Method & data handling
You don’t need to believe anything about us. Every engagement is structured so the result can be verified mechanically — before you pay and after.
| Intake screen | Encoding, delimiter, header hygiene and regulated/credential content are screened before any processing. Regulated data (SSN, PHI, payment cards, legal case files) is refused — across the entire file, not a sample of it. |
|---|---|
| Count identity | Input = clean + merged-away + exceptions. Every input row is accounted for exactly once — proven by a raw-row closure check, not by trusting the worker's own counts. ±0 tolerance, or the job does not ship. |
| Duplicate integrity | Zero duplicate keys survive in the clean file under the declared rule — verified by an independent recount, not by the worker. |
| Exception honesty | Conflicts and ambiguous matches are flagged with reason codes and their full original rows preserved — never silently “resolved”. What we refuse to guess is listed for your decision. |
| Recoverability | Every merged-away row carries its full original. Every edit is in the change log with before → after. The revert path is real. |
| Injection guard | Cells that could weaponize a spreadsheet (=, @, executable +/− payloads) are neutralized and logged. Phone numbers and dates are left alone. |
| Stranger test | Every package ships with a handoff README: counts, file inventory, how to audit any single change, how to revert. Written for someone who has never spoken to us. |
These checks run as scripts, plus a separate independent recount that does not reuse the worker's code. Acceptance criteria are written down before work starts and never softened afterwards.
outputs/ ├── clean_contacts.csv # import-ready dataset ├── changelog.csv # every change: row, field, before, after, action ├── exceptions.csv # rows needing your judgment + reason codes └── README.md # counts, rules applied, how to re-run or revert
A sample package on fully synthetic data is available on request — it shows the exact formats, including how conflicts and exceptions are surfaced rather than hidden.
| Access | Sandbox, staging or file-based exports only. We do not request production write access under any circumstances. |
|---|---|
| Retention | Your files are deleted on your request — deletion is verified against the disk and confirmed in writing at handoff by default. Raw input archived read-only during the job, then deleted. |
| Form transport | Website contact/sample forms are delivered to our mailbox via Web3Forms (web3forms.com), which processes only the form fields you submit. Data files are never sent through the form. |
| Tooling | Processing happens in our isolated working environment with local tooling. No third-party service touches your data without prior disclosure. |
| Review | Quality review may include an independent frontier-model pass over reason-coded exception summaries and redacted metadata. Column values stay local unless you approve otherwise in writing — we disclose before any such use. |
| Confidentiality | NDA-friendly: we review and sign your NDA before any file moves. One named operator with least-privilege access — every touch of your data is attributable to a single accountable person. |
| Incidents | Suspected data incident → work stops immediately and you are notified the same day. No exceptions, no quiet patches. |
| Declines | PHI, payment-card and legal-case data are declined outright. |