HiCourt
For your technical & knowledge-management staff
What it does today: research a question, assess a position or a risk, summarize documents, build a chronology, compare two versions, proofread a filing, check the citations in a draft, draft a document and more, working from your matter's files and your firm's permitted past work. It comes back in Word or text, with spreadsheets where the work has rows. A lawyer reviews, verifies and signs the work. Anything not built yet is marked not yet below.
Every authority shown on this page is invented, with fictional parties, a fictional jurisdiction, and reporter series that do not exist, cited to volume 000 at page 000. On a page about fabricated citations, printing a realistic-looking fake case would be the one unforgivable thing to do.
Pick a task, say what you need and add the matter's files. HiCourt drafts the work, looks up the cases it cites, checks them, sends flagged work back to be fixed, and hands your lawyer one finished document with a status on every claim in it. A check of a draft you already wrote comes back as a table and never rewrites your draft.
Written for the people who will be asked whether the firm should allow this in: technology and knowledge-management staff, and the associates who would use it. It assumes you want to check the mechanism rather than take it on trust, and it does not soften the limitations. Where the firm's library lives and what reads it are treated as the two separate questions they are. For the short version, see the summary.
The cases the work cites are looked up in the published case law: a lookup, not a guess. A separate checking stage, which wrote none of the work, then reads the opinion and tests each case against the claim it is cited for. Missing text and uncertain results stay visible. Quotations, pinpoints and whether a case is still good law are not checked yet.
Click it to open it full size. The whole run, end to end.
CLICK TO EXPAND
The problem, in firm terms
An AI tool produces citations that are structurally perfect (plausible parties, a plausible reporter, a plausible court, a plausible year) for cases that were never decided. It will also attach a quotation to a real case that the case does not contain, or describe a holding the court did not reach. The failure is invisible at the point it happens, because the output looks like every correct citation you have ever read.
For a firm the exposure is not that a draft is wrong. Drafts are wrong all the time; that's what review is for. The exposure is that this particular kind of wrong is engineered to survive review. It costs an attorney real time to run down each cite, and it looks correct until someone does.
On naming, and this holds across the whole site. No page here names another vendor, characterises another vendor's architecture, or quotes a figure about one. That is a deliberate standard rather than politeness: we do not run those systems, the products in this category ship changes constantly, and a comparison table written by a competitor is worth exactly what you would expect. So where these pages would otherwise draw a comparison, they do two things instead — state what HiCourt does, in terms of work performed and artifacts produced, and give you the question to put to every vendor you are evaluating, us included. A question cannot be unfair to anyone and does not go out of date. The answers are things you can watch happen in a pilot rather than take from us.
Read this before the demo
This is the part most likely to be glossed over in a pitch, so it goes above the demonstration rather than below it.
HiCourt looks up the cases the work cites in the published case law, as a separate step from any AI assessment. The checking stage is given the claim and passages of the opinion; it does not browse a reporter, a docket or a citator, and nothing it says can establish that a case exists or is still good law.
The checking stage assesses the claim sentence against selected passages of the opinion. It does not reproduce quotations from memory, and it is not a quotation check.
Confidence proves nothing
An AI can state a citation that does not exist with complete confidence, and confirm it if asked. A confident answer is not evidence that an authority exists, and this design does not treat it as evidence.
A lookup proves something
Whether a case exists is settled by looking it up, not by asking an AI. A case that cannot be found is flagged for your lawyer. A lookup still does not show that the case supports the claim; that is what the separate check is for.
Checks cover case citations in full reporter form. Short forms such as Id. and supra, case names on their own, statutes, regulations, rules and treatises are not looked up. Quotations, pinpoints, whether a case is still good law, and the facts in your own documents are not checked yet.
The design
Before your lawyer sees it, the cases are looked up and checked, flagged work goes back to be fixed, and the finished document is checked again.
