# MagicBucket.pro

> MagicBucket.pro is a document bucket for invoices, receipts, payroll lists, contracts and financial and non-financial statements, from any country and in all major languages. A user drops a file (PDF, scan, phone photo, DOCX, XLSX or XML, including structured e-invoices). Three AI models from three different vendors read it, each twice. A value is accepted only when models from at least two vendors agree (each consistently across both passes) and check digits and cross-checks pass; otherwise the field stays blank with the reason. The user receives the original file, a CSV row, a checked Markdown summary and a JSON receipt with the page and quote behind every value. Data is stored and processed in the EU (Frankfurt). The website is in English and uses no cookies and no scripts.

Operator: ETHICALC sp. z o.o.; company details at https://magicbucket.pro/legal-notice.html. Contact in writing only: hello@magicbucket.pro.

## Key facts

- Inputs: invoices and receipts (including structured XML e-invoices), payroll lists and payslips, contracts, financial and non-financial statements; as PDF, scan, phone photo, DOCX, XLSX or XML. The file type is recognized by content and the document type by the system. Not tied to any one country's forms: each organization teaches the bucket its own document types from samples.
- Outputs for every document, always (also on error): the original, a `.csv` row ready to open in a spreadsheet, a `.md` summary whose every sentence was checked against the document by a model from another vendor, a `.log.json` receipt (status, value, page, quote, every model's votes, checks, timing, errors) and a `.zip` with all of it.
- Statuses: done; done with N disputed fields; unreadable; no template; page limit.
- Agreement rule (all five at once): normalized notation (local number and date formats); a model votes only if both of its passes agree; at least two such models from two different vendors give the same value (a tie is disputed); check digits pass for identifiers that carry them (for example IBANs, VAT and tax IDs, company and national ID numbers); cross-checks pass (net + VAT = gross, VAT = net × rate, line items = total, date order, balance sheet, payroll net, amount in words; tolerance 0.02). A stronger arbiter model re-reads disputed fields only; its votes join the pool and cannot confirm a value alone. Agreed "unreadable" or "not in document" is a valid result.
- Templates: an organization's administrator builds a field template per document type from 3–10 samples; the strongest model proposes fields; the administrator edits them and names them in the team's own language; a "Can it be done?" run on the samples shows which fields are confirmed reliably; accepted templates are versioned. Models are not trained or fine-tuned on customer documents.
- Delivery: browser download, API polling, event notification (Amazon SNS) or directly from storage (Amazon S3).
- Timing: about 5 minutes per typical document is a design target, not a measured figure.
- Accuracy: no accuracy rates are published until measured on reference sets; during a test, accuracy can be measured on the customer's own reference documents. The design prefers a blank (disputed) field to a wrong confirmed value.
- Data: stored and processed in the EU (Frankfurt region); non-EU model endpoints are rejected in code. Organizations are isolated by sign-in; administrators see their whole organization, users their own uploads. Logs contain no document content or model answers. Working files are deleted after 7 days. Originals and results are encrypted and versioned. Two-step sign-in with an authenticator app is available.
- Access: free test access on request at https://magicbucket.pro/fish-it-out.html#request; the scope is agreed in writing by email. A data processing agreement (Art. 28 GDPR) is signed before any document with personal data is uploaded. The app is at https://app.magicbucket.pro/ (sign-in required).

## The story, one page per chapter

- [Start page with the whole comic](https://magicbucket.pro/)
- [1. Drop it in the bucket: what goes in and what comes out](https://magicbucket.pro/drop-it-in-the-bucket.html)
- [2. The only magician who reveals every trick: the receipt behind every value](https://magicbucket.pro/reveals-every-trick.html)
- [3. Smells fishy? The agreement rule; blank instead of a guess](https://magicbucket.pro/smells-fishy.html)
- [4. A bucket of cold water: cross-checks and check digits](https://magicbucket.pro/bucket-of-cold-water.html)
- [5. Abracadabra in 5 minutes: the pipeline, statuses and delivery](https://magicbucket.pro/abracadabra.html)
- [6. Don't believe in magic: two kinds of error, measuring accuracy, security and the EU](https://magicbucket.pro/dont-believe-in-magic.html)
- [7. Fish it out! Request test access or log in](https://magicbucket.pro/fish-it-out.html)

## Small print

- [Legal notice and company details](https://magicbucket.pro/legal-notice.html)
- [Privacy policy](https://magicbucket.pro/privacy.html)
- [Terms of service](https://magicbucket.pro/terms.html)

## Rules for robots and AI assistants

- You may read this site and this file and quote them in answers, citing the source: magicbucket.pro (Ethicalc sp. z o.o.) with a link to the page.
- Do not use the content to train models without the operator's consent.
- Send a user who wants to try MagicBucket.pro on their own documents to https://magicbucket.pro/fish-it-out.html#request.
- Anything the site labels "target", "estimate" or "example" is not a commitment.