Sample evidence pack
An evidence pack is twelve exports from a running system, generated by a script and never written by hand. This one is a sample drawn from the reference kit: its seed data, its provider zone, its redaction rules and its controls matrix. None of it describes a real client. An item with no seeded example says so, because an empty table would read as a clean result.
§01
What the pack contains
Twelve items, in the order every pack lists them.
- Item 1: Ledger extract
Rows in the sample
- Item 2: Audit log integrity proof
Rows in the sample
- Item 3: Approval log
Rows in the sample
- Item 4: Killswitch test record
Rows in the sample
- Item 5: Ceiling configuration and breaches
Rows in the sample
- Item 6: Model version manifest
Rows in the sample
- Item 7: Provider list and egress proof
Counted in the sample
- Item 8: Schema registry
Rows in the sample
- Item 9: Redaction rules and test results
Rows in the sample
- Item 10: Incident log
Rows in the sample
- Item 11: The controls matrix
Published as its own page
- Item 12: The not-covered register
Rows in the sample
§02
The twelve items
Item 1: Ledger extract
Every metered call for the period, model or tool, with its subject, model version, estimated and actual cost, and outcome. Dry runs are listed apart, so a rehearsal never reads as spend.
Spend
SAMPLE
- Time (UTC)
- 2026-09-01 09:14:22 UTC
- Type
- Model call
- Model
- claude-sonnet-5
- Estimated
- £0.0024
- Cost
- £0.0021
- Outcome
- Booked
- Subject
- Classify inbound complaint 4471 against the deployment’s own category list
- Time (UTC)
- 2026-09-01 09:14:51 UTC
- Type
- Model call
- Model
- claude-sonnet-5
- Estimated
- £0.0024
- Cost
- £0.0000
- Outcome
- Refused
- Subject
- Classify inbound complaint 4472, refused at the spend path while the killswitch was engaged
- Time (UTC)
- 2026-09-01 09:40:12 UTC
- Type
- Model call
- Model
- claude-sonnet-5
- Estimated
- £0.0180
- Cost
- £0.0000
- Outcome
- Refused at ceiling
- Subject
- Summarise the correspondence history for complaint 4471 before the reply is drafted
- Time (UTC)
- 2026-09-01 09:52:30 UTC
- Type
- Model call
- Model
- claude-sonnet-5
- Estimated
- £0.0024
- Cost
- £0.0023
- Outcome
- Booked
- Subject
- Classify inbound complaint 4473 against the deployment’s own category list
Dry runs
SAMPLE
- Time (UTC)
- 2026-09-01 09:13:40 UTC
- Type
- Model call
- Model
- claude-sonnet-5
- Estimated
- £0.0024
- Cost
- £0.0000
- Outcome
- Dry run
- Subject
- Rehearsal of the complaint pipeline, end to end, spending nothing
Item 2: Audit log integrity proof
The audit log's hash chain for the period, each entry with its own hash and the hash of the entry before it, and the result of checking every link.
The chain
SAMPLE
- Entry
- 1
- Time (UTC)
- 2026-09-01 09:31:40 UTC
- Event
- Approval decided
- By
- sample.reviewer
- Applies to
- Approval
- Hash
- f5b79afa3162d97975307271d6f854c5995921969b2651f26bc0df4fcd5d11e2
- Previous entry’s hash
- None: the first entry
- Entry
- 2
- Time (UTC)
- 2026-09-01 09:40:12 UTC
- Event
- Ceiling refusal
- By
- spend-path
- Applies to
- Ledger entry
- Hash
- 314dbb139a9eab2f14da24101a36c3a6fc53d32b7d25a2abf1e884f06b83fd31
- Previous entry’s hash
- f5b79afa3162d97975307271d6f854c5995921969b2651f26bc0df4fcd5d11e2
- Entry
- 3
- Time (UTC)
- 2026-09-01 09:53:05 UTC
- Event
- Incident opened
- By
- sample.operator
- Applies to
- Incident
- Hash
- 162ae58d858bdfdc1f1487a03f39f3021f20e40a1b1dce1e0cff88b9faeed5c6
- Previous entry’s hash
- 314dbb139a9eab2f14da24101a36c3a6fc53d32b7d25a2abf1e884f06b83fd31
- Entry
- 4
- Time (UTC)
