PILOTSix weeks, one change, on your own dispute history.
Bring one change to the software that decides your card disputes: a new rule, a prompt edit, a model swap. Scruple backtests it on your resolved cases, scores it in shadow on live traffic, and gates it in your CI. You sign the success criteria before any of your data is loaded.
For teams that already decide disputes with software and want to change it safely: a fintech or marketplace running its own rules or an agent on Stripe or Adyen, a processor whose operations team maintains representment rules, or a company that sells a dispute agent and has to show each release is controlled.
Status: a draft offer, November 2026. No pilot has run with a partner yet. Every number elsewhere on this site comes from the synthetic simulator or from public CFPB complaints. A pilot is how scruple gets its first number from a real dispute operation.
pilot card disputes, one workflow, one candidate change
beforetermsNDA and terms, planner inputs, where it runs
week 1kickoffcriteria signed, nothing loaded before
week 2exportingest report, rejected rows and why
week 3backtestchanged decisions per reason code
week 4shadowyour system decides, scruple acts on nothing
week 5gatescorecards on real pull requests
week 6readoutpass, fail or not enough data
reported whether the result passes or fails
deleted your data and keys, confirmed in writing
the pilot, as drafted in the engine's pilot kitdocs/pilot/SCOPE.md
01 / WHAT YOU GETSix things, each from a tool that exists today.
The gate, the CI integration, evidence packs, the audit chain, Stripe, Adyen and CSV ingest, the backtest and shadow mode exist and are tested, on Stripe's published fixtures and on synthetic histories. None has yet run on a partner's data.
Still to build or agree during the pilot: your own policy constraints as checks, the import mapping if you export CSV, and anything the scope lists as not built. Nothing planned is promised as finished until it is merged with tests.
- A backtest of one candidate change on your history.Which decisions change, in which reason codes, and which hard policy constraints it would break: contesting after the evidence deadline, auto-deciding a high-value case, accepting a duplicate. Reported per reason code, never only as an average.
- A shadow run on live traffic.Your production system keeps deciding. Scruple scores what the candidate would have done and acts on nothing.
- A gate in your CI.A GitHub Action or GitLab template that blocks a pull request when the candidate regresses, and posts a scorecard. Synthetic held-out cases and your own history are shown separately, each with its own verdict.
- A self-report check.What your agent or dashboard says it wins, against what the balance transactions say happened.
- An evidence pack per decision.JSON and PDF, with the audit chain proof, that a risk or compliance reviewer can read without running anything.
- A written readout.Results against the criteria you agreed before seeing any data. Reported whether they pass or fail.
02 / SIX WEEKSCriteria first, data second, readout last.
Each week has one output you can hold the pilot to. A failed criterion is not a reason to stop: it is a result, and it goes in the readout.
Outcomes lag. Decisions shadowed in week 4 mostly resolve after week 6, so the readout says which cost results are still open. A cost improvement claim is optional and sized up front; six weeks of traffic is often not enough to detect a small one, and we will say so before you start.
| Week | What happens | Output |
|---|---|---|
| 1 | Kickoff. Success criteria signed. Your current rules and the candidate translated into scruple's rule language. Export agreed. | Signed criteria, export plan |
| 2 | You run the export. Ingest validates every row, redacts, tokenises card numbers and encrypts chosen fields before anything is stored. | Ingest report: rows read, stored, rejected and why |
| 3 | The candidate and your current rules replayed on your resolved history. Planted regressions gated on the same history. | Backtest report per reason code |
| 4 | Your system keeps deciding; scruple scores the candidate alongside it and writes nothing. | Shadow summary |
| 5 | The gate installed on one repository you choose, as a required status check. | Scorecards on real pull requests |
| 6 | Readout with your reviewer. Your data and keys deleted. | Readout, evidence packs, deletion confirmation |
| Role | Over six weeks | Busiest |
|---|---|---|
| Owner | 16 hours | Week 1, signing the criteria |
| Engineer | 17 hours | Weeks 2 and 5, export and CI |
| Dispute expert | 8 hours | Week 1, current rules |
| Reviewer | 4 hours | Week 6, the readout |
03 / WHAT YOU PROVIDEA few thousand resolved disputes and one change.
Read-only data, a named owner, and an engineer for two short stretches. Scruple runs inside your environment, or on a scruple-operated machine with only tokenised, redacted data.
Pricing is set per pilot, agreed before kickoff.
- A named owner.One person who signs the criteria and the readout, about 16 hours over the six weeks.
- An engineer.For the export and the CI steps, mostly in weeks 2 and 5.
- A read-only export.Resolved disputes and their balance movements, through the Stripe API, Adyen notifications or CSV, for the window the planner asks for.
- One candidate change.A new rule, a prompt edit or a model swap, and the rule or agent version it would replace.
- A reviewer.From risk, compliance or the dispute team, for the readout.
- Criteria agreed before data arrives.Four kinds of hypothesis: the gate catches regressions planted in your history; the candidate adds no hard policy violations; the self-report gap is measured from the ledger; the scorecard reproduces byte for byte in your CI.
| Workflows other than card disputes: ACH returns, wires, reconciliation, AML, payables. |
| Writing to Stripe, Adyen or any network. The Stripe client only reads. |
| Acting on live traffic. Shadow mode acts on nothing. |
| Replaying an LLM agent on history. Python agents with a decide function can already be backtested. |
| SSO, role-based access, a web console. Training any model on your data. Legal or network-rule advice. |
04 / CONTACTTell us about your disputes.
The form writes an email in your own mail app, addressed to the founder. Nothing is sent from this page and nothing is stored.
Or write directly to dhegde@umich.edu.
Monthly dispute volume sizes the pilot: the planner uses it to say whether six weeks of history and traffic can reach the sample size your criteria need.
- Scruple
- A draft pilot offer. No partner data has been used.
- The one-page PDF
- Back to the home page