Example: Evidence Rail

An AIEP evidence rail is the ordered list of source artefacts that an AI response was built on. Each artefact is hashed and timestamped. Together they form a verifiable record of what the AI retrieved.

What a rail contains

Each entry in the evidence rail includes:

  • artefact_id — unique identifier within this rail (ev_001, ev_002 …)
  • source_url — where the document was retrieved from
  • confidence_tier — verified, high, medium, low, or unverified
  • content_hash — SHA-256 fingerprint of the source at retrieval time
  • retrieved_at — ISO 8601 timestamp
  • validation_status — passed, failed, or pending

Example: three-artefact rail

This is the evidence rail for the sample question “How can an AI answer be verified?”:

Artefact 1

artefact_id:       ev_001
source_url:        https://aiep.dev/schema/evidence.json
confidence_tier:   verified
content_hash:      f3a2c4e0b6a7d991d3e2f5a9c8b7e6f1d0c9b8a7e6f5d4c3b2a1908172635445
retrieved_at:      2026-04-25T00:00:00Z
validation_status: passed

Artefact 2

artefact_id:       ev_002
source_url:        https://aiep.dev/schema/response.json
confidence_tier:   verified
content_hash:      a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2
retrieved_at:      2026-04-25T00:00:00Z
validation_status: passed

Artefact 3

artefact_id:       ev_003
source_url:        https://aiep.dev/.well-known/aiep/index.json
confidence_tier:   verified
content_hash:      9f8e7d6c5b4a3928171615141312111009080706050403020100ffeeddccbbaa99
retrieved_at:      2026-04-25T00:00:00Z
validation_status: passed

Rail validation summary

Artefact count:         3
All hashes present:     true
All sources reachable:  true
Status:                 passed

How to use the hash

The content_hash is computed at the moment of retrieval. To replay the verification:

  1. Fetch source_url again
  2. Compute SHA-256 of the response body
  3. Compare to the recorded content_hash

If they match: the source has not changed since retrieval. If they differ: the source has changed. The discrepancy is now documented.

Machine-readable version

Download evidence-rail.json →

Download evidence.json (single artefact) →

Try it

Verification Playground →

View the full response example →

GitHub →

Machine endpoint →

See also: Verify AI Output · RAG vs AIEP · Deterministic AI · Example: Audit Trail