Fifteen worked examples
Every span below came from the real library. The page is generated by
site/gen_data.py and site/build.py, and the build fails if a single
recorded offset breaks source[start:end] == text. There is no hand-written result on this
page.
note
Every string is fake. Do not replace one with production text. Examples 10 to 14 need a model lane,
because a name, an address, a national id, and a login name have no grammar a rule can own.
How to read a block
| Field | Meaning |
|---|
| Source | Your string, with each span marked. Hover a mark for its offsets. |
| Result | The span table. ok means the recorded slice equals the recorded text. |
| Score | 1.00 means a checksum confirmed the slice. — means no validator exists for that form. |
| Masked | The same string with placeholders in place of spans. |
| Vault | Placeholder to original slice. A secret never appears here. |
1 A support ticket
Lane rules · 1 × EMAIL, 1 × PAYMENT, 1 × PHONE
Three structured forms, each confirmed by a grammar or a checksum. Order 48291 is not PII, so it stays visible.
Source
Customer Ada Lovelace called from +852 5555 0101. Email ada@example.com. Card 4242424242424242. Order 48291.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PHONE | 34 | 48 | +852 5555 0101 | — | ok |
| 2 | EMAIL | 56 | 71 | ada@example.com | — | ok |
| 3 | PAYMENT | 78 | 94 | 4242424242424242 | 1.00 | ok |
Masked output
Customer Ada Lovelace called from <PHONE_1>. Email <EMAIL_1>. Card <PAYMENT_1>. Order 48291.
Vault
| Placeholder | Value |
|---|
<PHONE_1> | +852 5555 0101 |
<EMAIL_1> | ada@example.com |
<PAYMENT_1> | 4242424242424242 |
2 A note with a password
Lane rules · 1 × EMAIL, 1 × PHONE, 1 × SECRET
A secret is irreversible on purpose. Restoring it would write a password into the next system.
Source
Ask Ada Lovelace <ada@example.com> to call +1 202 555 0147 about ticket 77. Do not include the password hunter2 in the note.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | EMAIL | 18 | 33 | ada@example.com | — | ok |
| 2 | PHONE | 43 | 58 | +1 202 555 0147 | 1.00 | ok |
| 3 | SECRET | 104 | 111 | hunter2 | — | ok |
Masked output
Ask Ada Lovelace <<EMAIL_1>> to call <PHONE_1> about ticket 77. Do not include the password <SECRET> in the note.
Vault
| Placeholder | Value |
|---|
<EMAIL_1> | ada@example.com |
<PHONE_1> | +1 202 555 0147 |
3 A wire transfer
Lane rules · 1 × EMAIL, 1 × IP_ADDRESS, 1 × PAYMENT
The IBAN passes the MOD-97 check. 4242424242424243 fails Luhn, so the rule refuses it rather than guessing.
Source
Wire to IBAN GB82WEST12345698765432 from card 4242424242424243. Reach ops at ops@example.org or 203.0.113.10.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PAYMENT | 13 | 35 | GB82WEST12345698765432 | 1.00 | ok |
| 2 | EMAIL | 77 | 92 | ops@example.org | — | ok |
| 3 | IP_ADDRESS | 96 | 108 | 203.0.113.10 | 1.00 | ok |
Masked output
Wire to IBAN <PAYMENT_1> from card 4242424242424243. Reach ops at <EMAIL_1> or <IP_ADDRESS_1>.
Vault
| Placeholder | Value |
|---|
<PAYMENT_1> | GB82WEST12345698765432 |
<EMAIL_1> | ops@example.org |
<IP_ADDRESS_1> | 203.0.113.10 |
4 A Chinese contact line
Lane rules · 1 × EMAIL, 1 × PHONE
The rule lane is language neutral for structured forms. The phone still passes libphonenumber.
Source
联系人张伟，手机13800138000，邮箱wei@example.com。
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PHONE | 8 | 19 | 13800138000 | 1.00 | ok |
| 2 | EMAIL | 22 | 37 | wei@example.com | — | ok |
Masked output
联系人张伟，手机<PHONE_1>，邮箱<EMAIL_1>。
Vault
| Placeholder | Value |
|---|
<PHONE_1> | 13800138000 |
<EMAIL_1> | wei@example.com |
5 Two people, two placeholders
Lane rules · 3 × EMAIL
The same address twice gets <EMAIL_1> twice. A second address gets <EMAIL_2>.
