Guide · Review and trust

Evidence-based review: approve tickets with the source in view

Evidence-based review means a reviewer decides on a ticket with the original quote beside it. Relay attaches the source line to every draft and holds it until a person approves. The result is a tracker that records decisions you can explain later, rather than a queue of unverified suggestions.

By the Relay maintainers · Last updated 19 November 2026 · Source on GitHub

What is evidence-based review?

A draft is a claim about what someone said or decided. Evidence-based review makes that claim checkable by showing the source text with it. In Relay, the source can be a chat message, a note, an email line or a transcript segment, quoted verbatim with speaker and timestamp where available. The reviewer is not asked to trust the model; they are asked to compare the draft against its own evidence and decide.

What does a review card show?

The card gathers everything needed for a decision in one place: the quote, the proposed fields, an owner with reasons and the resolved destination. That layout is deliberate, because a reviewer should not have to open three tools to find out where a ticket came from.

Fields shown on a Relay review card
FieldWhy it is shown
Source quoteLets you verify the draft against the original
Type and prioritySets the shape of the work before routing
Suggested ownerComes with reasons and a confidence score
DestinationShows the tracker resolved by policy
Risk and confidenceHelps decide between batch and individual review

How does approval work?

  1. Read the quote. Check that the source supports the proposed ticket.
  2. Adjust the fields. Edit the title, description, priority, owner or labels as needed.
  3. Decide. Approve, reject with a reason, delegate or merge the draft.
  4. Push. Approved drafts pass through an idempotent outbox to one tracker issue each.

In an illustrative card, a draft about a billing webhook shows the quoted sentence, a suggested owner and the destination tracker before any issue exists.

Illustrative demo data.

Why not auto-approve everything?

Extraction on messy multilingual fixtures scores about 0.78 precision and 0.75 recall, and live output varies between runs. Auto-approving every draft would push plausible errors straight into a shared tracker. Relay does allow low-risk drafts to be auto-approved under policy, so trivial items can move quickly while anything risky gets a person. The policy is explicit and testable, not a hidden threshold.

What does the record keep?

Decisions, commitments and their sources stay linked to the ticket they produced, so you can ask why an issue exists months later. Standup and weekly digests are built from the same record. Every write is attributed in an append-only, hash-chained audit log. A rejection is stored as feedback, which gives future suggestions a better starting point without hiding the original decision.

Next step

Run the credential-free demo and review a draft against its quote. Then read the triage guide to see where this gate sits in the full loop.


Self-host Relay Read the workflow guide