Errors are planted, with an answer key
We change quotations, briefs and chronology entries on purpose, record what we changed, and count how many the tool catches. We also count true items it wrongly refuses.
How we test Working prototype, in preview
Every number on this site, where it came from, and what it leaves out.
MoonView is a working prototype in preview, not a released product. The figures here are results from our own test runs, each with its date. None is a customer result. Where a test found a miss, the miss is listed beside the hit.
Method
How a claim earns its place here.
Each round uses files the tool has not seen
A fix that only fits the last test file proves little. The quote checker's rounds each used a new set of transcripts.
No client data
Tests use synthetic files and deposition transcripts that are public, most of them filed in court. Nothing from those public transcripts is shown on this site, because they name real people.
Simulated reviewers are a test method
Some rounds are run by simulated skeptical reviewers who play a litigator, a paralegal or a nurse consultant and try to break the tool. They are not customers and nothing they said is quoted here.
Misses are reported with hits
Each table below is followed by what the test missed or could not show.
A passing test is not a review
These numbers describe the checks. In use, a person still decides every point before anything is exported.
1 Quote and cite checker
The final round: quotations and briefs with planted defects.
On 2 Nov 2026 we ran the last of thirteen adversarial rounds. Each round used a fresh set of public court-filed deposition transcripts, with quotations altered on purpose and errors planted in briefs. Ten earlier rounds had each found between 2 and 10 defective quotations shown green, almost all of them word-perfect quotations whose meaning changed through what they left out. That is why green now means the whole answer only.
| Measure | Result | Date | Notes |
|---|---|---|---|
| Defective quotations shown green, final round | 0 of 63 | 2 Nov 2026 | Round 13, on transcripts new to the tool. |
| Errors planted in briefs that were caught, final round | 28 of 29 | 2 Nov 2026 | One planted error was missed. |
| Defective quotations shown green, the round before | 0 of 63 | 2 Nov 2026 | Round 12. Round 11, the first after green was narrowed to the whole answer, had 0 of 57. |
| New transcripts in the final round that loaded with no help | 2 of 7 | 2 Nov 2026 | The rest were refused or needed a working copy. |
| The same seven, read again after loading fixes | 3 of 7 | 10 Nov 2026 | 3 load on their own. 3 condensed files are read through the one-page-per-sheet copy. 1 needs the line-number copy. |
- Missed: 1 of 29 planted brief errors.
- Slow: building the checked copy of a condensed transcript took 5 to 10 minutes on some files.
- Not measured here: case law, exhibits, and whether an argument is sound. The checker does not look at them.
2 Deposition summary
Paid model runs on public transcripts.
On 30 Oct 2026 we ran public deposition transcripts through a commercial model (Claude Opus 5.5), twice. The first run covered seven files. After one fix to the instructions, the second run covered six of them again. Every kept point was then evaluated for its citation and its speaker.
| Measure | Result | Date | Notes |
|---|---|---|---|
| Kept points with a wrong citation, second run | 0 of 334 | 30 Oct 2026 | Six public filings. |
| Kept points that put words in the wrong mouth, second run | 0 of 334 | 30 Oct 2026 | |
| Kept points still needing a reviewer's attention, second run | 6 of 334 | 30 Oct 2026 | |
| Share of proposed points that code kept, second run | 27% | 30 Oct 2026 | The rest were refused. |
| First run: kept points with a wrong citation | 0 of 231 | 30 Oct 2026 | Seven public transcripts. Code kept 17% of 1,384 proposed points. |
| Meaning-changing edits accepted from a reviewer | 0 of 697 | 27 Oct 2026 | Probe edits written to change a point's meaning, on a synthetic transcript. The same run accepted 108 of 153 honest edits. |
| Meaning-changing probes accepted as a model's point | 2 of 574 | 28 Oct 2026 | Both drop a hedge the answer gives in a later clause. 14 more take words from the question and wait for the reviewer to confirm that context. Synthetic transcript. |
- Missed: 2 of 574 meaning-changing probes were accepted as a model's point.
- Thin: code kept about a quarter of what the model proposed. Some of what it refused was honest wording, such as
goes
forgo
. - Not repeatable to the word: two identical runs differ from each other. The model's output is saved with each review.
- Not tested on client transcripts: only public transcripts and synthetic files have been sent to a model service.
- The demo's summary is one saved run of 10 Nov 2026 on the invented deposition, with 12 points.
3 Medical chronology
Forty-two synthetic cases: checked, delivered and purged.
On 10 Nov 2026 we re-ran the checks, delivery and purge on 42 synthetic cases (OCR text for 37 of them was kept from the 7 Nov run), 2,824 pages and 1,136 visits. This run used no model. Each visit was given the entry a perfect drafter would write, quoted from the answer key, and then the same entries with errors planted. It measures the checker and the pipeline, not a model's drafting.