Source
Write to ada@example.com or grace@example.com. The desk address is ada@example.com.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | EMAIL | 9 | 24 | ada@example.com | — | ok |
| 2 | EMAIL | 28 | 45 | grace@example.com | — | ok |
| 3 | EMAIL | 67 | 82 | ada@example.com | — | ok |
Masked output
Write to <EMAIL_1> or <EMAIL_2>. The desk address is <EMAIL_1>.
Vault
| Placeholder | Value |
|---|
<EMAIL_1> | ada@example.com |
<EMAIL_2> | grace@example.com |
6 Text with no PII
Lane rules · no span
These look numeric, and none is a published PII form. A locator that spans them is guessing.
Source
Order 48291 ships as SKU-77219 from warehouse 4. The build finished in 12 minutes on node 3.
Result
No span. Nothing in this text matched a grammar on this lane, and that is the correct answer. A locator that invents a span here would be useless.
Masked output
Order 48291 ships as SKU-77219 from warehouse 4. The build finished in 12 minutes on node 3.
Vault
The vault is empty. A secret never enters it, and this text holds none.
7 Long text, original offsets
Lane rules · 1 × EMAIL
A long document is split into windows internally. The public offset is never a window offset.
Source
Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Email ada@example.com for the desk.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | EMAIL | 366 | 381 | ada@example.com | — | ok |
Masked output
Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Note. Email <EMAIL_1> for the desk.
Vault
| Placeholder | Value |
|---|
<EMAIL_1> | ada@example.com |
8 A log line
Lane rules · 1 × EMAIL, 1 × IP_ADDRESS
An IP address passes an RFC 4291 or RFC 3986 match and then a stdlib parse.
Source
2026-11-04T09:12:33Z level=warn user=ada@example.com src=198.51.100.7 path=/v1/orders action=denied
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | EMAIL | 32 | 52 | user=ada@example.com | — | ok |
| 2 | IP_ADDRESS | 57 | 69 | 198.51.100.7 | 1.00 | ok |
Masked output
2026-11-04T09:12:33Z level=warn <EMAIL_1> src=<IP_ADDRESS_1> path=/v1/orders action=denied
Vault
| Placeholder | Value |
|---|
<EMAIL_1> | user=ada@example.com |
<IP_ADDRESS_1> | 198.51.100.7 |
9 A prompt before it leaves
Lane en-closed · 1 × EMAIL, 2 × PERSON, 1 × SECRET
This is the product case: keep the sentence, remove the person, prove which characters left.
Source
Summarise this for the on-call engineer: Ada Lovelace reported that ada@example.com cannot reset the password hunter2 from Lisbon.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PERSON | 41 | 44 | Ada | 1.00 | ok |
| 2 | PERSON | 45 | 53 | Lovelace | 1.00 | ok |
| 3 | EMAIL | 68 | 83 | ada@example.com | — | ok |
| 4 | SECRET | 110 | 117 | hunter2 | — | ok |
Masked output
Summarise this for the on-call engineer: <PERSON_1> <PERSON_2> reported that <EMAIL_1> cannot reset the password <SECRET> from Lisbon.
Vault
| Placeholder | Value |
|---|
<PERSON_1> | Ada |
<PERSON_2> | Lovelace |
<EMAIL_1> | ada@example.com |
10 Two names
Lane en-closed · 4 × PERSON
A name has no grammar and no checksum. Rules own structured forms, models own names.
Source
Ada Lovelace met Grace Hopper at the lab to review the schedule.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PERSON | 0 | 3 | Ada | 1.00 | ok |
| 2 | PERSON | 4 | 12 | Lovelace | 1.00 | ok |
| 3 | PERSON | 17 | 22 | Grace | 1.00 | ok |
| 4 | PERSON | 23 | 29 | Hopper | 1.00 | ok |
Masked output
<PERSON_1> <PERSON_2> met <PERSON_3> <PERSON_4> at the lab to review the schedule.