Stage 0
Pick the task, say what you need and add the matter's files. Each file is read and indexed, and shows as ready, or says why it could not be read, before a run can use it. If a file seems to contradict a choice you made on the form, your choice wins and the run records it.
nothing unread goes in
Stage 1
Ethical walls and matter membership are applied before anything is searched. For research and analysis, the firm's permitted past work is then recalled, each item labelled as reviewed or not and as clean or with known problems. It informs the work; it never replaces checking the law.
only what this person may see
Stage 2
The work is drafted from a close reading of your request, your files and the permitted past work, and it looks for what hurts your position as well as what helps it.
built on your own material
Stage 3
The cases the work cites are looked up in the published case law. A case that cannot be found is flagged, not assumed. A separate checking stage, which wrote none of the work, reads the opinion and tests each case against the claim it is cited for. Missing text and uncertain results stay visible.
a lookup first, then a separate check
Stage 4
Anything flagged goes back to be corrected or withdrawn, and the fix is checked again. The loop is bounded, the run records why it stopped, and anything still flagged stays flagged for your lawyer. A check of a draft you pasted in never rewrites it.
bounded, and the stop is recorded
Stage 5
What survived the checks, and every flag still open, goes into one finished document in the form you asked for, with flags against the sentences they affect. A run that cannot finish keeps what it has, labelled incomplete; nothing is dressed up as done.
flags travel with the text
Stage 6
The finished document's cited claims go back through the checks, and a separate check asks whether it does what you asked. Every sentence gets a status, “not checked” included. Findings are reported to your lawyer; they never rewrite the document.
reported, never rewritten
Stage 7
Every finished run is filed into the firm's memory automatically, and a lawyer's signed review makes a new version. Later work can draw on it only where the person asking may see it.
the library grows with each matter
Authorities are checked, and flagged work repaired, before anything is written up, so the finished document is built from work that has already been tested.
Some citations checked early may be dropped later. That is recorded work, not a measured claim about savings or speed.
Every flag stays tied to the sentence it affects, with its reason, so review time goes to what is still open instead of being spread evenly across the document.
Built to test your position
The drafting looks for the authority that hurts you and the argument the other side will make, not only the case for your position. The checking then asks separate questions of each case the work relies on.
The work is written to answer the question you asked and to find what would undermine the answer.
Each case is checked on its own: does it exist, does it say what the work claims, does it apply here. Quotations, pinpoints and whether a case is still good law are not checked yet.
The stage that checks the work wrote none of it and has no stake in it. Its only job is to find what is wrong.
Every flag stays attached to the check that raised it and the reason it gave, so your lawyer can see why something was flagged, not just that it was.
Worked example
Matter brief submitted synthetic example
Client is a freight brokerage. After a cargo loss the counterparty has invoked the indemnification clause. Assess exposure, and whether the limitation-of-liability cap survives an allegation of gross negligence in the State of Example.
Invented example content · written by hand · not AI output
Written by hand to show the shape: four passages from one invented draft. They are not a run receipt.
The cap is the weaker half of our position, not the stronger one. Example courts have been willing to read a liability cap as unenforceable against conduct pleaded as gross negligence, and the clause here is drafted broadly enough that the argument is available to them. The memo should not tell the client the cap holds.
relied on: Nordwyn Freight Co. v. Halloway Terminals · Thelbrook v. Ainsworth Rail
Opposing counsel will plead gross negligence whether or not it can be proved, purely to put the cap in issue and move the settlement range. The indemnity clause and the cap are drafted in tension with each other, and that tension, not the loss itself, is where their brief will go.
relied on: Castermere Logistics v. Bay Ridge Holdings
Two of the authorities the analysis leans on are not controlling here. One is a federal district decision applying a different state's law, which is at best persuasive. The distinction matters more than usual in this matter, because the cap question is where the jurisdictions actually diverge.
relied on: Vantry Mutual Assurance v. Orrin Cartage · Example Commercial Code § 000-000
Framed for the client rather than the file: the indemnity is the larger number, but the cap is the larger uncertainty. The realistic worst case is not the loss. It is the loss with the cap struck out, and the client should be told the size of that gap before deciding whether to settle.
