For county jails
Every inmate message, screened overnight.
Readwarden checks each message against the official BOP TRULINCS rejection categories. Your staff starts the shift with a short, ranked queue: category, confidence, and the reason it was flagged.
No cost. No contract. Sits on top of your current messaging vendor.
Review queue
Screened this week · 6,184 messages · 22 flagged
- Sensitive identifiers97%
Full card number with expiry, CVV and PIN.
DC-40533 · B-Pod
- Escape95%
Arranges an outside vehicle during a scheduled medical transport and notes escort size.
DC-40502 · Medical
- Threats / extortion94%
Implied harm toward a named person linked to cooperating with police, tied to a court date (possible witness intimidation).
DC-40117 · B-Pod
- Contraband / criminal activity89%
Instructions to soak paper with a substance and mail it under another sender's name (drug-soaked paper).
DC-40177 · B-Pod
Nobody has time to read every message.
Tablets turned mail into thousands of messages a week. Most are family news, court dates and money on books. The few that matter, a witness named, a drop planned, a transport route shared, look the same until someone reads them.
Even the federal system fell behind: a 2020 DOJ Inspector General audit found BOP staff lacked the time to review inmate email. Read the audit
How it works
- 01
We pull your vendor's nightly feed
Each night we pull the day's messages from your messaging vendor: an API, an SFTP drop, or a nightly export. Nothing changes for residents or families.
API, SFTP drop, or nightly CSV export
- 02
Every message is screened
Each message is scored against the BOP TRULINCS rejection categories. Attorney and other privileged contacts on your exclusion list are skipped entirely.
Category · confidence · written reason
- 03
Your staff reviews the queue
By morning shift, flagged messages sit in a ranked queue. An officer reads each one and decides. Nothing is blocked or reported automatically.
Escalate, clear, or export for a case file
Screened against the federal standard
Categories follow the TRULINCS rejection list in BOP Program Statement 4500.13, §14.10(j)(5)(b).
Threats / extortion
Contains threats, extortion, or obscenity.
Violence / group disruption
Depicts, describes, or encourages activities that may lead to the use of physical violence or group disruption.
Weapons
Depicts or describes procedures for the construction or use of weapons, ammunition, bombs, or incendiary devices.
Escape
Depicts, encourages, or describes methods of escape.
Contraband / criminal activity
Encourages, instructs, or may facilitate criminal activity (e.g., introduction of contraband).
Directing a business
Constitutes unauthorized direction of an inmate's business.
Coded language
Is written in, or otherwise contains, a code.
Sexually explicit (security risk)
Constitutes sexually explicit material that poses a threat to safety, security, or orderly operation.
Drug / alcohol manufacture
Depicts or describes procedures for the manufacture of alcoholic beverages or drugs.
Sensitive identifiers
May facilitate criminal activity: SSNs, card and bank numbers, PINs, a third party's ID numbers.
Program violation
Otherwise violates the established parameters of the program (e.g. forwarding, three-way relays, unauthorized contacts).
Your facility rules
Add or drop categories to match your own message policy.
Measured on real inmate messages
Evaluated on real messages written by incarcerated people in federal facilities, sampled from a corpus of ~9.8 million. Today's labels come from two independent AI reviewers; human-labelled validation is in progress. Thresholds favor recall: missing a threat costs more than reading a false flag.
Threats / extortion
- Recall
- 100%
- 16 of 16 (95% CI 80.6 to 100%)
- Precision
- 53%
- 16 of 30
- Threshold
- 0.6
- False-flag rate
- 0.2%
Sensitive identifiers
- Recall
- 93%
- 50 of 54 (95% CI 82.4 to 97.1%)
- Precision
- 74%
- 50 of 68
- Threshold
- 0.3
- False-flag rate
- 0.3%
Contraband / criminal activity
In calibration
Other TRULINCS categories
In calibration
| Category | Recall | Precision | False-flag rate | Threshold |
|---|---|---|---|---|
| Threats / extortion | 100%16 of 16 (95% CI 80.6 to 100%) | 53%16 of 30 | 0.2% | 0.6 |
| Sensitive identifiers | 93%50 of 54 (95% CI 82.4 to 97.1%) | 74%50 of 68 | 0.3% | 0.3 |
| Contraband / criminal activity | In calibrationMeasured during your pilot and published once calibrated. | |||
| Other TRULINCS categories | In calibrationMeasured during your pilot and published once calibrated. | |||
Evaluation set: 1,298 messages (600 uniform-random, 698 keyword-targeted), each labelled by two independent blind AI reviewers against a written rubric, disagreements adjudicated. Human-labelled validation and full-corpus calibration are in progress. Recall is the share of real positives flagged; precision is the share of flags that were real; false-flag rate is measured on the random sample.
CJIS-aligned deployment
The pilot runs on controls modeled on the FBI CJIS Security Policy. Formal CJIS compliance work, including the Security Addendum and personnel screening, begins when an agency signs an agreement.
US-hosted, encrypted
Data stays in US data centers, encrypted in transit and at rest.
Named users, MFA
Role-based access for your staff only. Multi-factor sign-in required.
Full audit log
Every search, view, flag and decision is logged with user and time.
Privileged contacts excluded
Attorney and other privileged contacts on your list are never screened or stored.
No training on your data
Your messages are never used to train models. Written into the pilot agreement.
Your retention rules
Messages are held only as long as your agency's policy says, then deleted.
Questions sheriffs ask
- Does it block or reject messages?
- No. It flags. Your staff reads every flag and makes every decision, under your existing policy.
- Why the BOP categories?
- TRULINCS rejection reasons (BOP Program Statement 4500.13, §14.10(j)(5)(b)) are the most detailed public standard for message review. Your own policy can add or drop categories.
- What about pretrial detainees?
- Most jail residents are unconvicted. That is why decisions stay with staff, privileged contacts are excluded, and every access is audited.
- What happens after 90 days?
- You get the readout and decide. No auto-renewal, no obligation. If you continue, pricing is set per average daily population.
Free 90-day pilot
Judge it on your own traffic.
Run it alongside your current process for 90 days. At the end you get the numbers and decide. No cost and no contract.
You get
- Screening of every message in your vendor's feed for 90 days
- Review queue accounts for your intel unit and supervisors
- Weekly summary: volume, flags by category, what your staff escalated
- A closing readout with measured accuracy on your own traffic, for every category
We need
- Read access to your message feed, or a nightly export
- One point of contact (intel sergeant or jail commander)
- Your privileged-contact list, so we can exclude it
Request a pilot
Tell us about your facility. We reply within two business days to schedule a 30-minute call.
Or email admin@readwarden.com