RePath Labs is the research and development team at eHawk, Inc. We test new technology on real community corrections work, publish what we find, and build what holds up into RePath, so agencies can get better outcomes while spending less time and money.
The idea in 60 seconds
Five short scenes on the problem our first three studies set out to solve, and what they found.
1 · The gap
The best evidence says what happens inside a visit is the strongest lever an agency has on outcomes. In our experience, a trained auditor can carefully score four to eight visits a day, so many agencies can review only a few of every 100 by hand.
2 · The catch
Three experienced auditors scored the same 30 visits. On average, two experts' scores for one visit landed 19 to 33 points apart on a 100-point scale. Feedback that depends on who grades it is hard to coach from.
3 · Standard set, standard met
First, three experts adjudicated ten visits item by item and agreed on 429 of 430 rubric decisions: an answer key the field rarely has. Graded against it, the RePath AI workflow matched 90.21% of decisions. The experts' own grades from before the panel met matched 50% to 86%.
4 · Coaching
An expert reviewer rated 120 coaching notes without knowing who wrote them. The top AI coach averaged 3.58 of 5, against 3.30 for the higher-rated human coach, and wrote the most notes ready to show an officer as written.
5 · Every visit
Caseloads of officers trained in core correctional practices recidivated at 36.2% vs. 49.9% for comparison officers, and those skills fade without feedback. Our studies point to a way to give that feedback on every recorded visit. Field pilots will test whether it changes outcomes.
Figures from RePath Labs Papers 1 to 3 (2026) and Chadwick, DeWolf & Serin (2015). Review rates are illustrative. Coaching examples are invented for illustration, not study text.
Why Labs exists
AI models now read, reason and write well enough to help with work that once needed a trained expert reading every visit. That opens real opportunities in community corrections, and real risk of tools that promise more than they deliver. Labs finds out which is which on real supervision work, before anything reaches an agency.
Most of an officer's week goes to documentation and system work, not people. Every hour technology gives back can go to the conversation that changes outcomes.
8 to 10 hoursa week back per officer across RePath agencies
Decades of research show that officers who use core correctional practices with fidelity get better outcomes, and that the skills fade without ongoing feedback.
36.2% vs. 49.9%recidivism, trained vs. comparison officers' caseloads (Chadwick et al., 2015)
When supervision fails, the bill goes to jails, prisons, courts, victims and taxpayers. Better practice inside the visit is the cheapest place to prevent it.
$10 billiona year spent locking people up for supervision violations (Council of State Governments)
Our north star
Visits spent on what drives change: goals, problem solving and the next step, not paperwork.
Lower recidivismThe outcome the training research links to officer skillsSpecific, evidence-linked coaching on their real visits, plus hours back from documentation.
8 to 10 hours a weekReturned per officer across RePath agenciesCapacity without new hires, and practice quality leaders can see across every officer and office.
765 officer hoursReturned in one state DOC's 15-week, three-office pilotEvery revocation avoided can mean a jail or prison bed and a court hearing that never happen.
Nearly 1 in 4 state prison bedsFilled by supervision violations (CSG Justice Center)What we bring to every study
Labs is not a lab in the abstract. It sits inside a company that supports 125+ agencies and has spent a decade learning how supervision really runs, visit by visit.
More than 10 million data points and more than 10,000 participants and office visits, combined with our own direct research and our research library.
Alexandria
Alexandria is RePath Labs' compiled knowledge base of community corrections research, standards, and agency policies and procedures, with profiles of each jurisdiction's supervision system and the assessment instruments agencies use. Labs projects start there, so we build on what the field already knows instead of rediscovering it.
It also lets us answer practical questions fast: how a state structures graduated responses, what a policy requires of a visit, or which instruments an agency already uses.
How Labs works
Every idea runs the same path. The bar to reach an agency is evidence, not a demo.
Alexandria, the published literature and national standards set the question and the yardstick.
Fast builds on synthetic or de-identified data, kept separate from the production roadmap.
Qualified practitioners grade the same work. We set targets in advance where we can, and say so when we did not.
Working papers with DOIs, methods and every limitation stated, free to read.
What survives moves into RePath through eHawk's full product development process.
Assistive by design. RePath gives officers and supervisors better information. People make the decisions. Visit grading and coaching are built for skill development, never for hiring, discipline or other personnel decisions.
The RePath Labs Evidence Series
The training research is clear that officer skills drive outcomes and that the skills need feedback to last. The missing piece has been feedback at scale. Our first three studies test that missing piece, link by link, on real supervision visits. The fourth is the first study we know of to compare decision models, a new class of AI, with large language models against an expert-agreed standard.
Paper 1 · Measure
Yes, within the range of three expert auditors.
Paper 2 · Coach
Blinded review rated it as highly as human coaching.
Paper 3 · Standardize
429 of 430 decisions agreed; AI matched 90.21%.
Paper 4 · Compare
80 to 83% agreement once adapted, at a fraction of the cost.
Next
Improving officer coaching, improving assessments, and much more.
Tested in RePath Labs studiesNext in the lab
Paper 1 · Measure
Published · DOIThree experienced auditors and an AI grader scored the same 30 recorded visits against a 43-item rubric built from national standards and core correctional practices. The experts disagreed with one another more than most people expect. The AI landed inside their range.
