Sample evidence pack
Sample ISO 42001 evidence pack: Each control checked against its record
A Statement of Applicability extract and evidence index for our own demo organisation, not a client, with real records from two demo AI systems and the control-to-record check run on them. Veriprajna wrote the documents, built the systems and ran the check.
https://veriprajna.com/ai-services/iso-42001-implementation/sample-evidence-pack
The pack: Ten included controls, each checked against its record
Nine control areas in our own words, seven included, and ten controls traced to records of three kinds. Four rows are gaps: no person approved those documents.
- An internal review, not an audit. Veriprajna wrote the demo organisation's documents, built its AI systems and wrote and ran the check. No person audits work they built or run, so this check is labelled an internal review. It issues no certificate, attestation or letter.
- AI spot-check, not a person's review. An AI reviewer that did not build the demo re-checked the extract, the expected outcomes, the documents, the index rows and every figure against the raw run files. It changed no label and no figure. No person has reviewed the labels.
- Not the records behind the service page's example. That example shows an approved policy and approved assessments for a support assistant. This sample is a separate set from our own demo, and none of its documents is approved.
- The shops, customers and tickets are invented; no client system or data appears here.
Statement of Applicability with justifications
Each control area with its decision and reason, and what each included area points to. Here the areas are our own list, and the reasons are the demo's.
Evidence index
One row per included control: the record, its owner role, where it lives, when it was last produced and, for AI system controls, the AI system it came from.
Readiness check: Every included control against its record
Each row checked and what is missing reported. In this sample the missing items are four approvals.
The check failing on purpose
Five rows changed by us (planted by us), so each rule is seen failing as well as passing.
Source: The demo's run files: one counted run of the real index and one of the planted copy (planted by us), which replay byte for byte from the same files. No model was called. The operational records are read in place from four sibling demos and never edited.
- Summary
Our own demo organisation, not a client. Veriprajna's demo set-up runs two demo AI systems on fake data, locally and never deployed: a support agent for an invented shop, and a separate ticket triage assistant for another invented shop. The figures describe this demo only.
Result on the evidence index as written. 10 rows: 6 pass and 4 are gaps. All four gaps are the same: no recorded approval, on the AI policy (two rows), the impact assessment and the supplier assessment.
Why those four are gaps. Veriprajna's AI build agent wrote those documents for the demo and no person approved them, so the check reports the missing approval instead of a made-up one.
What held on every row. All 10 records exist unchanged, have an owner role, sit where the index says and fall within their review interval. All 6 operational records came from the AI system the index names.
Planted gaps (planted by us). On a copy of the index with five rows changed by us, each planted gap was flagged by exactly the rule expected for it. Those counts are kept apart from the result above.
- Scope and AI systems
Two AI systems are in scope. Each uses one model bought in from the same model provider, unchanged, and the two systems use different models; the provider and the models are not named in this sample.
AI system and owner role Version in the records, and records it produced Support agent: answers customer messages and acts on orders through tools. Owner role: Support agent owner As delivered in the evaluation demo; and a later version with an action gate, in the guardrails demo. Records it produced: C03, C05, C07, C08 (as delivered); C09 (with the gate) Ticket triage assistant: helps a support agent work through their open ticket queue (summaries, next steps, policy answers, draft replies); no tools. Owner role: Ticket triage assistant owner The configuration counted in the red-teaming demo. Records it produced: C06 Owners are roles, never people. The demo's roles register has five roles: AIMS owner, support agent owner, ticket triage assistant owner, platform engineering lead and head of support operations. Every role is unfilled; no person is named anywhere in the pack.
Two systems, not one. Three of the four sibling demos run the same support agent; the red-teaming demo runs a different assistant. Its record is attributed to the ticket triage assistant, never to the support agent.
- Statement of Applicability extract
Area (our list), decision and what it points to Why (the demo's text; Area 8 reworded) Area 1: AI policy. Included: C01 Required by our own AI policy commitments; every other control traces back to it. Area 2: Internal organisation. Included: C02 Each AI system needs an owner with authority, and people need a way to report a concern. Area 3: Resources for AI systems. Included: C03 Both systems depend on a bought-in model, libraries and a run environment that must be known to re-test them. Area 4: Impact assessment. Included: C04 Both systems act for or on customers; their possible effects on people must be assessed. Area 5: AI system life cycle. Included: C05, C06, C07, C08 Both systems are changed and re-run; each change must be verified, and operation must be monitored and logged. Area 6: Data for AI systems. Excluded: nothing No AI system in scope is trained, fine-tuned or built on data the demo organisation acquires: both use a bought-in model, unchanged. The test cases they are checked against are covered under verification and validation (C05, C06). Area 7: Information for interested parties. Excluded: nothing The demo AI systems have no external users, customers or other outside parties: they run on fake data in a local test set-up and are never deployed. Area 8: Use of AI systems. Included: C09 The support agent can move money and change orders; its actions must stay within the limits set in its gate rules. Area 9: Third parties and customers. Included: C10 Both systems depend on one model provider. Our list, in our own words. The nine control areas are our own list, in the order the service page uses, and the control wording is ours. The numbers 1 to 9 are our list order only. No control ID is used and no text of ISO/IEC 42001:2023 is quoted; IDs are added only after a re-check against the standard's text.
