Competitive bake-off
Cover vs
Season 2, same 18 PDFs: Cover caught 100% of 149 private details, 0 leaked, 12 over-redacted, 3.2 s per PDF. Alibaba Qwen 3 caught 75%, 50 leaked, 188 over-redacted, 4.5 min per PDF. Of 12 engines ranked by mistakes (leaked private details weigh most, and over-redactions count too), Alibaba Qwen 3 places 4th and Cover 1st.
Sending PDFs with private identity information to a cloud Ai like Claude or ChatGPT is a bad idea. But could a local Ai on your own computer redact them instead? This bake-off runs Cover’s engine and the leading local Ai engines on the same 18 bank, tax and medical PDFs. So far the local Ai engines fall behind Cover in three ways: they over-redact, most are slow, and none offer a fast way to fix their mistakes.
Seasons
Season 2 · 18 PDFs · 149 private details (PII)
PDF selection focuses on money, home and insurance documents. Real-world samples for anyone asking Ai about their own money, or accountants and tax pros who redact a client's documents before asking Ai for a second opinion. Watch Cover's engine improve with each release, even outpacing Google's tech.
- Money 9
- Home 5
- Health & insurance 3
- Taxes 1
Engine quality over time
Local redaction landscape
| # | engine | PII accuracy | Over-redaction | Speed |
|---|---|---|---|---|
| 1 | 🥇 Cover | Excellent | Very good | Excellent |
| 2 | 🥈 Gemma 4Google | Very good | Very poor | Good |
| 3 | 🥉 Phi-4Microsoft | Good | Very poor | Poor |
| 4 | Qwen 3Alibaba | Good | Very poor | Very poor |
| 5 | Qwen 2.5Alibaba | Poor | Very poor | Good |
| 6 | DeepSeek R1DeepSeek | Poor | Very poor | Very poor |
| 7 | Llama 3.1Meta | Poor | Very poor | Good |
| 8 | Apple IntelligenceApple | Very poor | Very poor | Excellent |
| 9 | Gemma 3Google | Poor | Very poor | Good |
| 10 | Privacy FilterOpenAI | Poor | Very poor | Very good |
| 11 | RampartAmerica.gov | Very poor | Very poor | Excellent |
| 12 | NemotronOpenMed | Very poor | Very poor | Excellent |
Per-page side-by-sides
Check out the actual pages below and how each engine redacted them, side by side with the original.
Red outline: private detail left visibleBlue outline: over-redaction
4th place
Worth flagging
Alibaba Qwen 3
A reasoning model that spends most of its time thinking: about four and a half minutes per PDF, the slowest here, and it still catches only 76% of the PII.A reasoning model that spends most of its time thinking rather than detecting. It is the slowest engine in the field at about four and a half minutes per PDF, just ahead of DeepSeek, and the thinking does not buy much: it catches 76% of the PII to Cover's 100%, over-redacts about thirteen times more (188 against 14), and takes roughly ninety times longer per document.
Cover’s edge
Holds even if a model matches Cover’s accuracy.
- One-click edits fix any box instantly; no re-prompting (LLMs over-redact).
- Fully offline files never leave your device.
- Seconds per file no datacenter hardware to buy or maintain.