Synthetic records only. No real client record has been through the chronology tool.
| Measure | Result | Date | Notes |
|---|---|---|---|
| Planted errors the checker refused | 5,368 of 5,368 | 10 Nov 2026 | Eight kinds: wrong date, another provider, a cut finding, an added No, another matter's page, an invented cost, another visit's page, a visit type the provider contradicts. |
| True entries that passed the checker | 1,128 of 1,136 | 10 Nov 2026 | 8 refused. All trace to OCR reading ]as lin a medication list. |
| Visit pages given the right provider and date | 2,511 of 2,511 | 10 Nov 2026 | On our own synthetic layouts. |
| Bill pages found and matched to their visit | 132 of 132 | 10 Nov 2026 | None matched falsely. |
| Lost visits flagged as possibly missed | 15 of 15 | 10 Nov 2026 | |
| Duplicate pages found | 48 of 48 | 10 Nov 2026 | In one pair the copy came first in the file, so the later page was named the repeat. |
| Cases delivered, read back and purged clean | 42 of 42 | 10 Nov 2026 | |
| Blurred test pages marked low for a person to check | 120 of 120 | 7 Nov 2026 | Synthetic pages at the heavier blur strengths. At the lightest strength tested, 0 of 30 were marked. |
| One small run with a model | 11 entries, 11 visits | 10 Nov 2026 | The 28 synthetic pages of the demo matter. |
| Real 1963 hospital pages: handwritten pages marked low | 10 of 10 | 10 Nov 2026 | 22 public-domain pages. No typed page was marked low. |
| Real 1963 hospital pages: provider suggestion wrong | 8 of 13 | 7 Nov 2026 | Of the 13 pages that name a provider. |
- Synthetic: every record in the large test was generated. The sorting rules fit those layouts.
- No model in the large test: it does not show how often a model drafts a wrong entry, only that the checker refuses the kinds of error we planted.
- Handwriting: OCR cannot read it. The model cannot cite a handwritten page until a person types its text.
- Real pages: On the 9 typed real pages that carry a date, the date was read right on 7. The 22 real pages were split into 20 documents where 13 was right.
- Light blur: a slightly blurred page can be misread without a low mark.
- Intake: a matter is still created and read in by script, not from a screen.
4 Reading transcripts
The same 24 public transcripts, before and after.
Everything depends on reading the page and line numbers correctly. The set is 24 publicly available full-size deposition transcripts: 21 from court dockets and 3 published by U.S. House committees. We read them at the start of the project and again after the fixes.
| Measure | Result | Date | Notes |
|---|---|---|---|
| Transcripts read with no wrong page or line | 23 of 24 | 30 Oct 2026 | The other one was refused, not misread. |
| The same set at the start | 8 of 24 | 13 Oct 2026 | 3 more were read with wrong page numbers and 13 were refused. A court filing stamp had been taken for the printed page number. |
- Refused: 1 of 24 could not be read at all.
- Layouts beyond these: condensed, plain-text and image-only transcripts are read through checked working copies. See the loading results under the quote checker.
5 The demo matter
The invented case you can try.
The demo is one invented plaintiff, Quinn Mockridge: a 5-page deposition, a brief with five quotations, four of them carrying a planted error, and 28 pages of synthetic medical records. On 10 Nov 2026 we walked it through every tool in a real browser.
| Measure | Result | Date | Notes |
|---|---|---|---|
| Demo quotations given the planned verdict | 5 of 5 | 10 Nov 2026 | Exact, dropped qualifier, wrong page, counsel's question, words never said. |
| Pages of the deposition read as testimony | 3 of 5 | 10 Nov 2026 | Pages 2 to 4. The caption page and the certificate are listed as left out. |
| Chronology entries the model gave for the demo records | 11 for 11 visits | 10 Nov 2026 | 3 further pages were accounted for: 2 fax covers and 1 page from another patient, which the screen flagged. |
| Delivered files read back against approved entries | 3 of 3 | 10 Nov 2026 | Word, PDF and Excel. |
- Quotation marks change the label: Typed without quotation marks, words that are not in the transcript are labelled Not a quote (amber). In quotation marks, or in a brief, the same words are Not found (red).
- A misfiled page is flagged, but has no label of its own to choose.
Open gaps
What is not done, in one list.
- It is a prototype. The Windows build is an unsigned preview. Windows SmartScreen warns before it runs.
- Some transcript layouts still need a person to start a working copy or confirm how pages read, and condensed copies are slow to build.
- The chronology has never seen a real client record. Real handwritten pages cannot be cited by the model, and real forms confuse the provider suggestion.
- The summary is conservative. About three quarters of what a model proposes is refused.
- Client transcripts are not sent to outside model services. Summary development uses synthetic and public transcripts only.
- The reporter tools are a frozen prototype. The preview program still lists it under Other tools; it is not being developed.
- None of this is legal advice. MoonView does not check case law, give strategy or judge credibility.
Your files
What leaves your computer.
The answer is different for each tool, so here it is for each one.
- No AI Quote checker
- Nothing. The program asks the download site once at start whether a newer version exists. It sends its version and build number, and nothing about your files.
- Model Summary
- The witness's answers, the question each responds to and your issue list go to the model service whose key you set; never the PDF, page or line numbers. The summary is switched off in the preview build.
- Reviewed service Chronology
- Today the chronology is run as a reviewed service: a firm sends records, and a reviewer approves every entry beside its page. OCR runs on the machine that holds the records. To draft entries, the text of each visit's pages is sent to an outside model service; for client records that step stays off until a business associate agreement is signed. So far only synthetic records have been through it.
Test it yourself.
The demo runs the real quote checker in your browser. Type a quotation that is nearly right and see what it says.