Vault
| Placeholder | Value |
|---|
<PERSON_1> | Ada |
<PERSON_2> | Lovelace |
<PERSON_3> | Grace |
<PERSON_4> | Hopper |
11 A postal address
Lane en-closed · 2 × ADDRESS, 1 × EMAIL, 2 × PERSON
An address is a model type. The email beside it is a rule type. The locator merges both.
Source
Ship the replacement to Grace Hopper, 12 Rue de Rivoli, Paris. Confirm on grace@example.com before Friday.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PERSON | 24 | 29 | Grace | 1.00 | ok |
| 2 | PERSON | 30 | 36 | Hopper | 1.00 | ok |
| 3 | ADDRESS | 38 | 54 | 12 Rue de Rivoli | 1.00 | ok |
| 4 | ADDRESS | 56 | 61 | Paris | 0.87 | ok |
| 5 | EMAIL | 74 | 91 | grace@example.com | — | ok |
Masked output
Ship the replacement to <PERSON_1> <PERSON_2>, <ADDRESS_1>, <ADDRESS_2>. Confirm on <EMAIL_1> before Friday.
Vault
| Placeholder | Value |
|---|
<PERSON_1> | Grace |
<PERSON_2> | Hopper |
<ADDRESS_1> | 12 Rue de Rivoli |
<ADDRESS_2> | Paris |
<EMAIL_1> | grace@example.com |
12 A national id
Lane en-closed · 1 × GOV_ID, 2 × PERSON
V1 ships no national id pattern, by design. This row shows what the model lane buys.
Source
Applicant Ada Lovelace, SSN 121-45-6789, signed the form today.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PERSON | 10 | 13 | Ada | 1.00 | ok |
| 2 | PERSON | 14 | 22 | Lovelace | 1.00 | ok |
| 3 | GOV_ID | 28 | 39 | 121-45-6789 | 0.90 | ok |
Masked output
Applicant <PERSON_1> <PERSON_2>, SSN <GOV_ID_1>, signed the form today.
Vault
| Placeholder | Value |
|---|
<PERSON_1> | Ada |
<PERSON_2> | Lovelace |
<GOV_ID_1> | 121-45-6789 |
13 An account handle
Lane en-closed · 1 × EMAIL, 1 × USERNAME
USERNAME is a model type. It is the string that marks an account, not the person.
Source
User lkernan signed in and asked to reset access for lkernan@example.com.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | USERNAME | 5 | 12 | lkernan | 0.99 | ok |
| 2 | EMAIL | 53 | 72 | lkernan@example.com | — | ok |
Masked output
User <USERNAME_1> signed in and asked to reset access for <EMAIL_1>.
Vault
| Placeholder | Value |
|---|
<USERNAME_1> | lkernan |
<EMAIL_1> | lkernan@example.com |
14 A finance note
Lane en-closed · 1 × EMAIL, 1 × PAYMENT, 2 × PERSON
The hardest case in the public score. A finance document does not look like a persona note.
Source
The transfer from Grace Hopper uses card 5490 3479 1287 6543 and routing number 271210785. Receipt to grace@example.com.
Result
| # | Type | Start | End | Text | Score | Slice |
|---|
| 1 | PERSON | 18 | 23 | Grace | 1.00 | ok |
| 2 | PERSON | 24 | 30 | Hopper | 1.00 | ok |
| 3 | PAYMENT | 41 | 60 | 5490 3479 1287 6543 | 1.00 | ok |
| 4 | EMAIL | 102 | 119 | grace@example.com | — | ok |
Masked output
The transfer from <PERSON_1> <PERSON_2> uses card <PAYMENT_1> and routing number 271210785. Receipt to <EMAIL_1>.
Vault
| Placeholder | Value |
|---|
<PERSON_1> | Grace |
<PERSON_2> | Hopper |
<PAYMENT_1> | 5490 3479 1287 6543 |
<EMAIL_1> | grace@example.com |
What these examples do not prove
An example shows that one string works. It does not measure accuracy. A library can pass every
string on this page and still score badly on a new domain, and that is exactly what the
public score found: the closed encoder's person score falls from 0.995
in domain to 0.048 on finance text.
Use these pages to understand the shape of a result. Use the
evaluation guide to decide whether the library is good enough for
your text.