Totals. 9 areas listed: 7 included and 2 excluded, each exclusion with a written reason. The 7 included areas hold 10 controls.
- The reasons are the demo's text, written by Veriprajna's AI build agent; Area 8's reason is reworded to name the gate rules that set its limits. Area 1's reason points back to the AI policy itself, so read it as the demo's wording, not as a model justification.
- Both exclusions are decisions for our own demo organisation, which trains no model and has no outside users; a real organisation decides them from its own risk treatment. "Data for AI systems" is excluded although both systems read customer records and ticket text when they run.
- Evidence index
Control; record and where it lives; owner, last produced, review interval Check C01 An AI policy is documented, sits beside the other policies, and is reviewed. (Organisational) Record: AI policy. Lives in: Document system: Policies. Owner: AIMS owner. Last produced: 2026-11-18. Interval: 365 days. Gap: no recorded approval C02 AI roles and responsibilities are assigned, and there is a way to report concerns. (Organisational) Record: AI policy, section 3 (roles) and section 4 (reporting a concern). Lives in: Document system: Policies. Owner: AIMS owner. Last produced: 2026-11-18. Interval: 365 days. Gap: no recorded approval C03 The model, settings, libraries and environment each AI system relies on are documented. (Operational) Record: Run configuration of the support agent's test runs (model, settings, libraries, environment). Lives in: Evaluation demo: results. From: Support agent, as delivered in the evaluation demo. Owner: Support agent owner. Last produced: 2026-11-09. Interval: 365 days. Pass: all 4 checks C04 The impacts of each AI system on individuals, groups and society are assessed and the results documented. (Assessment) Record: Impact assessment: Support agent and ticket triage assistant. Lives in: Document system: Assessments. Owner: Head of support operations. Last produced: 2026-11-18. Interval: 365 days. Gap: no recorded approval C05 Verification and validation: the support agent is tested against a pass threshold set in advance. (Operational) Record: Test run against the pass threshold (48 cases, 3 trials each). Lives in: Evaluation demo: run store. From: Support agent, as delivered in the evaluation demo. Owner: Support agent owner. Last produced: 2026-11-09. Interval: 90 days. Pass: all 4 checks C06 Verification and validation: the ticket triage assistant is tested against a frozen attack suite. (Operational) Record: Attack suite result before the fix (5 frozen attacks, 30 attempts each). Lives in: Red-teaming demo: run results. From: Ticket triage assistant, the configuration counted in the red-teaming demo. Owner: Ticket triage assistant owner. Last produced: 2026-11-11. Interval: 90 days. Pass: all 4 checks C07 Operation and monitoring: the support agent's alert rules run and their alerts are logged. (Operational) Record: Alert log over a quiet period (8 simulated hours, no fault planted). Lives in: Observability demo: run store. From: Support agent, as delivered in the evaluation demo. Owner: Support agent owner. Last produced: 2026-11-16. Interval: 30 days. Pass: all 4 checks C08 Event logs: every decision of the support agent is logged in a store that shows later edits. (Operational) Record: Decision log of the same quiet period (hash-chained audit store). Lives in: Observability demo: run store. From: Support agent, as delivered in the evaluation demo. Owner: Support agent owner. Last produced: 2026-11-16. Interval: 30 days. Pass: all 4 checks C09 Responsible use: the support agent's actions pass a gate, and gated actions wait for approval. (Operational) Record: Gate decision records of the gated agent (24 tasks, 3 runs each). Lives in: Guardrails demo: run store. From: Support agent, with the action gate (guardrails demo). Owner: Support agent owner. Last produced: 2026-11-13. Interval: 30 days. Pass: all 4 checks C10 Suppliers are assessed before use and reviewed. (Assessment) Record: Supplier assessment: The model provider. Lives in: Document system: Assessments. Owner: Platform engineering lead. Last produced: 2026-11-18. Interval: 365 days. Gap: no recorded approval Record types. 2 organisational, 2 assessment and 6 operational rows: 10 rows for 10 controls, pointing at 9 distinct records, because the AI policy serves C01 and C02.
