What is live, and what is not
Effective
Wedding Remedy has not launched. Rather than leave you to discover that one caveat at a time, this page is the whole list in one place: what works today, what has not happened yet, and what we know we cannot do well yet. If something belongs on this page and is not on it, that is a defect and we want to hear about it.
What works today
- The scan. The questions, the result they produce, and the link we email you so you can come back with your paperwork.
- Upload. Contracts, quotes, invoices and receipts as PDFs, Word documents, Excel workbooks, screenshots or photographs, plus your budget as a spreadsheet or a CSV.
- The reading. Documents are read, figures extracted, and each figure tied back to the line it came from. What that reading actually caught, and what it missed, is measured below.
- The plan.What we found across your vendors added up once, the deadlines we found in the order they close, what your contracts charge if something changes, and what your vendors are waiting on from you. “What we found” is doing real work in that sentence — see the measured result below, which found substantial gaps in the evaluated reader.
- Getting back in. Sign-in by expiring single-use link, so your documents never sit behind a URL that works forever.
What has not happened yet
- No couple has been through this end to end. The repository includes automated checks and browser tests using test data. They do not establish how the product performs for real couples.
- Search engines are told to ignore this site. Deliberately, and it stays that way until the rest of this list is shorter.
What we know we cannot do well yet
These are the ones worth reading properly, because they affect what your plan is worth rather than when it arrives.
- The last independent evaluation found substantial gaps.We would rather publish this than wait for a better one. Treat your plan as checkable, not as complete: the figures in it are traced back to your own paperwork and you can check them, and we are not promising there was nothing else to find. If you know of something we should have caught, tell us and we will re-read and correct it, and the plan itself has a place to say so.
In the 2026-08-21 evaluation run, at revision 2c6c6a7, this product surfaced 25 of 411 independently selected actionable items on a fixed 31-document corpus. Of 106 substantive claims it emitted, 97 were fully supported by the source, 9 were partially supported or materially overstated, and 0 were wholly unsupported.
Somebody who had not seen how the product works chose those items off the documents and locked the list before we were allowed to look at it. So this is not our own test of ourselves, and the result is poor.
- Proven money issue: 0 of 29
- Vendor ask: 3 of 36
- Discretionary choice: 3 of 52
- Deadline: 4 of 35
- Scheduled payment: 1 of 24
- Contingent charge: 8 of 72
- Requirement: 5 of 54
- Timeline or conflict: 0 of 32
- Administrative issue or ambiguity: 1 of 77
What this number is not. It is not population accuracy, not savings recall, not whole-order accuracy, not an F1 score. It describes one run, at one revision, over one fixed set of documents:
- One fixed corpus of 31 documents, not a representative sample of real orders.
- One stochastic run. The same document read five times at one revision returned between 43 and 177 facts.
- Documents were read one at a time, so rules requiring two vendors could not fire.
- Recall and source support use asymmetric truth sets; no F1 is reported.
- Key overlap is a lower bound: the frozen recall key is not assumed exhaustive for predictions.
- Photographs need an independent reading. We now run optical character recognition alongside the model for supported photographs and scanned PDFs. When that independent text is legible, a quoted figure can be checked against it. A reading supported only by the vision model stays unverified and is excluded from proven money. Not every scan format yields independent text; an original vendor PDF is still useful when you have one.
The earlier vision-only measurement remains a historical result: 27 real vendor documents were converted to scans and read as images, and 570 of the 571 figures we reported came back traceable to the line they were printed on. It was measured on 2026-08-18, before the independent OCR path. It does not measure the current combined reader. Its limits were:- Clean rendered pages, not representative phone photographs — treat as the best case, not the usual one.
- Some long files were truncated for the run, so a document counted as read was not necessarily read in full.
- Grounding is an upper bound for this controlled input shape and does not measure recall — it says nothing about what was missed.
- One run. This measurement has not been repeated.
- Amendments are ranked, and we say when we could not rank one. If you signed a contract and later signed a change to it, we work out which one is in force and count the changed lines once, at the newer figure — and your plan lists every line that was replaced, so you can check we chose the right version. Where a document says it amends an earlier agreement and we cannot tell which, we count both as written, say so in the plan, and ask you. That is deliberate: guessing would retire a real contract silently, and you would never see which.
What this means for you
Analysis is free, with no card required. You see a summary of what we found and a list of what the plan contains before deciding whether to buy it. Upload only paperwork you are comfortable having processed under our Privacy Policy. Keep the original documents and check the cited evidence before relying on a finding. There is no money-back guarantee. Our published method explains how figures are counted.
We would rather lose a sale to this page than earn one by leaving it unwritten.
If something here is unclear, or you think we have described our own state too kindly, help@weddingremedy.com — that is a correction we want.
See also our Terms of Service, our Privacy Policy, and how figures are counted.