Skip to content

The model never holds the facts

Inside the operating system is the engine: the deterministic layer that holds the record and your jurisdiction’s rules. Agents put language around it. They never invent what is in it.

Most AI in government is asked to be careful. This one is built so the mistake has nowhere to come from.

A general assistant generates its answer, which is why it can produce a hearing date that never existed. On eCourt AI a date is not generated. It is computed by the engine from the rule that governs it, or read from the record, and returned with its source attached. The model is in the path for two things only: understanding what a person asked, and putting the answer into their language at a reading level they can use.

How an answer is produced on eCourt AI. A question arrives from a person, in their own words. The model reads it and works out what is being asked. That is the only thing the model is asked to do on the way in. The question then goes to the engine, the deterministic layer holding the record and your jurisdiction rules, which computes or retrieves the answer along with the rule or record it came from. The model then puts that answer into plain words in the reader language. The model never sources a fact and is never asked to remember one, so a fabricated date has nowhere in this path to originate. The answer leaves carrying its source.

This is the answer to a question your IT team will ask in the first ten minutes: what happens when the model is wrong about a fact. It cannot be, because it was never asked.

What the engine governs

Jurisdiction policy is built in, not configured on arrival by your staff. Each domain below is held per jurisdiction, and each one publishes what agents may do with it and what no agent does with it, ever. In the domains where a mistake would cost most, the restraint is the point.

Codes, statutes, and ordinances

The codes your jurisdiction enforces, with their text, classifications, effective dates, and amendment history, so a citation always resolves to the version that applied on the day in question.

What agents may do

  • Retrieve the provision that applies and quote it with its citation
  • Explain a provision in plain words, in the reader’s language
  • Flag when a record cites a provision that was not in force on that date

What no agent does

  • I never paraphrase a provision as though it were the provision. The text and its citation travel together.
  • I never give legal advice about what a provision means for someone’s case.

Violations and charges

Violation and charge definitions as your jurisdiction codes them: elements, classifications, degrees, and the schedules that attach to each.

What agents may do

  • Look up how a charge already on the record is classified
  • Prepare filings and notices that carry the charge exactly as recorded
  • Flag a mismatch between a charge on the record and its statutory classification

What no agent does

  • I never suggest what to charge. Charging is a prosecutor’s discretion and it is not mine to inform.
  • I never infer a charge from a narrative or a set of facts.

Sentencing and disposition rules

Statutory ranges, mandatory minimums and maximums, enhancements, credits, and the disposition and program options your jurisdiction allows, each traceable to its provision.

What agents may do

  • Present the statutory range that applies, with its source, to the bench and to counsel
  • Prepare a worksheet a judicial officer reviews, changes, and signs
  • Flag when a disposition already entered falls outside the statutory range

What no agent does

  • I never recommend a sentence, and I never rank the options.
  • I never score a person, and I never predict what a person will do.
  • I never apply the law to a person’s facts to reach a conclusion about that person. That is a judicial act and it belongs to a judge.
  • Nothing I prepare here ever reaches a member of the public.

Deadlines and timelines

Every clock your jurisdiction runs: filing and response periods, service and notice windows, speedy trial and statutory limits, tolling, and the court holidays and counting method that decide where a date actually lands.

What agents may do

  • Compute a date from the rule and show the rule it came from
  • Flag a deadline that is approaching or has passed, to the person responsible for it
  • Recompute every affected date when one event moves, and route the change for approval

What no agent does

  • I never move a date on my own. I compute it, and a person decides.
  • I never waive, extend, or toll anything. Those are decisions.

Fines, fees, and payment rules

Fine and fee schedules, surcharges and distributions, payment plan terms, hardship and waiver criteria, and the indigency standards your jurisdiction applies.

What agents may do

  • Compute a balance and itemise it against the schedule that produced it
  • Offer the payment plans and hardship processes your agency has enabled
  • Reconcile every transaction to the record

What no agent does

  • I never decide a waiver, a reduction, or a finding of indigency.
  • I never threaten, and I never imply a consequence the record does not support.

Calendars and scheduling rules

Session structures, courtroom and judicial availability, case-type durations, hearing sequencing requirements, interpreter and accommodation needs, and the conflict rules that govern them.

What agents may do

  • Find and flag conflicts a proposed date would create
  • Prepare a calendar change for approval, with the downstream effects listed
  • Hold the accommodations a case requires so a setting cannot quietly drop them

What no agent does

  • I surface conflicts. I do not resolve them.
  • I never set, continue, or vacate a hearing without the approval your policies require.

Notice and service requirements

Who must be told, of what, by when, and by which method, including the form and content requirements your jurisdiction imposes and the proof of service each method needs.

What agents may do

  • Determine which notices an event requires and prepare each one from the record
  • Produce them in the recipient’s language, at a plain reading level, in the required form
  • Track service and flag anything unserved before it becomes a defect

What no agent does

  • I draft only from the record. I never invent a fact to complete a notice.
  • I never certify service. A person does that.

Records, retention, and confidentiality

Sealing, expungement, and confidentiality rules, juvenile and victim protections, redaction requirements, public access categories, and retention and destruction schedules.

What agents may do

  • Apply the access category a record carries before anything is shown or sent
  • Redact to your jurisdiction’s rule and record what was redacted and why
  • Flag records reaching a retention or sealing milestone

What no agent does

  • I never release a record whose access category does not permit it, to anyone, including staff.
  • I never decide a sealing or expungement petition.

Communication and consent rules

Consent status per person and per channel, contact-time restrictions, frequency limits, opt-out handling, and the language and accessibility requirements that attach to each recipient.

What agents may do

  • Check consent and channel eligibility before every message, without exception
  • Honour an opt-out immediately and everywhere
  • Reach a person in their language, on the channel they chose

What no agent does

  • I never contact anyone the rules say I may not, and no setting in the system can override that.
  • I never treat silence as consent.

Your policy, written by you, enforced by the engine

Every domain arrives configured to your jurisdiction. From there your agency customises what ships or authors its own policy outright, and the engine enforces it the same way it enforces the rules that came with it. One policy layer governs both sides of the house: what your staff can do, and what a person outside the agency sees, is told, and is offered.

Customise what ships

Start from your jurisdiction’s rules as configured, then change what your agency does differently. Local practice stays local.

Author your own

Write policy the engine did not ship with, in your agency’s own terms, and it governs from the moment you turn it on.

One layer, both sides

The same policy governs staff-facing processes and client-facing experiences, so what a clerk may do and what a person is told can never drift apart.

Changes are on the record

Every policy change is attributed, timestamped, reviewable, and exportable, the same as any other action in the system.

Rules are dated, because cases are

Codes change, schedules are amended, and a case is governed by the rule that applied on the day in question. The engine holds effective dates and amendment history, so an answer about a two-year-old matter uses the two-year-old rule. A system that only knows today’s rules is wrong about every case that is not new.

Every answer carries its source

A computed date shows the rule that produced it. A balance itemises against the schedule behind it. A quoted provision travels with its citation. This is what makes an agent’s work reviewable rather than merely plausible, and it is why a wrong answer is visible instead of silent.

Bring a rule you think is hard

A local counting method, an amendment nobody implemented, a fee schedule with three exceptions. The engine either holds it or it does not, and we will tell you which on the first call.

Not sure where to start?Get an AI readiness plan