From RAP sheet to usable case record
A RAP sheet is the California Department of Justice's record of a person's arrests, court cases, and sentences. RAP Scanner reads these dense, inconsistently formatted PDFs and handles the slow, error-prone work an attorney would otherwise do by hand — extracting each conviction, charge, disposition, and sentence into a structured case file that legal aid teams use for record-clearance work.
RAP sheets are printed in a terse, teletype-era format, arrive as scans of variable quality, and layer later court and custody events over earlier ones. RAP Scanner reads them deterministically, using explicit, inspectable rules. The same input produces the same result, and every value can be traced to the line that produced it.
Phases from PDF to structured record
Fields captured per charge before interpretation
Explicit parsing and interpretation rules
Of output values trace back to a source line
RAP Scanner structures the record. The Access Project's Clean Slate Engine uses that structured information to analyze eligibility under California record-clearance laws. An attorney reviews the results before anything is filed.
The pipeline
Four phases
Text extraction
AWS Textract converts the PDF to text and preserves the page structure.
Parse
The parser turns the OCR text into structured rows, one per charge or event, while preserving the wording on the sheet.
Interpret each row
Rules separate statute codes, identify courts and counties, and read custody and probation terms from the sentence text.
Connect the record
Rules link later court actions to earlier convictions, merge duplicate case entries, trace probation history, and identify dismissed counts.
Send to review
The cases populate the client's Clean Slate Engine profile, with a review document and a color-highlighted copy of the RAP sheet.
An optional AI assist can be enabled for a small number of unresolved fields. It fills blank fields only and never replaces a value produced by the rules. See AI & privacy.
The parser
How the parser reads a RAP sheet
The parser works through the OCR text from top to bottom. It recognizes the recurring labels and sections used on California RAP sheets, then copies the relevant text into defined fields. Parsing runs locally after OCR and uses the same ordered rules each time.
Structure
Names and aliases
The subject's legal name and aliases appear at the top of the record. The parser groups them so later entries can be tied to the same person.
Court and custody blocks
Each COURT: or CUSTODY: block records an event cycle with a date, location, and case number. Prison and juvenile custody entries are kept distinct.
Counts and later events
Each CNT: marks a charge. Later dated lines may record a dismissal, reduction, probation revocation, or sentence change. The parser links those events to the count they affect.
Steps
The parser cleans common OCR noise, reads the name and alias lines, splits the record into court and custody blocks, reads each block's date, location, and case number, and then extracts its counts and later events. It keeps convictions, relief, wardships, and limited contextual entries while excluding arrest-only and administrative material. Each retained count or event becomes its own row.
Example
This illustrative court block shows how a conviction moves from OCR text to transcription and then interpretation. It does not depict a real person.
COURT : NAM: : 002 20040907 CASC MCEL CAJON CNT 001 #C243209DV 273.5 (A) PC-INFLICT CORPORAL INJ:SPOUSE DISPO: DISMISSED CNT 002 SEE COMMENT FOR CHARGE TOC:N *DISPO: CONVICTED CONV STATUS : MISDEMEANOR SEN: 003 YEARS PROB, 020 DAYS JAIL, FINE COM: CNT 02 CHRG-242 (A) (1) PC DCN T6011582980437000
group_type COURT
block_type count
case_number C243209DV
date 9/7/04
location CASC MCEL CAJON
count_no 002
section_desc SEE COMMENT FOR CHARGE
disposition CONVICTED
conviction_type MISDEMEANOR
sentence 003 YEARS PROB,
020 DAYS JAIL, FINE
comment CNT 02 CHRG-242
(A) (1) PC
county_1 San Diego courthouse_1 El Cajon conviction_date_iso 2004-09-07 code_1 PC section_1 242 (A)(1) description_1 BATTERY conviction_type_1 MISDEMEANOR custody_period_1 20 days jail probation_granted Yes, 3 years
Count 001 was dismissed, so it is excluded from the conviction output. Count 002 says only "SEE COMMENT FOR CHARGE." Phase 2 preserves that wording and the related comment. Later rules identify 242 (A)(1) PC from the comment, map MCEL CAJON to the El Cajon courthouse, and read three years of probation and 20 days in jail from the sentence. Each change is recorded with its rule and source.
How each field is read
Each field has its own rules, based on patterns found across California RAP sheets. The rules account for common OCR errors while preserving the source text for review.
