Review & proof

A result is stronger when you can see why it passed.

Ysra separates implementation from review and binds evidence to the candidate that was actually checked. When it can’t prove something, it says so.

The same claim cannot approve itself.

Four rules

What makes the evidence mean something.

01

Acceptance criteria become the contract

What you asked for — and the acceptance criteria you gave — define what review checks. Not what the implementation happened to do.

02

Evidence is bound to one candidate

Checks, screenshots, and findings attach to the exact files that were checked. Stale evidence cannot approve different code.

03

Checks and reviewers have different jobs

Deterministic checks — tests, types, lint, build — report facts. Independent reviewers judge the outcome against the request. Product policy decides what passes; no vote ships software.

04

Partial work stays inspectable

An incomplete result is preserved, labelled, and deliverable as what it is — never rewritten into a success.

Honest failure

Not every red is a bug.

Failures are classified before anything happens next. Only real product findings send code back for repair. An unavailable browser or reviewer is never silently turned into a code defect — or into a pass.

  1. ProductThe candidate doesn’t do what was asked.Back to repair
  2. TestThe check itself is wrong or flaky.Fix or flag the check
  3. ReviewerA review step failed or disagreed with the evidence.Re-review, not rewrite
  4. InfrastructureA browser, runtime, or service wasn’t available.Marked unverified
  5. Missing evidenceThere is no proof either way.Marked unverified

Independent review

Specialist review, scaled to the risk.

Ysra’s Review Council assesses a candidate against the acceptance contract and the surfaces the change touches. Low-risk changes get a light path; changes to authentication, payments, persistence, migrations, external APIs, deployment, or background jobs get more depth.

Which specialists run depends on the change and the review profile. Not every specialist runs on every task, and review can be advisory or required by policy.

  • Changed behaviour and blast radius
  • Code correctness
  • Repository architecture
  • Native tests, lint, typecheck, and build
  • Requirements acceptance
  • API behaviour
  • Browser workflows
  • Security boundaries
  • Visual and accessibility obligations
  • Deployment or job reliability
  • Evidence completeness and candidate identity

Browser evidence

The journey, replayed.

For runnable web products, browser verification keeps the scenario, the navigation steps, screenshots, the viewport, and the requests observed — with a clear pass, fail, or unverified.

Example verification replay
localhost:3000/dashboard
Your tasksNew task
  • Draft Q4 plan
  • Review onboarding copy
  • Call the venue about Friday

↻ reloaded · 200 OK · item still present

Candidate 7f3a2c1Screenshot 6/6Requests observed: POST /api/tasks · 201

The Build Report

When Ysra can’t prove a requirement, it says so and keeps the candidate.

Partial and unverified states are first-class results, not error screens. You decide whether to resume, deliver as unverified, or stop.

Build Report

Organization invites

candidate 7f3a2c1acme/web · ysra/session-invites
  1. 01

    Native checks

    Uses the repository’s own test, type, lint, and build commands.

    SHELL✓ passed
  2. 02

    Browser workflows

    Exercises applicable user journeys in the running candidate.

    BROWSER✓ 4 of 4 steps
  3. 03

    Independent review

    Reviews the requested outcome separately from implementation.

    REVIEW✓ no findings
  4. 04

    Exact candidate identity

    Keeps evidence attached to the files that were actually checked.

    IDENTITY✓ 7f3a2c1
  5. 05

    Honest stops

    Preserves partial work and says what remains unverified.

    PARTIAL◐ 1 unverified

Why it stopped: “Invite emails render in the mail preview” could not be verified — no mail sandbox was configured. The candidate is preserved; you decide.

Review the changesOpen UNVERIFIED PRResume from checkpoint

Illustrative Build Report. Labels match the product; the session is fictional.

Judge the work by its evidence.

Start a software task and read the Build Report — including the parts that didn’t pass.