relied on: no authority cited in this passage
Invented example content · fictional authorities · not AI output
Invented, not a run. Some rows show checks that are not built yet (quotations, whether a case is still good law, statutes) so you can see where they would sit; each is marked not yet.
| Authority as cited in the draft | Flag | Check | What the check found |
|---|---|---|---|
| Nordwyn Freight Co. v. Halloway Terminals, 000 Ex.Supp. 000 (D. Example 2019) | ⚠ UNCERTAIN | does it say that | The case was found, but the passages available do not clearly support the proposition it is cited for. The check could not settle it either way, so it goes to the lawyer marked, not passed. |
| Castermere Logistics v. Bay Ridge Holdings, 000 Ex.3d 000 (Example Sup. Ct. 2021) | ✕ NOT FOUND | does it exist | Not found in the published case law. That alone does not prove the case was invented, and a missing, ambiguous or failed lookup is each shown differently. |
| Vantry Mutual Assurance v. Orrin Cartage, 000 Ex.App.2d 000, 000 (Example Ct. App. 2017) | ≠ QUOTE MISMATCH | the quotation (not yet) | The draft quotes a sentence at a pin cite. Quotation checks are not built yet, so a quotation is always something to read at source before it is used. |
| Thelbrook v. Ainsworth Rail, 000 Ex.3d 000 (Example Sup. Ct. 2013) | ⚑ NEGATIVE TREATMENT | still good law (not yet) | Not yet: this needs a licensed citator. When it comes, overruled, vacated, abrogated, limited, superseded, questioned and distinguished will stay distinct rather than collapse into one warning. |
| Example Commercial Code § 000-000 | ⇄ JURISDICTION MISMATCH | statutes (not yet) | Cited in the draft as controlling, from a different jurisdiction from the one this matter is in. Statutes are not looked up yet; this row shows where that check would sit, and today the provision is left for the lawyer. |
| (no authority cited) for "a broker owes a heightened duty when it selects the carrier" | ∅ UNCITED | no citation to check | A legal proposition asserted in the draft with no authority attached at any point. It may well be right. Nothing in the run supports it, and the finished document marks it that way. |
No result is ever labelled "verified". Whether a case exists, whether it supports the claim, whether a quotation is right and whether the case is still good law are separate questions; the last two are not checked yet.
Invented example content · written by hand · not AI output
Open questions are recorded rather than smoothed over. A memo that reads as though the question were settled is less useful to the attorney than one that shows exactly where it is not.
Invented example content · written by hand · not AI output
Invented. Each point in the memo shows what it rests on and the status of those authorities.
Every sentence of the finished memo is given a status, and its cited claims are checked again before it reaches the lawyer.
What lands with the attorney is a memo in which the argument that matters most to the client rests on two authorities the run could not stand behind. That is the honest output for this matter, and a version of this page where the demo resolves cleanly would be a worse advertisement for the design, not a better one.
Invented example · illustrative counts, not measurements
Illustrative, not measured. What drives the cost of a run is the number of distinct authorities, the length of the documents and how much goes back for repair. The AI processing runs on commercial infrastructure, so a library kept in your office would not remove that cost.
Illustrative: one matter run
| Stage | Work | How it scales |
|---|---|---|
| Intake | read the request and the files; apply permissions; recall permitted past work | one pass |
| Drafting | draft the work from the request, the files and the past work | one pass |
| Authority Check | look up each distinct authority, then test it against its claim | grows with distinct authorities |
| Repair & Re-check | fix what was flagged and check it again, within a bounded loop | grows with what was flagged |
| Written up | one finished document, flags against the sentences they affect | one pass |
| Final Check | the finished document's cited claims, and a check against the request | grows with the document |
Each authority is checked once, however many times it is cited, and repairs are checked again before the document is written up.
Actual usage comes from a particular run and configuration. This table provides no latency, savings or comparative-quality measurement.
The flags
A flag a firm can't act on precisely is worse than no flag, because it spends attention without directing it. Every result keeps its source, its reason and its status separate. Some labels below belong to checks that are not built yet, and say so.