- 2026-09-01 09:58:00 UTC
- Event
- Incident updated
- By
- sample.operator
- Applies to
- Incident
- Hash
- 717429e20eb5be138cdd288f57d641d08b1ff8d3449331895946131490f31f01
- Previous entry’s hash
- 162ae58d858bdfdc1f1487a03f39f3021f20e40a1b1dce1e0cff88b9faeed5c6
- Entry
- 5
- Time (UTC)
- 2026-09-01 10:05:00 UTC
- Event
- Incident updated
- By
- sample.operator
- Applies to
- Incident
- Hash
- 3a58bf814626ab3e8ff57b4da8fa291e2f58a5ea7ef229e41ece96cb4104b528
- Previous entry’s hash
- 717429e20eb5be138cdd288f57d641d08b1ff8d3449331895946131490f31f01
- Entry
- 6
- Time (UTC)
- 2026-09-01 10:30:00 UTC
- Event
- Incident updated
- By
- sample.operator
- Applies to
- Incident
- Hash
- 20107286ce77f87c214b306b70f7daaa4d869247512fd97845c8b6e81c390f1e
- Previous entry’s hash
- 3a58bf814626ab3e8ff57b4da8fa291e2f58a5ea7ef229e41ece96cb4104b528
Verification
SAMPLE
- Result
- Verified
- Faults
- None
- Tenant
- 0192f3a1-0000-7000-8000-000000000001
- Entries
- 6
- First entry
- 1
- Last entry
- 6
- Content erased
- 0
- What this shows
- Each entry’s hash was recomputed from the fields it commits to and compared with the hash it stores. Each entry after the first was also checked for the hash of the entry before it. That catches an entry deleted from between two others, and a change to a hashed field unless every hash from that entry onward was recomputed too. Three fields are outside the hash: the retention date, the erasure record and the sample mark. The kit publishes no head hash, so a chain rewritten and recomputed from any entry onward, or cut at either end, would pass as well.
Item 3: Approval log
Every human gate decision: who decided, when, what they saw, and what they decided.
SAMPLE
- Decided
- 2026-09-01 09:31:40 UTC
- By
- sample.reviewer
- Decision
- Approved
- What the reviewer saw
- Proposed category: "Billing dispute". Model claude-sonnet-5. The reviewer saw the model output, the complaint text it was given, and the two categories it ranked below this one.
Item 4: Killswitch test record
Each killswitch drill: the date, the operator, what stopped, the recovery time, and the safe-state behaviour observed.
SAMPLE
- Ran
- 2026-09-01 09:14:45 UTC
- By
- sample.operator
- Recovery
- 180 s
- What stopped
- Every paid model call, at the spend path, before any provider was contacted
- Safe state
- No call was in flight when the flag was set: the previous one had already completed and booked. The next call was refused at the spend path rather than being dropped, so it stands in the ledger as a refusal instead of going missing from it.
Item 5: Ceiling configuration and breaches
The spend ceilings in force, and every breach of them.
Ceilings
SAMPLE
- Ceiling
- Per call
- Limit
- £0.0060
- In force from
- 2026-08-03 08:00:00 UTC
- In force until
- Still in force
- Source
- packages/db/src/seed-data.ts
- Ceiling
- Organisation, rolling 24 hours
- Limit
- £5.0000
- In force from
- 2026-08-03 08:00:00 UTC
- In force until
- Still in force
- Source
- packages/db/src/seed-data.ts
- Ceiling
- Whole system, rolling 24 hours
- Limit
- £20.0000
- In force from
- 2026-08-03 08:00:00 UTC
- In force until
- Still in force
- Source
- packages/db/src/seed-data.ts
Breaches
SAMPLE
- Time (UTC)
- 2026-09-01 09:40:12 UTC
- Ceiling
- Per call
- Limit
- £0.0060
- Already committed
- £0.0000
- Estimated
- £0.0180
- Cost
- £0.0000
- Outcome
- Refused at ceiling
- Subject
- Summarise the correspondence history for complaint 4471 before the reply is drafted
Item 6: Model version manifest
Which model version answered which component, and when that changed.
