Flywheel verification pilot

Ten hours, fixed at $1,250. Flywheel on one of your AI workflows, with checks you keep and receipts anyone can recheck.

Zain Dana Harper, sole proprietorKent, Washingtonzaindharper@gmail.com

You have an AI workflow whose answers someone else has to trust: a reviewer, an auditor, a customer or your own team next quarter. In the pilot I set up Flywheel on that workflow and write the checks that decide whether its answers hold. You get back sealed receipts and a report that says what each check shows and what it leaves open. The price stays the same if the work runs long.

What you get

Flywheel set up on your workflow

Flywheel installed and pointed at one workflow you choose, on the model you already use, hosted or local. Runs go through Flywheel's local endpoint, and each step lands on a hash-chained ledger on the machine that ran it.

Proof: what Flywheel does, run records, release 1.9.0

Written checks

Each check compares a value in an answer with the source that decides it, such as a rate table, a spec, a test suite or a script you supply. A check reports which fields agree with the source, which disagree and which nothing could confirm. The check files are yours and run again on every later answer.

Proof: check-output contract, Form 1040 worked example

Receipts a stranger can recheck offline

Every checked result is sealed into a receipt. Anyone you hand it to can recheck the seal with the Python standard library alone, with no account, no network and no call to me.

Proof: receipt and proof formats, recheck manifests, route, receipt, tamper and refusal demo

A findings report

A written report on what the checks found, with every result marked as confirmed, disagreeing or unconfirmed. Each check carries a line on what it does not prove. The report closes with the commands that rerun every check and recheck every receipt.

Proof of the reporting style: Flywheel offline benchmark record, benchmark evidence status

Time and price

Fixed price: $1,250 for a timebox of 10 hours of work.
The price does not rise. If the 10 hours run out before every check is written, you get the report on what was finished, with the unfinished checks listed by name.
The work starts once we both sign a short engagement letter with the scope and the acceptance criteria. Payment terms go in that letter.
No part of the fee depends on what the checks find.

How it works

Scope. You email the workflow, the answers that matter and the decision they feed. I reply with a one-page scope that names the checks I plan to write and what falls outside them, or with a decline and the reason.
Setup. I install Flywheel and connect it to your workflow, either on a copy you send me or in a working session on your machine.
Checks. I write a check for each value that matters, against the source you name as the one that decides it.
Runs and receipts. The workflow runs through Flywheel, every result is checked, and each checked result is sealed into a receipt.
Handover. You receive the check files, the receipts, the findings report and the commands to rerun all of it without me.

What you provide

One workflow you own or hold a licence to test: a script, an agent task, a pipeline step or a repository.
The sources that decide a correct answer for it: tables, specs, tests or reference outputs.
A few sample inputs, with the answers your team expects where you have them.
The model access the workflow already uses. Your keys stay in your environment and never reach me.
One person who can answer questions about the workflow during the 10 hours.

Leave out credentials, customer data and personal identifiers. If the workflow needs them to run, we set it up on your machine and the data stays there.

What this pilot does not prove

A receipt shows that a result has not changed since it was sealed. It cannot show that the result is correct. Flywheel's stated limits
A check is only as good as the source it compares against. If the source is wrong, the check agrees with a wrong answer.
Ten hours covers one workflow and a small set of checks. It cannot cover a whole system.
The pilot does not certify safety, attest compliance or audit security. The work I do not take applies here too.

The record so far

No client has run this pilot yet. There are no case studies, testimonials or third-party reviews of it, and no AI lab uses Flywheel. The evidence you can check today is the tool and its public record: the latest release with its checksums, the documentation, the package on PyPI and the CI history. Any income from a pilot goes on the public income ledger, under the independence policy.

Request a pilot

Email zaindharper@gmail.com with the subject "Flywheel verification pilot". This shape is enough:

Workflow:
What runs, and on which model?

Answers that matter:
Which values in its output
does someone have to trust?

Deciding source:
What table, spec, test or script
says whether a value is right?

Decision:
What will you decide once
the checks are in place?

For a smaller start in public, you can send a public test case. That route is free, takes public material only, and returns a MATCH, DRIFT or UNVERIFIABLE packet.