Not a label HiCourt uses. A case that was found and passed its checks is shown as exactly that, because a lookup does not establish that a quotation is right or that the case is still good law.
Not found in the published case law the lookup searches. That does not by itself prove the case was invented, and an ambiguous or failed lookup is shown differently.
The check could not settle, from the passages available, whether the case supports the claim. It directs your lawyer's review; it is not proof the citation is wrong.
Not yet: quotation and pinpoint checks are not built. Until they are, the lawyer compares any quotation against the source and page.
Not yet: this needs a licensed citator. When it comes, overruled, vacated, abrogated, limited, superseded, questioned and distinguished stay distinct, each with its source and date.
The check assesses whether the case applies in your forum. An uncertain answer is shown as uncertain, not as a mismatch, and a case cited as persuasive is not flagged just for not being binding.
A sentence in the finished document that rests on no citation is marked, so it is not mistaken for a checked one. Checking uncited assertions against the law is not built yet.
Pricing
There is no price list. The commercial shape is genuinely undecided, and the in-office deployment is part of why.
TBD
Flat annual fee, per-attorney licence, per-matter charge, or usage-based. All four are live options. Each suits a different size and shape of firm, and getting it wrong is expensive to undo.
A live decision, not a settled one.
a subscription
A flat fee with generous limits, rather than a bill that moves every month. A meter would make you weigh whether a matter deserves the full treatment, which is the opposite of the point. Current thinking, not a settled decision.
You should not have to ration the thing you bought.
hard limits
Each run has a spending ceiling and stops when its recorded cost reaches it. Billing against a firm's allowance is not built yet.
A run must be unable to surprise you.
We do not price against attorney time. "This costs less than N hours of associate review" is the natural next sentence and it is an outcome claim wearing arithmetic. It asserts how much work this replaces, which nobody has measured. What a run costs to operate and what it is worth to a firm are different questions, and only the second one matters to you. That one is yours to answer.
What would move a number, once there is one: the count of distinct authorities in a matter, document length, and how much goes back for repair, rather than the number of lawyers using it. A long brief with 90 authorities is a different amount of work from a short memo with 12, and any pricing that ignores that will be wrong for somebody.
If your firm has a strong preference about how it would want to be billed, that is one of the more useful things you could tell us, because it is a live decision rather than a settled one.
The firm's library
Your firm's library is in use today: the files added to a matter are read and indexed, and past work is recalled where the person asking may see it. Each firm runs in its own dedicated hosted environment; keeping the library on a machine in your office or on each workstation is not offered yet.
A file added to the library exists in several forms, and every one of them is protected the same way as the original.
| A · Private store we run | B · Machine in the office (not yet) | C · Per workstation (not yet) | |
|---|---|---|---|
| Files at rest | Our infrastructure, single-tenant to the firm | Firm-owned drives on the firm's LAN | Each individual workstation |
| Index scope | One index across everything the firm has uploaded, which is what makes it answerable | One index per machine; no firm-wide question is possible | |
| Key custody | On the firm's own server volume, outside the backed-up database; our access none by default, support access requested and logged | Firm-held, firm-managed | Whatever that workstation's disk encryption is |
| Backups | Ours to run, to a schedule the firm agrees | The firm's to run, and the most common real-world failure | Usually nobody's, in practice |
| Deletion | Has to reach the file, its passages, the search index, past work and backups | Same, times the number of machines | |
| Exposure if a device is lost | None, since nothing is local | One machine, in a controlled room | A partner's laptop in a taxi, holding whatever they had uploaded |
Option C is more contained per device and often worse overall. Twelve unmanaged copies with twelve backup stories is not obviously safer than one managed store, and it forfeits the shared library entirely. It is a legitimate choice for particular matters and a poor default for a firm.
This is the part that matters more than where the disk is, and it is the part we would push hardest on in anyone's product, including ours. A retrieval index does not naturally respect an ethical wall. Ask it a question and it returns whatever is semantically nearest, which may be from a matter the asker must not see. A screened person asking an ordinary question is not doing anything wrong, which is exactly why the control has to sit in the system rather than in people's judgement.