How to read the check column. An operational row is checked on R1 to R4 and a policy or assessment row on R1, R2, R3 and R5 (the rules are in the next part). "Pass: all 4 checks" means every rule that applies held. Each gap row passed R1, R2 and R3 and failed R5.
- C01, C02, C04 and C10: written by Veriprajna's AI build agent for our own demo organisation; no person wrote or approved them, so the check reports no recorded approval.
- C03 covers one system. The C03 record covers the support agent only; the ticket triage assistant's configuration is not indexed. The check does not judge whether a record covers everything its control names.
- C07 and C08: recorded API replies of the same demo agent, recorded by the evaluation demo and replayed; not live traffic.
- Weak dates. Three records carry no date of their own; their dates come from the producing demo's build notes (C03, C07, C08). Of the 10 dates, 4 come from the record itself, 2 from a run manifest, 1 from a run ledger and 3 from build notes. The check reports this as a count and does not fail it.
- The review intervals are the demo organisation's stated policy, set after the records' dates were known; the date rule is shown failing only in the planted gaps (planted by us).
- The check, rule by rule
Deterministic code, no model. The rules, the extract, the expected outcome of every row and the planted gaps (planted by us) were written to files before the check code existed.
Rule What it checks on the row's record Rows, passed R0 Every included control has exactly one row, and every excluded area has a written reason The whole index: passed R1 The record exists at the indexed place and is unchanged (its file hash matches the index) 10 rows, 10 passed R2 The owner is a role in the roles register, never a person, and the record lies in the stated location 10 rows, 10 passed R3 The last-produced date, read from declared evidence, is within the control's review interval 10 rows, 10 passed R4 Operational rows: the model and prompt identity in the record or its run's files is the named AI system version's 6 rows, 6 passed R5 Policies, role assignments and assessments: an approval is recorded in the approval register 4 rows, 0 passed - R1 stops the row. If a record is missing or changed, R2 to R5 are not checked for that row.
- R4 counts what carries no identity. Entries that carry no identity are counted and reported, not failed; any entry naming another model or prompt fails the row. Identity was shown by the record itself for 2 rows, by the record and its run's files for 2, and by the run's files only for 2 (C07, C09).
- R5 needs an approver role. An approval counts only from a role in the roles register, dated on or after the record. An approval by the AI build agent does not count. The approval register is empty, because no person approved anything.
- Expected outcomes matched on all 10 rows of the real index and all 10 rows of the planted copy (planted by us). A difference would have been reported, never fixed by editing the expected outcome.
- What the operational records show
Read from each operational record by the check. These figures are read from the sibling demos' own run files. The C05, C06 and C07 figures also appear in the sibling reports, which were fact-checked; the C08 and C09 counts appear only here and were re-derived from the same run files by an AI spot-check. The check confirms only where each record came from.
Control and record Read from the record, and its label C05: Test run against the pass threshold 144 of 144 test episodes passed; score 1.0 against a threshold of 0.979167, set in advance by a fixed rule from 20 unchanged runs. Label: Results on a test set, not a measured rate C06: Attack suite result, before the fix 0 leaks in 150 attack attempts; 1 refusal. Label: Planted weakness (planted by us): planted, open C07: Alert log over a quiet period 32 alerts over 8 simulated hours and 4,800 requests. Label: No fault planted in this period C08: Decision log 6,688 log records, of which 4,800 are decisions. Label: References to content, not the content C09: Gate decision records 186 gate decisions over 72 task runs; 0 approval requests. Label: Main-gate arm only; results on a test set - C06: 0 leaks in 150 attack attempts: not observed for these attacks, not proof the assistant is not vulnerable. The planted weakness is planted, open: the chat window fetches images from any host (planted by us). It is not a finding.
- C07: 32 false alerts in a quiet period with no fault planted; not incidents.
- C08: 4,800 of 6,688 log records are decisions, each naming the support agent; the other 1,888 carry no identity. None names another system.
- C09: 186 gate decisions over 72 task runs, all allowed; 0 approval requests, so this record does not show the approval step. The same run also holds arms with a prompt planted by us; they are not in this record.
Where each record comes from (tests run by Veriprajna, which built this stack):
- C03 and C05: the evaluation demo, in the sample LLM evaluation report.
- C06: the red-teaming demo, in the sample AI red teaming report. It tests the ticket triage assistant (that report calls it a customer-support assistant), a different AI system from the support agent.
- C07 and C08: the observability demo, in the sample drill report.