2025_Tax_Return_Okafor-Brennan.pdf
Page 1Cover wins
Page 2Cover wins
Page 3Cover wins
Page 4Cover wins
Page 5Tie
Page 7Cover wins
Page 8Tie
WellsFargo-sample-statement.pdf
Page 1Tie
Page 2Qwen 3 wins
capitalone-sample-estatement.pdf
Page 1Cover wins
commercebank-sample-statement.pdf
Page 1Cover wins
Fidelity-sample-statement.pdf
Page 1Cover wins
Page 2Tie
Page 3Cover wins
Page 4Tie
Page 5Cover wins
Page 6Cover wins
Page 7Tie
Page 8Tie
Page 9Cover wins
Page 10Cover wins
Page 11Cover wins
Page 12Cover wins
Page 13Tie
Page 14Tie
Page 15Cover wins
Page 16Cover wins
Page 17Tie
Page 18Cover wins
Page 19Tie
Page 20Cover wins
Page 21Cover wins
Page 22Tie
Page 23Tie
Page 24Tie
Page 25Tie
Page 28Qwen 3 wins
fidelity-large-print-sample-statement.pdf
Page 1Tie
Page 3Tie
pershing-2024-sample-tax-yearend.pdf
Page 1Cover wins
Page 2Cover wins
Page 3Cover wins
Page 4Cover wins
Page 5Tie
Page 6Cover wins
Page 7Cover wins
Page 8Cover wins
ubs-portfolio-summary-sample.pdf
Page 1Cover wins
Page 2Cover wins
Page 3Cover wins
Page 4Cover wins
Page 5Cover wins
Page 6Cover wins
Page 7Cover wins
Page 8Cover wins
Page 9Cover wins
Page 10Cover wins
Page 11Cover wins
Page 12Cover wins
Page 13Cover wins
Page 14Cover wins
Page 15Cover wins
Page 16Cover wins
Page 17Cover wins
Page 18Cover wins
Page 19Cover wins
Page 20Cover wins
Page 21Cover wins
Page 22Cover wins
cfpb-sample-credit-card-statement-handout.pdf
Page 1Tie
allstate-auto-declarations-sample.pdf
Page 1Cover wins
Page 2Cover wins
cfpb-ficus-bank-loan-estimate-sample.pdf
Page 1Cover wins
Page 2Tie
Page 3Tie
cigna-eob-example.pdf
Page 2Cover wins
medicare-sample-part-b-msn.pdf
Page 2Cover wins
Page 3Cover wins
Page 4Cover wins
Page 5Cover wins
Page 6Tie
cfpb-draft-periodic-mortgage-statement.pdf
Page 1Cover wins
Lease_Agreement_signed.pdf
Page 1Cover wins
transunion-cameo-credit-report-sample.pdf
Page 1Cover wins
Page 2Cover wins
Page 3Cover wins
Page 4Tie
Page 5Cover wins
Page 6Cover wins
Page 8Cover wins
Page 9Cover wins
Page 10Cover wins
att-uverse-sample-bill.pdf
Page 1Tie
Page 2Tie
Page 3Cover wins
Page 4Cover wins
pge-sample-bill-sparky-joule.pdf
Page 1Qwen 3 wins
Page 2Tie
Page 3Tie
Page 4Cover wins
Page 5Tie
Page 6Tie
Full results
Cover v1.0.2 vs Alibaba Qwen 3 qwen3:latest · 18 real-world sample documents.
| metric · all 18 docs | Cover | Qwen 3 |
|---|---|---|
| Recall: PII correctly redacted | 100% | 75% |
| Leaks: confirmed PII left visible | 0 | 50 |
| Over-redaction: non-PII redacted by mistake | 12 | 188 |
| Speed: median time per PDF | ~3.2 sec/PDF | ~4.5 min/PDF |
Per-document breakdown: all 18 docs (one outlier can dominate the totals)
| document | Cover leaks | Qwen 3 leaks | Cover over | Qwen 3 over |
|---|---|---|---|---|
| ubs-portfolio-summary-sample | 0 | 4 | 0 | 90 |
| transunion-cameo-credit-report-sample | 0 | 18 | 0 | 35 |
| Fidelity-sample-statement | 0 | 3 | 2 | 22 |
| pershing-2024-sample-tax-yearend | 0 | 0 | 0 | 18 |
| 2025_Tax_Return_Okafor-Brennan | 0 | 1 | 0 | 14 |
| commercebank-sample-statement | 0 | 0 | 3 | 4 |
| allstate-auto-declarations-sample | 0 | 0 | 0 | 3 |
| att-uverse-sample-bill | 0 | 5 | 0 | 1 |
| medicare-sample-part-b-msn | 0 | 1 | 4 | 1 |
| Lease_Agreement_signed | 0 | 9 | 0 | 0 |
| WellsFargo-sample-statement | 0 | 0 | 2 | 0 |
| capitalone-sample-estatement | 0 | 2 | 0 | 0 |
| cfpb-draft-periodic-mortgage-statement | 0 | 1 | 0 | 0 |
| cfpb-ficus-bank-loan-estimate-sample | 0 | 4 | 0 | 0 |
| cfpb-sample-credit-card-statement-handout | 0 | 0 | 0 | 0 |
| cigna-eob-example | 0 | 1 | 0 | 0 |
| fidelity-large-print-sample-statement | 0 | 0 | 0 | 0 |
| pge-sample-bill-sparky-joule | 0 | 1 | 1 | 0 |
| TOTAL | 0 | 50 | 12 | 188 |
How the test works
Test set: 18 public sample documents (149 verified PII items), held fixed across runs so results stay comparable. Every engine redacted the same documents; each output is scored against a Claude-authored gold standard (confirmed + ambiguous tiers), and every counted over-redaction is kept auditable. Pages and the leaderboard use the same weighting: leaked private details weigh most, and over-redactions count too.