HiCourt applies the asker's permissions before anything is scored: ethical walls and matter membership decide which documents a search may even consider, and anything outside them is dropped before it is read. Having been read recently does not make a document eligible.
Each run keeps a protected record of what it did, readable only by people who are still allowed to see it. Deleting a file everywhere it has been copied, including backups, and handling legal holds across every copy, are not finished yet.
Every run records the request, the files and past work it drew on, each stage's outcome and every finding. The record is protected like the work itself and is checked against the reader's permissions each time it is opened.
Where the library sits is not the same question as who reads it. Whichever arrangement a firm picks, when someone asks a question the matching excerpts are sent to whatever produces the answer. Storage location bounds where material rests; it does not by itself bound where material travels at query time. That is the next section, and the two decisions should be made together.
Processing
Drafting and checking use commercial AI, so the text of the work is sent out for processing. Only what the person asking may see is sent. Where your files are stored and where the work is processed are separate decisions.
HiCourt is in testing with one test firm. Setting out the AI processing and its data terms in writing before a firm sends anything is still to come; this page does not attest to contracts.
| How it works | |
|---|---|
| The AI | Commercial AI, used for drafting and checking. HiCourt chooses and configures it; your firm does not have to. |
| Matter text crosses | Your firm → the AI processing, for each run |
| Egress during a run | Required, because that is the mechanism |
| Who holds your library's keys | In the hosted arrangement, on your firm's own server volume, outside the backed-up database. Where the library would sit on your hardware (not offered yet), you would. That is a separate question with a separate answer, and this section does not change it. |
| Who operates what | Your firm uses it through the browser. We operate HiCourt. The AI itself runs on commercial infrastructure. |
| Whose AI accounts | Ours. HiCourt holds the accounts and makes every call. Your firm supplies no API key, no AI access and no separate contract — there is nothing for you to procure, and nothing for you to be billed for separately. |
| Named before you upload | Not yet: the AI processing and its data terms set out in writing before you upload. |
| Capability | Upgraded as better AI becomes available, each change tested end to end before it is used. |
| Blast radius if breached | Depends on the AI infrastructure's own controls, and is not the firm's to control |
What to hold us to, given the trade is settled. The useful questions are not whether matter text is sent out for AI processing — it is, on every run — but where it goes, on what terms, and what you can see afterwards. Ask where a run's text is processed, the retention and training terms we have contracted for, and the record of each run. Those are things you can check. "Your data is safe with us" is not.
The one exception, named because it exists. A firm that wants the AI itself on its own hardware is talking about buying and running a server capable of it. That is a different conversation and a different shape of product, and we would rather have that conversation than pretend the standard arrangement is something it isn't. It is not what this page describes. on the table
Layers 1–3 are what routing settles. Layers 4–8 are where your choice of where the library lives does its work, and are where most of the engineering is.
Each firm's dedicated hosted environment runs its tasks, holds its library and keeps its run records. A run's record is readable only by people still allowed to see it; finishing a run does not mean its content vanished.
That last sentence is a property to check rather than a promise to accept. Work in progress necessarily passes through this layer, so the questions worth asking are what is written to disk, what appears in logs, and how long a completed run's working state survives. Ask us to run a matter and then show you.
by design Separate tenancy per firm, so a defect or a compromise is bounded to one firm rather than shared across all of them.
Commercial AI does the drafting and checking. The case lookup is not AI at all; it searches the published case law.
The AI processes only the text a run is allowed to send. No speed, cost-saving or quality figure is claimed.
HiCourt configures the AI; a firm does not have to choose or manage it.
A run sends the text of the work, and the permitted passages it needs, out for AI processing, and nothing the person asking may not see. The files a run uses are fixed when you press Run and checked against permissions again before material is sent.
Not yet: the AI processing and its data terms set out in writing before submission.
Matter text, drafts, the authority table and the audit record all persist somewhere. Encryption at rest, with the keys kept outside whatever gets backed up, is the requirement; the firm's existing disk and backup policy should govern it rather than a second, parallel regime nobody audits.
by design The firm's existing backup policy governs it, rather than a second parallel regime nobody audits.
