Step 4 of 6 · Testing - practitioner kit
Evidence Record Template
Captures one piece of evidence, its stage, status, authority, integrity and handling.
Use this template
A template is not evidence. Fill in sanitized identifiers only.
Purpose
Makes every claim in a run or finding traceable to an attributable source with known limits.
When used: Whenever evidence is captured during a run, and when reviewing a finding or retest.
Identity and capture
| Field | What to record |
|---|---|
| Evidence record ID | Unique identifier, for example EV-0142. |
| Assessment / case / run IDs | Links to the plan, case and run. |
| Captured by | Person or automated collector. |
| Capture time / timezone | When it was collected, not when the event occurred. |
| Source system | Component that produced the record. |
| Source type | Application log, control decision log, tool gateway log, provider audit, system-of-record state, detector alert, transcript. |
| Correlation / request / trace ID | Identifier linking the evidence to the run. |
Classification
| Field | What to record |
|---|---|
| Evidence stage | One of the eight stages below. |
| Evidence status | Observed, unobserved, unavailable or not applicable. |
| Authoritative vs model-visible | Authoritative records come from application, provider or control systems. Model output and narration are model-visible and cannot prove stages 5 to 8. |
| Normalized observation | One factual sentence describing what the record shows. |
| Confidence / limitations | What could make this record misleading: sampling, clock skew, missing fields. |
Integrity and handling
| Field | What to record |
|---|---|
| Raw source location / reference | Path or query to the raw record. Never paste real secrets or personal data; reference them. |
| Integrity / hash / version | Hash or immutable version where the source supports it. |
| Retention period | As defined in the Test Plan. |
| Chain-of-custody / handling note | Who accessed or transformed the record and how. |
| Related evidence IDs | Records that corroborate or contradict this one. |
Valid and invalid evidence by stage
| Stage | Valid evidence / not sufficient on its own |
|---|---|
| 1. Source and provenance | Valid: ingestion log with source and trust label. Not sufficient: the model naming the source. |
| 2. Model-visible context | Valid: orchestration log of assembled context or retrieved chunk IDs. Not sufficient: the fixture existing in the corpus. |
| 3. Model or plan influence | Valid: recorded plan or output compared with baseline. Not sufficient: a single output without a baseline. |
| 4. Proposed action | Valid: structured tool-call proposal with parameters. Not sufficient: prose saying it will act. |
| 5. Control decision | Valid: policy decision record with policy version. Not sufficient: absence of an error message. |
| 6. Submission / execution | Valid: tool gateway or provider execution log. Not sufficient: the model stating 'done'. |
| 7. Downstream effect | Valid: state read from the system of record or provider audit. Not sufficient: 'I changed the record' in model output. |
| 8. Security impact | Valid: stage 7 evidence classified against policy for principal, resource, action and destination. Not sufficient: a detector alert, which supports observation but does not establish impact. |
Evidence status meanings
| Status | Meaning |
|---|---|
| Observed | An authoritative record shows the state occurred. |
| Unobserved | The source was available and shows no such state. Not the same as safe or pass. |
| Unavailable | The source was missing, incomplete or not retained. Leads to Inconclusive where the stage is required. |
| Not applicable | The stage cannot occur for this case design; state why. |
How to use this template
- Create one record per source per stage; do not bundle.
- Write the normalized observation as fact, without interpretation.
- Classify authority before using the record in a conclusion.
- Reference sensitive material by location; keep secrets out of reports.
Common mistakes
- Treating transcripts as authoritative for execution or effect.
- Recording unavailable logs as unobserved.
- Copying real tokens or personal data into evidence notes.
- Using a detector alert as proof of impact.
- No timestamps or time zone, making correlation impossible.