Skip to content
Flows

What Flows just did

Based on your request to build Evals for Your Harness, Flows created 4 implementation stages and 12 checks.

You asked for: Stand up a real eval system: representative task sets, automatic and rubric grading, statistical honesty, and eval-gated harness changes.. Flows generated 4 stages because the application requires An eval task set representative of real usage, tiered by difficulty; Automatic grading where possible (tests pass, output validates) and rubric grading elsewhere; Multiple runs per task with variance reported - single runs lie. Each stage tells your AI builder what to do, tests whether it actually worked, and gives a repair instruction when it fails. These checks are designed to catch: Evals that test toy tasks nothing like real usage; Single-run scores treated as truth despite run-to-run variance; Grading by the same model that did the work, uncalibrated.

Example failure

The form reported success, but no database record was created.

What Flows does

Flows detects the failure, generates a repair instruction, reruns the test, and stores the passing evidence.

Stages — Plan → Build → Test → Fix → Prove → Ship

4 stages · 12 checks · progressive disclosure — open a stage for details

Copy to external builder

Copies a prompt for Claude/Cursor/etc. Records prompt_copied only — not execution, commits, or checks.

Execute through Oort

Requires Sign in with Oort + a provider connection (BYOK). Keys stay on Oort. A model name in the dropdown is not the same as a live connection.

Next

Check off: Tasks mined from real usage

Add a proof note on each check — a checkbox alone is weak proof.

Probes & advanced tools are under “More verification tools”. Experience level: Settings.

Evals for Your Harness
Harness Engineering90-150 minutes

Project (saved with proof)

Project binding: Unbound — complete project binding before trusting this run · missing repositoryUrl, repositoryStartCommit, environment, stack

More verification tools (probes, timeline, adapters)Show

App persistence (create → read → assert token)

Body template: {"probe":"{{token}}","title":"flows-probe-{{token}}"}. Needs CORS-readable public create/list endpoints (or inconclusive). Badge only when create→read→refresh finds the token (detects false success).

Two-account authorization (User B cannot see/change User A)

Authorization evidence source: Cross-account ownership test (trust level 5/6). These are not equivalent. Unauthenticated probe alone is weaker (source: unauthenticated API probe). Tokens stay in this browser session only — not uploaded to Flows servers.

Generated edge-case tests (review / approve)

  • api_probe Submitting the form empty is rejected with a clear error
  • api_probe Each required field can be omitted and fails validation
  • api_probe Malformed email is rejected
  • api_probe Logged-out users cannot open protected pages
  • api_probe Protected API rejects requests without credentials

Generated from this route’s signals — you approve; you do not invent security tests.

Content probe (GET body contains text)

In-product content probe reads response text when CORS allows — not full DOM/click automation. Playwright script is optional and external.

Proof strength

thin · 0/100

Heuristic from notes + probes + gates — not a production certification.

Verification adapters

  • Deploy URL probe · idle
  • API path probe · idle
  • Authz probe (unauth → 401/403) · idle
  • Browser persistence · idle
  • App persistence (create→read) · idle
  • Content / body text probe · idle
  • In-product Playwright DOM · not embedded (download external script)

Filled circles = available here. Empty = not shipped. Probes are not a production audit. Adapter docs

0/4 steps0%
Recording plan identity…

Step 1 of 4

Not started

Design the task set

Representative, tiered, and partially held out.

Flows keeps plan, checks, failures, and next steps connected

Your AI tool writes or changes the code

Repository grounds the plan only after inspect

Checks always carry a validation tier

Not checkedChecks not locked

How this plan was created

Plan from project details

Flows creates steps using the goal, requirements, stack, platform, and details you provide.

  • Project description
  • Selected template / route
  • User-entered stack
  • Constraints and notes
  • Evals for Your Harness

No repository inspection implied. Connect and inspect a codebase to ground steps in real files.

Codebase access

Checking GitHub…

Or paste a repository URL manually
Plan instructions

Use these instructions in Claude, ChatGPT, Cursor, Emergent, or another builder.

Build the eval task set from real usage. Mine transcripts/issues for 15-25 tasks across the classes that matter: multi-step builds, bug fixes with verification, research/navigation questions, permission-sensitive operations, summarizations, and refusal-appropriate requests. Tier by difficulty (smoke: must always pass; standard: the working band; stretch: currently unreliable). For each task write: the exact starting state (repo fixture, files), the request, and success criteria (machine-checkable where possible: tests to pass, artifacts to exist, strings to appear). Hold out 20% of tasks - never used when tuning prompts, only for validation. Package fixtures reproducibly.

The AI tool writes code — Flows only organizes the plan · Complete implementation prompt with explicit requirements

Project context

Copies the project goal, technical details, current step, previous progress, and checks so your AI tool understands what it is working on.

Expected after this step

A tiered, fixtured task set with held-out validation tasks.

Should not happen

  • Evals that test toy tasks nothing like real usage
  • Single-run scores treated as truth despite run-to-run variance
  • Grading by the same model that did the work, uncalibrated
  • A suite that exists but never gates anything

Verify gate — prove this step works before continuing

Do not move on until every check is true. Add proof notes — a checked box alone is weak proof. Checks are manual by default; deploy/API probes live under “More verification tools” and do not auto-check boxes.

Do not continue if…

  • !Evals that test toy tasks nothing like real usage
  • !Single-run scores treated as truth despite run-to-run variance
  • !Grading by the same model that did the work, uncalibrated
  • !A suite that exists but never gates anything

Repair path — if this step fails

Use the Repair Prompt when a verify gate fails. Loop: Detected → Diagnosed → Repair issued → Changed → Retested → Resolved

Show repair prompt textShow

Tasks are toy-sized. Take three real sessions from history and convert them into fixtures verbatim - real mess (ambiguity, partial context) is exactly what the suite must contain.

Your notes for this step

Definition of Done

Final route-level requirements. Manual checks — separate from per-step verify gates. Completing step gates does not auto-check these. Route completion is not a production-ready claim.

0/4 · manual