Dates
Eight-digit date strings are cleaned and converted into standard dates. The original text remains available for comparison.
Statute citations
A charge is separated into its section, code, and description. Joined statutes stay together, and qualifiers such as ATTEMPTED are preserved.
Dispositions
The rules recognize about 50 disposition terms, including common OCR versions of words such as "convicted." These terms determine which counts represent convictions, dismissals, or relief.
Custody and sentence
Sentence text is split into prison, jail, probation, and fine terms, with the duration of each stored separately.
Case numbers
Spaces, dashes, and leading symbols are removed for matching. The rules also account for common OCR substitutions when the same case appears elsewhere.
Courts and counties
The location text is compared with a maintained directory of California courts and common OCR variations, then mapped to a courthouse and county.
Complex records
Handling changes across a record
The most difficult records contain related events separated by years, pages, or OCR errors. The rules connect those entries and preserve the current status of each conviction.
Later court actions
A later entry may revoke probation, change a sentence, dismiss a conviction, or reduce a felony. The scanner links that entry to the affected count and updates the final case history.
Dismissals and reductions
The scanner recognizes set-asides and dismissals under statutes such as Penal Code 1203.4, as well as reductions under Proposition 47, Proposition 64, and Penal Code 17(b). It records the legal basis when the sheet provides enough information to identify it.
Multiple counts and repeated cases
Each count remains separate. When a case appears in more than one block, the scanner combines the entries into the most complete record while preserving count-level details. Sentencing enhancements are separated from the underlying charge.
OCR errors and wrapped text
The scanner repairs recognized OCR errors, rejoins wrapped charge and comment lines, and removes form text that has been attached to a case entry.
Maintained reference tables
Court names, statute information, sentencing enhancements, and wobbler status live in reviewable reference tables. They can be updated as the law and court practices change.
Every row has a permanent ID. Each interpreted value is stored with the rule and source text that produced it. A reviewer can trace a final value from the case file to the rule, the source line, and the page of the RAP sheet.
Integration and outputs
Built into the Clean Slate Engine
RAP Scanner is built into the Clean Slate Engine. A user uploads a RAP sheet, and the extracted cases populate on the client's profile for review and eligibility analysis. The scanner also generates a RAP Summary and a highlighted copy of the record.
Clean Slate Engine case file
Identifiable convictions are loaded onto the client profile for review and analysis.
RAP Summary
A Word document summarizes the client's case history, missing information, extraction warnings, and any entries that need closer review.
Highlighted RAP sheet
The original PDF is returned with color-coded highlights showing convictions, dismissals, reductions, and custody notes. Reviewers can compare the extracted data directly with the source page.
A reviewer can edit any field in the web app.
AI and privacy
Rules first, with optional AI for unresolved fields
Criminal-history records contain sensitive personal information. AWS Textract performs the initial OCR. After that step, the standard parsing and interpretation process runs locally through explicit rules. The optional AI assist is off by default.
Rule-based process
Default- Provides the source of truth for extracted and interpreted values.
- Produces the same output from the same input.
- Runs locally after OCR.
- Stores the source text and rule behind each value.
Optional AI assist
Off by default- Uses Anthropic's Claude only when a rule leaves a field blank or uncertain.
- Fills blank fields and never replaces a rule-derived value.
- Applies a confidence threshold to each limited task.
- Logs each request and response for review.
When the AI assist is enabled, names are replaced with placeholders and case numbers receive temporary labels such as CASE_1. The mapping stays in memory. Dates are reduced or removed when the task does not require them. Statute text and legal terms remain so the system can interpret the entry. The exact content sent with each request is logged for review.
Verification
How accuracy is checked
Hand-labeled benchmark
A reviewer labels a curated set of about 50 RAP sheets against the source PDFs. The scanner's results are compared with those labels, with extra weight given to fields such as the charge and case number.
Regression checks
A fixed set of reference sheets is run again after code changes. Unexpected differences are flagged for review. When a rule change is intentional, the expected results are updated and documented.
Automated tests
More than 800 tests cover parsing and post-processing. They run without cloud services or API keys, so the rule-based behavior can be checked independently.
The rules, source links, benchmark, regression checks, automated tests, and attorney review make each result verifiable before use.
Illustrative examples do not depict real people or records. RAP Scanner structures RAP-sheet data. Eligibility is determined separately and reviewed by an attorney before filing.
