trace one failed task end to end with redacted inputs, configuration versions and permission decisions.
Not in place: Troubleshooting can create a second sensitive-data store.
Check: Trace one failed task end to end with redacted inputs, configuration versions and permission decisions.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Trace one failed task end to end with redacted inputs, configuration versions and permission decisions.
Have peak load, latency and consumption limits been measured?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Load-test the actual workflow and verify per-user and total request, token and action limits.
Partly in place: Coverage is incomplete. Average behavior may conceal peak failures and uncontrolled usage.
Check: Load-test the actual workflow and verify per-user and total request, token and action limits.
Not in place: Average behavior may conceal peak failures and uncontrolled usage.
Check: Load-test the actual workflow and verify per-user and total request, token and action limits.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Load-test the actual workflow and verify per-user and total request, token and action limits.
Is operating cost measured per accepted outcome, including review and rework?
Documented and tested: You report supporting evidence for this control. Confirm it covers the current workflow and release.
Check: Measure accepted unique tasks and the complete cost boundary before comparing alternatives.
Partly in place: Coverage is incomplete. Low model cost can hide expensive retries and manual recovery.
Check: Measure accepted unique tasks and the complete cost boundary before comparing alternatives.
Not in place: Low model cost can hide expensive retries and manual recovery.
Check: Measure accepted unique tasks and the complete cost boundary before comparing alternatives.
Not assessed: The control is unverified. Do not treat an unknown as a passed check.
Check: Measure accepted unique tasks and the complete cost boundary before comparing alternatives.
Hypothetical worked example
These illustrative inputs are not your result, a customer case, or a benchmark. Change the inputs above to review your own situation.
- Is the workflow bounded by a measurable outcome and an explicit non-AI alternative?
- Not assessed
- Are release, risk acceptance and incident decisions assigned to accountable owners?
- Not assessed
- Have you assessed who could be harmed by an incorrect answer or action?
- Not assessed
- Are data permissions and retrieval scope explicit?
- Not assessed
- Are source quality, freshness and deletion rules verified?
- Not assessed
- Are sensitive inputs, provider handling and retention boundaries reviewed?
- Not assessed
- Do you have representative failure cases and a release gate?
- Not assessed
- Are unsupported claims, missing evidence and abstention tested?
- Not assessed
- Have malicious instructions in prompts, retrieved content and tool results been tested?
- Not assessed
- Are model outputs validated before use by another system?
- Not assessed
- Can the system change records or trigger external actions?
- Scope undecided
- Is every write independently authorized outside the model?
- Not assessed
- Are duplicate, delayed and partially completed actions recoverable?
- Not assessed
- Can a user challenge a result and reach a qualified human?
- Not assessed
- Are model, prompt, retrieval and tool configurations versioned together?
- Not assessed
- Have provider failure, throttling and unavailable dependencies been rehearsed?
- Not assessed
- Can you stop or roll back a release without losing the action history?
- Not assessed
- Do provider or knowledge-source changes trigger re-evaluation?
- Not assessed
- Can operators detect quality, access and action failures in production?
- Not assessed
- Are traces useful without exposing secrets or unnecessary personal data?
- Not assessed
- Have peak load, latency and consumption limits been measured?
- Not assessed
- Is operating cost measured per accepted outcome, including review and rework?
- Not assessed
Resolve the flagged control gaps before widening production use
22 applicable prompts reviewed. 13 flagged prerequisites need owner review before the affected use or commitment. These are self-reported observations, not proof of readiness, a maturity grade or a recommendation to buy a service. Even an all-positive review needs supporting evidence.
- Is the workflow bounded by a measurable outcome and an explicit non-AI alternative? The control is unverified. Do not treat an unknown as a passed check.
- Are release, risk acceptance and incident decisions assigned to accountable owners? The control is unverified. Do not treat an unknown as a passed check.
- Have you assessed who could be harmed by an incorrect answer or action? The control is unverified. Do not treat an unknown as a passed check.
- Are data permissions and retrieval scope explicit? The control is unverified. Do not treat an unknown as a passed check.
- Are source quality, freshness and deletion rules verified? The control is unverified. Do not treat an unknown as a passed check.
- Are sensitive inputs, provider handling and retention boundaries reviewed? The control is unverified. Do not treat an unknown as a passed check.
- Do you have representative failure cases and a release gate? The control is unverified. Do not treat an unknown as a passed check.
- Are unsupported claims, missing evidence and abstention tested? The control is unverified. Do not treat an unknown as a passed check.
- Have malicious instructions in prompts, retrieved content and tool results been tested? The control is unverified. Do not treat an unknown as a passed check.
- Are model outputs validated before use by another system? The control is unverified. Do not treat an unknown as a passed check.
- Can the system change records or trigger external actions? The authority boundary is unresolved.
- Is every write independently authorized outside the model? The control is unverified. Do not treat an unknown as a passed check.
- Are duplicate, delayed and partially completed actions recoverable? The control is unverified. Do not treat an unknown as a passed check.
- Can a user challenge a result and reach a qualified human? The control is unverified. Do not treat an unknown as a passed check.
- Are model, prompt, retrieval and tool configurations versioned together? The control is unverified. Do not treat an unknown as a passed check.
- Have provider failure, throttling and unavailable dependencies been rehearsed? The control is unverified. Do not treat an unknown as a passed check.
- Can you stop or roll back a release without losing the action history? The control is unverified. Do not treat an unknown as a passed check.
- Do provider or knowledge-source changes trigger re-evaluation? The control is unverified. Do not treat an unknown as a passed check.
- Can operators detect quality, access and action failures in production? The control is unverified. Do not treat an unknown as a passed check.
- Are traces useful without exposing secrets or unnecessary personal data? The control is unverified. Do not treat an unknown as a passed check.
- Have peak load, latency and consumption limits been measured? The control is unverified. Do not treat an unknown as a passed check.
- Is operating cost measured per accepted outcome, including review and rework? The control is unverified. Do not treat an unknown as a passed check.
Checks for this example
- Compare accepted task outcomes with the current process on the same task set.
- Assign who may launch, stop, change and accept residual risk.
- Record plausible harms and the controls required before each exposure level.
- Test allowed and denied users against the same record, including stale access and cross-tenant requests.
- Test source updates, deletion propagation and citations back to the exact supporting record.
- Review the exact provider configuration and every storage or logging destination with the data owner.
- Build a representative evaluation set and define failure conditions that stop release.
- Check whether the system declines unsupported claims and escalates instead of inventing evidence.
- Test instruction attacks with real access controls and validate the output before any privileged use.
- Reject malformed or out-of-scope outputs at the consuming boundary and test rejection behavior.
- Classify every tool and downstream operation as read, write or approval-only.
- Attempt unauthorized writes and approval bypasses at the tool boundary.
- Rehearse an ambiguous timeout and reconcile actual state before retrying.
- Test escalation with the responsible team and verify the user can stop or decline the proposed action.
- Reconstruct the deployed configuration and repeat its acceptance checks.
- Simulate provider and retrieval outages and verify bounded retries and safe handoff.
- Rehearse rollback with pending requests and preserve the records needed to reconcile effects.
- Define which model, schema, source and permission changes require a new evaluation.
- Connect operational traces with accepted outcomes, denied access and tool side effects.
- Trace one failed task end to end with redacted inputs, configuration versions and permission decisions.
- Load-test the actual workflow and verify per-user and total request, token and action limits.
- Measure accepted unique tasks and the complete cost boundary before comparing alternatives.
Further reading
References inform these review topics. This is Ampity’s own worksheet, not certification or a complete framework assessment.