Permissions are applied before anything is searched: ethical walls and matter membership decide which documents a run may draw on, and anything outside them is dropped before it is read. They are checked again before material is sent for processing.
by design See below. This is the layer most worth interrogating, and the one we would rather you pushed on hardest.
Every run records the request, the files and past work it drew on, each stage's outcome and every finding. The record is checked against the reader's permissions each time it is opened.
A step whose outcome was not recorded is shown as unknown, never assumed to have succeeded.
Run records are kept and protected like the work itself, and are readable only by people still allowed to see them.
Not yet: deleting a file everywhere it has been copied, including its passages, the search index, past work and backups, with a tested end-to-end path.
Matter documents reach the AI's input and the rendered file only. They never reach tool execution or automation, and a document containing something that looks like an instruction is rendered as text, not obeyed. This is a standing project rule rather than a feature, and it is why the final stage assembles a document and cannot act.
decided The rule is fixed, and the architecture is built around honouring it.
Independent review is something your firm should require, not something we should assert. Penetration test, certification, third-party assessment: whichever arrangement you pick for your library, insist on them and insist on reading them. If there are specific ones your firm would require before adopting something like this, tell us which, because that is the gap between a design and anything you could responsibly rely on. Nothing on this page is a security assurance. It is a description of what would have to be built and independently checked before one could be given.
Confidentiality
The work is processed by commercial AI on every run. Where your library is stored decides where your files rest; it does not stop the work being sent out for processing.
What that solves, and what it leaves. Keeping the library on your own hardware means the standing collection of everything your firm has uploaded is under your control, your encryption and your retention schedule. It does not mean matter text stays in the building: the excerpts a task needs still go out for AI processing on every run. "Our library is on our own server" and "nothing we work on leaves the building" are different sentences, and only the first is a design property here. A firm that needs the second is describing a different product, and we would rather say so than let the distinction go unnoticed until someone's IT lead finds it.
Whether output like this is protected work product is a legal conclusion about your matter. This page does not make it, and a vendor asserting it would be telling you something it is not in a position to know. It depends on purpose and circumstances, and the third-party-transit question is squarely inside that analysis under every arrangement offered here. It is your call, and it should be made before anything goes in.
The professional duties in play (confidentiality, competence with the tools you use, supervision of work done with them, and candour about the authorities in a filing) are described here rather than cited. Quoting a rule number or a bar opinion from memory on a page about fabricated citations would be an unusually stupid way to prove the point.
The firm's memory is in use: every finished run is filed into it, and research and analysis recall permitted past work. That makes the permission line matter more, not less.
On the main pitch
Cross-project memory is the moat: a corpus built out of your own work, which a competitor can't buy.
In a firm
Cross-matter memory is the moat here too — a firm's precedent files always have been — but a retrieval layer that lets one matter's material surface in a run it should not is a confidentiality and conflicts problem, and keeping the machine in the building does nothing to fix it, because the leak is between people and matters rather than out of the door.
HiCourt keeps that line in the software: past work is recalled only where the person asking may see it, labelled with where it came from and whether it was reviewed, and it never replaces a fresh check of the law.
This page is static: no server, no cookies, no analytics, no external requests. The form below opens your own mail client and sends nothing on its own.
It composes to a personal mailbox at a consumer email provider, because this static page has no server of its own behind it. That is a fine channel for "here's what my firm would need" and an unacceptable one for anything about a client, a matter, or a party. Please don't put confidential facts in it.
The reader this page is for is professionally trained to notice what a document quietly asserts. So here is exactly what each part of the output means, stated tightly enough to rely on.
We're looking for the firms this is right for, and for what it would have to do before your firm would let it in. A clear no, because… is genuinely the most useful reply.
Any pilot would run on synthetic or already-public material only (a moot problem, a published opinion, a redacted precedent from your own library) rather than a live privileged matter. A first conversation should not need one, and we would rather earn the real thing after your team has been through the architecture above.