A receipt for every task: what was done, what was checked

The idea

A receipt is the record that closes every Crewshal task. It says what was done, which checks ran and how they ended, how long the work took and what it cost (an estimate at list price, not an invoice). The platform seals it from records it already keeps, so the receipt shows exactly what was checked, and a later change to the work's output, checks or recorded activity marks it stale.

The problem
An AI agent's "done" is a claim; without a record you must trust it or check the work again yourself.
Why it matters
You read the output, the checks, the time and the cost in one place, and a check that could not run shows as not passed.
Where it stands
In production as of A production test task ended with a receipt that said Verified, 3 of 3; the worker checked its own work, and the run as a whole is recorded as partial.

Crewshal team · · 3 min read

Recorded run · Production, reserved test workspace · November 12, 2026

From one request to a sealed receipt

One request in production goes through triage, a start, three stage gates and a saved file, and ends with a receipt that says Verified.

Step 1 of 7Request

PersonOne request

A person types one request in Flow and sends it: a saved Markdown checklist for a fictional bicycle repair shop, with exactly four steps in a set order.

InputThe request, as sent

Write a saved Markdown onboarding checklist for a fictional bicycle repair shop. Use only these four steps: receive the bike; record the requested repair; inspect the brakes; call the owner when it is ready. Include one heading and exactly four checkbox rows in that order. Keep the result under 120 words. Work within this task yourself. Do not hire, delegate, create child tasks, contact anyone, publish, connect accounts, run commands, or spend money.

Where it ran. Production at crewshal.com, in a reserved test workspace with one AI employee; a script played the person's clicks in a real browser. The run as a whole did not pass: three required checks failed (the request preview was slower than its 100 ms target, the first automatic reading of the request found no type and no acceptance criteria within 5 seconds, and the record does not tie every requirement to one exact check), and 4 more did not run. The record is marked partial; a complete pass of the whole request flow, and customer acceptance, are still open.

The recorded files behind this replay
  • Production run record: gates F01 to F17 with times sha256 5732757656ca
  • Acceptance record: what the run proved, and what stayed open sha256 0d78491d2d5f
  • The saved checklist, as downloaded sha256 0d3030d761c3
  • Production capture of the receipt sha256 6e9dffbdf483

Why it matters

"Done" from an AI agent is a claim. Without a record you must trust the claim, or check the work again yourself.

A receipt turns the claim into something you can read: the output, the checks and their results, the time and the cost. When a check could not run, the receipt says so instead of hiding it.

How it works

  1. The work saves real output. A document revision with its content hash, a file with its size and SHA-256, or a recorded action result such as a change or a message.
  2. The checks run on that output. The platform records each result and the exact output it checked. For Verified, each acceptance criterion of the plan must be covered by a passing check.
  3. The receipt is sealed. When the task moves to done, the platform collects the output, the checks, the actions, the time and the cost from those records, in the same step.
  4. The state says how far the proof goes. Verified needs real output, a passing platform completion check, current checks that all passed before completion, and coverage of every acceptance criterion in the current plan. A check that failed or could not run leaves the receipt Unverified, with the reason; it is not counted as a pass.
  5. It stays honest over time. When the output, the checks or the recorded activity change later, the receipt is marked stale. Earlier receipts stay as history.

What the run showed

The replay above is a production run on November 12, 2026, in a reserved test workspace with one AI employee. A script played the person's clicks in a real browser.

  • The request asked for a saved Markdown checklist for a fictional bicycle repair shop, with exactly four steps in a set order.
  • Nova, the workspace's chief executive, triaged it by rules in 7 seconds: an idea, rank 57 of 100, held for a person to start.
  • After Start, Nova planned, built and checked the work. Each stage ended at a gate with Nova as its recorded owner.
  • The saved checklist was revision 1, 178 bytes, with one heading and the four steps in order.
  • The task was done and its receipt sealed about 1 minute 50 seconds after Send. The receipt said Verified: 3 of 3 recorded checks passed before completion.
  • Nova checked the work itself. Two checks were the platform's: the exact saved revision and the requirement parts. The third was an AI review of the content in Nova's own run. No other reviewer and no person checked it.
  • The download then had the same hash and size, and the same revision was still there after a reload, 2 minutes 1 second after Send.

Limits

  • A test workspace. The run used a reserved production workspace and a scripted person, not an outside customer.
  • The run as a whole did not pass. Three required checks failed: the request preview was slower than its 100 ms target, the first extraction found no kind and no acceptance criteria within its 5-second window, and the record does not bind every requirement to one exact check. Four more required checks did not run. The record is marked partial, and the whole Flow remains open.
  • Self-checked. In this run the worker checked its own work; no independent reviewer took part.
  • A receipt proves the recorded checks. It does not prove that a person approved the work or that a customer accepted it.

Frequently asked questions

What does "Verified" mean on a receipt?

The work saved real output, the platform's completion check passed, every check passed before completion, and every acceptance criterion of the current plan is covered by a passing check: a file check, a platform requirement proof or an AI content review. It does not mean that a person approved the work. Anything less shows as Unverified, with the reason.

Is a receipt a chat message from the AI?

No. The platform builds it from its own records of the output, the checks and the actions. The AI worker does not write it.

What happens to a receipt when the work changes?

The receipt is marked stale, and it stays as history. A new receipt is sealed when the task completes again, for example after a reopen.

  • Native change-proof engine

    One engine finds the change kinds from the diff and the outputs. Each kind needs its own typed proof: a UI recording, a real HTTP exchange, tests, a query, or exact document bytes. Only a person can waive a check, and a waived check shows as not passed.

  • Exact bytes behind every check

    A file check binds the criterion, tenant, item, digest, size and producing action, and the bytes are hashed again before use. Selected command output is kept whole and checked again by identity and hash.

  • Ticket moves and completion reports

    The runtime moves the source ticket at the events that the team's workflow names, never back or sideways. It posts one template report (outcome, evidence, caveats, what needs a person) to each place that asked.

  • Acceptance criteria as executable checks

    At plan time, each acceptance criterion maps to a test, script, query, HTTP check or screenshot check. Verification runs all of them in each round.

  • Worker insights and verified work per token

    Reports built from the call records show how each worker spends time, tokens, context, models and tools. The Performance tab shows verified results beside the tokens they took.

  • Platform proof for each requirement

    Each part of a requirement needs proof that the platform made. A saved document gets an automatic check of its exact revision and bytes. Semantic criteria need a recorded model review.

  • Independent verifier

    The platform checks an AI worker's pull request in its own container, out of the worker's reach, and approves it only when a new test fails without the fix.

    Recorded walkthrough

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.