How far apart two experts' scores landed on the same visit
Average gap between each pair of auditors, 100-point weighted visit score, 30 visits (one auditor scored 25)
23 of 30 visits (77%) passed the test: the AI's score was no farther from the three-expert average than the median expert's own score.
Paper 2 · Coach
Published · DOIA score alone does not change practice. Feedback does. An external expert reviewer rated 120 coaching notes on the same 30 visits, four per visit, without knowing which were written by expert coaches and which by AI.
In RePath today: Officer Coach in RePath Intelligence is built on this work. See Officer Coach →
Blinded expert rating of coaching quality
Mean composite of eight 1 to 5 dimensions, 30 visits per source, 95% interval
Paper 3 · Standardize
Published · DOIPaper 1 could only show the AI landed inside the experts' spread. So the three experts met, reviewed all 43 rubric items on ten visits together, and agreed on 429 of 430 decisions. That gave the field something it rarely has: an answer key.
Agreement with the expert-adjudicated standard
Share of 429 rubric decisions matched, ten development visits
Taken together
What changes for every 100 recorded supervision visits when grading and coaching no longer depend on how many visits a supervisor can review by hand.
reviewed by hand (illustrative)
every eligible recorded visit
experts' grades matched the agreed standard 50 to 86%
agreement with the expert-agreed standard (ten-visit development study)
reviewed visits coached by a supervisor (illustrative)
coaching rated as highly as human coaching in a blinded 30-visit study
The "Today" columns are illustrative; many agencies review one visit per officer per quarter or year. "With RePath" shows capacity. Quality evidence comes from RePath Labs Papers 1 to 3 (2026).
In RePath today. Visit quality scores and Officer Coach in RePath Intelligence are built on this research. See visit quality scores → See Officer Coach →
“Continuous automated scoring is, in effect, a permanent booster session.”
Beyond large language models
In September 2026 a new class of AI arrived: decision models, which answer each question by choosing from a fixed list of allowed answers instead of writing text. Within four weeks one leaderboard tracked more than 170 of them. Paper 4 is the first study we know of to test decision models against an expert-agreed standard in a professional field.
Paper 4 · Compare
Submitted for publicationWe gave three decision models, TypeSafe Jev, OpenAI Decisions and Microsoft Decision-1, the same ten visits and 429 expert judgments behind Paper 3. First we dropped them into our LLM-built grading system unchanged. Then we adapted the task to each model, without changing the rubric or the answer key.
Agreement with the same expert standard
Share of 429 expert judgments matched, ten development visits
Frontier LLM workflowDecision model, task adaptedHuman experts
| Property | Frontier LLM | Decision model |
|---|---|---|
| What it returns | Open-ended text, images or code | One answer from a fixed list, usually with a probability |
| Can it make things up? | Yes: quotes, citations, reasons | No, only allowed answers, though it can still return one experts would not choose |
| Best suited to | Holistic judgment and nuanced evidence | Bounded questions with concrete, observable evidence |
“Systems should be built around the strengths of the tools they use and tested against a human reference to choose the right tool for each task.”
Questions
RePath Labs is the research and development team at eHawk, Inc., the company that builds RePath, a smartphone-based supervision platform used by 125+ probation, parole, pretrial, treatment court and juvenile agencies. Labs tests new technology, especially AI, on real community corrections work, publishes the results, and moves what holds up into RePath. Ted Green directs RePath Labs.
Not yet. They are working papers (preprints) posted with permanent DOIs so they can be read, cited and challenged now. Each discloses its methods, data handling, conflicts of interest and limitations. A DOI makes a paper permanent and citable; it is not peer review, and we do not describe it as one.
No. Visit grading and coaching are built to help officers develop skills and to help supervisors coach. The papers state plainly that scores are not intended for hiring, firing, promotion, discipline or any other personnel decision. Officers and supervisors make every decision about a case.
While the well-accepted literature documents a strong relationship between the quality of visits and reduced recidivism, we need more field data over time to establish that relationship for AI coaching. That research also shows officers' skills fade without feedback. In our studies, on small samples, AI grading tracked expert auditors and a blinded expert rated AI coaching notes as highly as expert coaches' notes. Whether that feedback, delivered on every visit, changes officer practice and lowers recidivism is what our field pilots will measure.
Alexandria is RePath Labs' compiled library of community corrections research, standards, and agency policies and procedures, with profiles of 55 jurisdictions and the assessment instruments agencies use: more than 5,000 sources in all. It is where Labs projects start.
We are recruiting agencies for field pilots that measure officer skill use over time, then completions and recidivism against a comparison group. Researchers can request methods detail or discuss collaboration. Email Ted Green at tedgreen@ehawksolutions.com.
Partner with RePath Labs
We are recruiting agency partners for field pilots that track officer skill use over time, then completions and recidivism against a comparison group. Partners shape the study, see results first, and help the field answer its most important open question.
Ted Green
Director, RePath Labs · Co-founder and COO, eHawk, Inc.
tedgreen@ehawksolutions.com