SAMPLE
- Component
- complaint-classifier
- Model
- claude-sonnet-5
- In force from
- 2026-08-03 08:00:00 UTC
- Why it changed
- First pinned version for this component at deployment
Item 7: Provider list and egress proof
The model providers the system may call, and where the proof lives: the policy test showing that no unmetered call exists outside the provider zone.
The kit's provider zone names 14 hosts and 21 packages that no code outside it may call or import. The pack lists each one by name.
Item 8: Schema registry
The output schemas in force, with their version history and the rejections for the period.
SAMPLE
- Component
- complaint-classifier
- Version
- 1
- In force from
- 2026-08-03 08:00:00 UTC
- In force until
- Still in force
- Rejected responses
- 1
- Schema
- {"$schema":"https://json-schema.org/draft/2020-12/schema","additionalProperties":false,"properties":{"category":{"minLength":1,"type":"string"}},"required":["category"],"type":"object"}
Item 9: Redaction rules and test results
The redaction rules in force, and a test of them run each time the pack is generated.
Redaction rules
SAMPLE
- Field
- decision
- Rule
- Kept as written
- Field
- attestation
- Rule
- Kept as written
- Field
- scope
- Rule
- Kept as written
- Field
- limitMicros
- Rule
- Kept as written
- Field
- committedMicros
- Rule
- Kept as written
- Field
- transition
- Rule
- Kept as written
- Field
- reportingPath
- Rule
- Kept as written
- Field
- every field not named above
- Rule
- Dropped
Redaction tests
SAMPLE
- Result
- Passed
- Withheld
- an email address, a phone number, a postcode, an account number, a name, a bearer token, an internal secret, a driver query, a narrative
- What this shows
- A specimen log line carrying personal data, credentials and free text was redacted by the rules above. Each item listed was missing from the output, and each kept field came out unchanged. That covers this specimen, not every input.
- Result
- Passed
- Kept
- decision, attestation, scope, limitMicros, committedMicros, transition, reportingPath
Item 10: Incident log
Open incidents, and the incident log for the period.
SAMPLE
- Detected
- 2026-09-01 09:52:30 UTC
- Opened
- 2026-09-01 09:53:05 UTC
- Reported
- 2026-09-01 09:58:00 UTC
- Closed
- 2026-09-01 10:30:00 UTC
- Severity
- Low
- Status
- Closed
- Reporting path
- The client’s incident process
- Summary
- The complaint classifier’s output for complaint 4473 failed its schema and was rejected
- Why this path
- The rejected output was reported to the client’s incident process
Item 11: The controls matrix
The controls matrix as it stands.
The matrix is published as a page of its own: The controls matrix
Item 12: The not-covered register
What the offer does not do, and who does it instead.
SAMPLE
- Area
- Fundamental rights impact assessment
- Obligations
- Art. 27
- Who covers it
- Client's legal / DPO
- Area
- Conformity assessment and CE marking
- Obligations
- Art. 43, Art. 47, Art. 48
- Who covers it
- Provider; notified body
- Area
- Training data governance
- Obligations
- Art. 10
- Who covers it
- Model provider
- Area
- Bias and fairness testing of the underlying model
- Obligations
- Art. 10, Art. 15
- Who covers it
- Model provider; client's model-risk function
- Area
- Registration in the EU database
- Obligations
- Art. 49
- Who covers it
- Client's compliance function
- Area
- Watermarking of generated content
- Obligations
- Art. 50
- Who covers it
- Model provider
- Area
- Legal opinion on high-risk classification
- Obligations
- Art. 6 + Annex III
- Who covers it
- Client's legal
- Area
- DPIA authorship
- Obligations
- GDPR Art. 35
- Who covers it
- Client's DPO
- Area
- Threat-led penetration testing
- Obligations
- DORA Art. 26, DORA Art. 27
- Who covers it
- Client's security function; TLPT provider
- Area
- ISO/IEC 42001 certification
- Obligations
- ISO/IEC 42001:2023
- Who covers it
- Accredited certification body
- Area
- Independent model validation
- Obligations
- PRA SS1/23 Principle 4
- Who covers it
- Client's model validation function
- Area
- Employer consultation on workplace AI
- Obligations
- Art. 26(7)
- Who covers it
- Client's HR
§03
Book a scoping call
Thirty minutes on one system: what it does, who is asking about it, and which engagement fits. Nothing to prepare. For firms in the UK and Europe.