Policy Central
An open-source agent for writing and reviewing policy documents. Ask it to check a manual, adapt or revise a policy, or draft a new one. It finds clauses that conflict with or repeat your other policies, and the clauses a change affects. Each finding quotes both clauses, and each change waits for your approval.
Orlando PD’s body camera policy checked against CFA standard 32.02M: one wording error, quoted from both documents, fixed as a tracked change.
Works with: Word documents · PDFs, including scans · web pages · web search.
The beta is by invitationEach person gets their own link; nothing to install and no account. Write to support@policinglab.org for one.
Try it now
Opening to invited reviewers first
Who it is for
Ten people, ten checks. Pick one to see their situation, what they would type, and what comes back. The documents named are public.
1 of 10
Policy unit lead: the law changes overnight
- SITUATION
- In September 2025 a Florida appeals court struck down the state’s open-carry ban (F.S. 790.053), and open carry became legal statewide within weeks. The court’s opinion and the police chiefs’ association’s summary of the Attorney General’s guidance, against a department’s policies on stops, weapons and courthouse security.
- THEY ASK
- “Open carry became legal in Florida in September 2025. Which of our policies need updating, and what should they say?”
- THEY GET
- A plain answer: no policy depends on open carry being illegal. One line worth adding, to the courthouse weapons rule, naming the places guns stay banned (F.S. 790.06(12)), tracked in Word.
2 of 10
Accreditation manager: before an accreditation visit
- SITUATION
- The assessors are coming. The Florida accreditation standards (CFA) against the agency’s policies.
- THEY ASK
- “Which standards don’t our policies cover?”
- THEY GET
- Every standard marked covered, partly or not found, with quotes, and a crosswalk to download.
3 of 10
Special events planner: a big event is coming
- SITUATION
- The 2028 Olympics in Los Angeles, a Super Bowl or Mardi Gras. The Justice Department’s planning checklist for large events against the fire department’s special-event rules and the police department’s crowd policy, each in its own vocabulary.
- THEY ASK
- “Our big event is 90 days away. What does our plan not cover yet from the planning checklist, and what should we sort out first?”
- THEY GET
- A run-up list: the checklist items neither plan covers yet, most urgent first, with each plan’s own words quoted.
4 of 10
Emergency manager: before hurricane season
- SITUATION
- Hurricane season starts on 1 June. Florida’s Emergency Management Act (ch. 252) and the state law-enforcement annex against Orlando PD’s emergency operations plan.
- THEY ASK
- “Does our hurricane plan cover what state law expects of us?”
- THEY GET
- Each duty marked covered or not found, with the law and the plan quoted.
5 of 10
Legal advisor: after a court settlement
- SITUATION
- A settlement sets what the policies must say. The Minneapolis settlement agreement (2023) against the police manual’s use-of-force chapter.
- THEY ASK
- “Which paragraphs on crowd control and force don’t our policies meet?”
- THEY GET
- Every paragraph marked covered, partly or not found, with quotes, and a crosswalk to download.
6 of 10
Chief: the first 100 days
- SITUATION
- A new chief wants to know what the manual really says. From a real manual: children 12 and under are not handcuffed, against a rule that every prisoner is.
- THEY ASK
- “Where does our manual contradict itself?”
- THEY GET
- Each conflict, with both clauses side by side.
7 of 10
Neighboring agencies: a pursuit crosses the line
- SITUATION
- Officers from two agencies end up in the same pursuit. Orlando PD’s and the Orange County Sheriff’s Office’s vehicle pursuit policies (both 2026).
- THEY ASK
- “Where would an Orlando officer and a deputy act differently?”
- THEY GET
- The differences that matter, with both policies quoted.
8 of 10
Training lead: after a policy change
- SITUATION
- The policy changed; the lesson plans may not have. Sacramento PD’s use-of-force training against the final policy.
- THEY ASK
- “Where does our training still teach the old rule?”
- THEY GET
- Each outdated passage, with the current rule beside it.
9 of 10
Oversight board or monitor: a monitor’s report
- SITUATION
- A report says what should change. The NOPD monitor’s final report (2025) against current policies.
- THEY ASK
- “Which findings do our policies address?”
- THEY GET
- Each finding marked addressed or not found, with the policy quoted. One table both sides can read.
10 of 10
HR director: beyond policing Illustrative
- SITUATION
- State employment laws change too. A city personnel manual against Florida employment laws: domestic violence leave, E-Verify, minors’ hours.
- THEY ASK
- “What in our manual no longer matches the law?”
- THEY GET
- Tracked changes, with the law quoted.
Free during the beta
Free during the beta. Your workspace is temporary and is deleted after an hour without use.
Privacy
- Kept
- What you type, your accept and reject choices, your feedback, the organisation and role you give if you choose to, and for uploads the title, the page count and the findings’ one-line summaries.