- C09: the guardrails demo, in the sample AI agent guardrail testing report.
- The four gaps
Four rows, one gap: no recorded approval.
Control, record and owner role Why it is a gap C01: AI policy. Owner role: AIMS owner No approval in the register C02: AI policy, roles and reporting sections. Owner role: AIMS owner The same unapproved policy C04: Impact assessment: both AI systems. Owner role: Head of support operations No approval in the register (it was also written without any person; no rule tests that) C10: Supplier assessment: the model provider. Owner role: Platform engineering lead No approval in the register (it was also written without any person; no rule tests that) Each document says in its first lines that the AI build agent wrote it for the demo and that no person approved it. The check does not count a document we wrote as approved.
Neither document is complete. The policy refers to information security and data protection policies that are not in the pack. The supplier assessment leaves three of its four checks "Not yet checked".
- Planted gaps (planted by us)
Five rows changed by us on a copy of the index, to show each rule failing. Every change is planted by us; these counts are kept apart from the real result and are never a rate.
Row and planted change (planted by us) Rule that flagged it, and result C01: The row points at a copy of the AI policy edited after indexing, keeping the original's hash (planted by us) R1 record exists and is unchanged. Flagged by the expected rule; later rules not checked C03: The row points at a record that does not exist (planted by us) R1 record exists and is unchanged. Flagged by the expected rule; later rules not checked C05: A support agent row points at the ticket triage assistant's attack-suite record, still naming the support agent (planted by us) R4 came from the named AI system. Flagged by the expected rule C07: The row points at a stale alert-log record dated 2026-07-01 (planted by us) R3 last produced within the review interval. Flagged by the expected rule C08: The owner is removed from the row (planted by us) R2 owner role and location. Flagged by the expected rule - 5 planted gaps (planted by us), 5 flagged by exactly the expected rule, and by no other.
- On the planted copy (planted by us), 2 rows pass and 8 are gaps: the 5 planted gaps plus the 3 real approval gaps it keeps (C02, C04, C10).
- The stale record (planted by us) is 140 days old against a 30-day interval.
- The edited policy copy (planted by us) differs from the indexed policy by one sentence.
What this check shows, and what it does not
What a pass means, and what it does not
- A pass is about the record's place in the system. The record exists unchanged, has an owner role, sits where the index says and is within its review interval; for an AI system control, it came from the named AI system.
- It does not judge content. Whether a record is correct or sufficient is a person's review, and the check does not do it.
- It does not judge the sibling demos' figures. It reads them from the run files and confirms where each record came from.
- One fixed as-of date. Every record's date is judged against one fixed date, so a replay on a later day gives the same result. With a later as-of date the shorter intervals start to fail: that is the date rule working.
- Row by row. The limits of particular rows sit beside them in the pack above.
Who wrote the labels, and when they were fixed
- One author. The extract (every include and exclude decision and its reason), the roles register, the review intervals, every index row, the expected outcome of every row with a reason, the planted gaps (planted by us) and the three documents were written by Veriprajna's AI build agent.
- Fixed before the check existed (UTC, clock times on the run day): the extract at 11:30:38, the rules at 11:30:54, the expected outcomes at 11:32:28 and the planted gaps (planted by us) at 11:32:42. A freeze log holds the hash of every input file. The check code followed, and the one counted run of both arms was at 11:34:11.
- No model judged anything. Every outcome is decided by code.
Runs excluded, re-run or changed
Each arm ran once, and nothing was re-run to change a result. One freeze entry was rewritten before the counted run, because timestamps typed into six input files were ahead of the clock; they were corrected to the files' save times, and no rule, decision or label changed.
Run configuration
A local macOS workstation with Python 3.9.6, standard library only (SQLite 3.51.0). No model was called, nothing was spent, no network was used, and the four sibling demos were read in place and never edited.
Replay. 11 replay tests (pytest 8.4.2) passed. They fail if a replay differs by one byte, if an input file changed after the freeze, or if a sibling record changed after it was indexed.
In an engagement
On the service page, the evidence index names, for each control you include, the record, its owner, where it lives and the date it was last produced, and for AI system controls the AI system it came from.
Before your internal audit we check every included control against its record and report what is missing; that is a readiness check, not an audit. Organisational controls point to an approved policy or role assignment, and assessment controls to an assessment written with your people and approved by a named owner.
Know which of your included controls has a record behind it
This pack is our own demo organisation. For yours, we build the AI management system and check each included control against its record, starting with a short discovery call.
- Email or WhatsApp us
- Reply within 1 business day
- Free 30-minute discovery call
- Written scope and fixed fee within 2 business days of the call