- Not kept
- Document text, full quotes, your address. There are no accounts. Uploads are deleted with the workspace, after an hour without use.
The app runs in The Policing Lab’s own AWS account; pairs of clauses go to the ranking model through OpenRouter. Don’t upload confidential drafts.
Open source
The code will be published under the AGPL-3.0 licence. Run it on your own machine with your own model keys, or read how it works.
Research
- Policy Central: Detecting Conflicts, Redundancy, and Similarity in Deployed Police Policy (Holly Baker, Michael Kuo and Kyana Pollas, 2026): the first Policy Central system, piloted with a police department, and the benchmark used to test it.
Who made it
Policy Central is made by The Policing Lab, a nonprofit in Miami that works with police departments on data and policy.
How well does it work?
General AI assistants are good for quick questions about a document. Policy Central is built for one job, reviewing policy: it reads the whole manual, checks policies against laws and standards, quotes both sides of every finding, and drafts the fix as a tracked change.
These are early results from a small test; a larger, independent evaluation is under way. In November 2026 we planted 10 contradictions in each of two public police manuals and asked each tool, on its default settings, to find where the manual contradicts itself.
1,596-page manual
| Tool | Found of 10 | False findings | Real issues we did not plant | Time | Cost per run | How much it read |
|---|---|---|---|---|---|---|
| Policy Central | 7 | 0 | 4 | 6 min | about $1.80 | every page screened; the likeliest conflicts read closely |
| AI coding assistant | 0 to 1 | 1 to 2 | 0 | 2 min | about $0.35 | opened under 1% of the lines; keyword searches for the rest |
| AI office assistant | 0 | 4 | 0 | 1 min | monthly licence | about 4%, by its own account |
455-page manual
| Tool | Found of 10 | False findings | Real issues we did not plant | Time | Cost per run | How much it read |
|---|---|---|---|---|---|---|
| Policy Central | 6 | 0 | 3 | 3 min | about $1.60 | every page screened; the likeliest conflicts read closely |
| AI coding assistant | 6 | 3 | 4 | 7 min | about $2.30 | every page |
| AI office assistant | 1 | 4 | 0 | under 1 min | monthly licence | not stated |
How we tested
- The manuals
- Two public US police manuals: the Los Angeles Police Department’s Department Manual (1,596 pages) and the Fort Worth Police Department’s General Orders (455 pages), as plain text.
- What was planted
- In each manual we altered 10 clauses so that each one contradicts another clause in the same manual: a duty reversed (“shall” to “shall not”), a time limit or an age changed, a definition turned around. Each tool was asked to find where the manual contradicts itself, with nothing said about planted clauses.
- How a finding counted
- “Found” is a strict count: a planted contradiction counted only when one finding quoted both the altered clause and the clause it was planted against. A script checked every quote word for word against the manual. A finding was false when the two clauses it quoted do not conflict when read in context. Findings we could not settle either way are left out of both counts.
- Real issues we did not plant
- Contradictions already in the manuals, which some tools found as well. One reviewer checked each one by hand against the manual.
- Runs and settings
- Each tool was run once or twice, on its default settings, and we wrote the planted contradictions ourselves. With stronger settings on the 1,596-page manual, the coding assistant found 1 in 7 minutes for about $2.90, having opened about 4% of the manual, and the office assistant found none in about 3 minutes, in each of two runs. Each also caught one planted contradiction against a different passage from the one we planted it against, which the strict count leaves out. Only default settings were run on the 455-page manual.
- Cost
- Model charges for one run, at list price. The office assistant’s cost is part of a monthly licence, so there is no per-run cost.
How we measured how much each tool read
- Policy Central. It breaks every manual into its sections and rules, and screens every one of them. It then reads closely the places most likely to conflict. Its record lists every comparison it made, and which ones it read closely.
- AI coding assistant. We ran it with its detailed activity log switched on. The log records every action it takes: each time it opens part of the file (with the exact line range) and each keyword search (with the words searched and the lines that came back). We added up the line ranges it opened. On the 1,596-page manual that came to about 300 to 400 of 44,497 lines with default settings (under 1%), and 1,681 lines with stronger settings (about 4%). Its searches also showed it matching lines from elsewhere in the file, but it never read the manual through.
- AI office assistant. We can’t see inside it, so after its answer we asked it how much of the manual it had read. With stronger settings it replied that it had read lines 0 to 799 of 22,248, about 3.6%, and had run an automated comparison over the rest; with default settings it said it had opened the first 800 lines. These figures are its own account.
- Why this matters. No keyword marks a planted contradiction: each is an ordinary-looking clause that disagrees with another one far away. A tool that searches rather than reads tends to miss them.
Feedback and contact
There is a feedback button inside the app. For anything else, write to support@policinglab.org.