Everything about the Fault-Line Radar in one place — what it tracks, how it scores, how evidence gets in, what the record already requires of a firm, and how a reading becomes an instruction.
Law firms are rushing to use artificial intelligence, because it drafts and researches in seconds what used to take a junior lawyer half the night, and the pull to lean on it is enormous. Nobody has settled which of the mistakes AI makes will actually get a lawyer in real trouble, and the rules that will govern it are still being written, one court case and one ethics ruling at a time. So every firm is quietly betting on what the rules will say before those rules exist.
This is a tool for making that bet with your eyes open. It watches where something new that AI can do bumps against an old rule about how lawyers must behave, and tells you which collision the legal world is about to crack down on next. It measures pressure, the way a weather service tracks a storm building at sea, so a firm gets ready before landfall instead of scrambling after it.
Every week it reads a wide pile of real, dated, sourced material: court punishments, ethics opinions, new laws, what the software companies are building, what insurers demand before they'll cover a firm. A ruling from an actual court counts for a great deal; a marketing blog counts for almost nothing, so ten breathless blog posts can never outweigh one real decision.
Before any fact counts for anything, three separate checkers try to prove it wrong, and it only gets in if it survives at least two of them. That is how the system has already caught its own mistakes, like a court rule that only one website mentioned and a claim about two companies that fell apart the moment someone looked closely. It also goes back and grades its own past readings against what actually happened.
The tool comes in three parts, and each answers a different question.
Here is an example, worked all the way through.
A lawyer has a hard duty to keep client secrets private. An AI that reads a firm's documents can now do the work in seconds. Type the secrets in, though, and they quietly belong to the company running the tool. Secrecy against speed. That collision is a fault line.
The pressure on it is already high, because the duty is absolute while the tools are new and loose. What would settle it is the law making a move. Until then the honest read is pressure building, and the fix a firm can stand up today is data-flow mapping and vendor attestation, so a lawyer knows where the secrets go before they go anywhere.
The same call, scaled to the whole field of eleven, follows on the board below.
These are the six controls the law already requires of any firm, in the order to act on them. The order is the shortest measured warning window first: a control whose antecedents historically bound in 192 days gives less notice than one that took 907, so it comes first. It is not an order of importance, and it is not the number of sources. The four kinds of authority on the left flow into a bar for each control, so you can see what the duty actually rests on. The other five fault lines, the ones nothing requires yet, are listed underneath rather than charted.
The two axes disagree on purpose. convergence has ten sources and sits fourth; verification has twenty and sits last, because it carries the longest warning window on the board at 907 days. Reading down the list is a plan; reading the bars is a measure of proof.
Five fault lines sit below the flag threshold. Their ruling-eligible counts run from 3 down to zero, so drawing them as bars beside six well-supported duties would read as a finding about them rather than about the record. They are listed instead, with the reason each is still worth watching.
| Fault line | Sources | Why it is on the watch list |
|---|
A reading is only as good as the evidence under it, so the radar refuses to lean on any single source. Every item that counts is one of four kinds of authority, and the kinds are kept apart because they are not interchangeable: a circuit court can sanction a lawyer, and a bar association can discipline one, but a state ethics opinion carries a different weight from a ruling that binds.
Every item is real, dated, and traceable, and weighted by how much authority it carries before it counts. A court ruling moves a reading far more than a company's press release ever could. A reading is the sum of its evidence, collected, weighted, and added up.
Circuit courts and state appellate benches. When a higher court rules, the reading moves further than it does for anything below it.
District orders and sanctions. Lower authority, but this is where enforcement usually starts, so it moves early.
Binding primary law that needs no court to decide anything. A statute binds without any AI capability driver behind it, which is why a line can carry a high ruling reading and a low capability one.
Ethics opinions and bar guidance. The profession telling itself what is required. Counted, but never weighted as if it were a ruling.
Authority over volume, set by one test: how much can this speaker actually do to a lawyer? A court can sanction you. A bar can discipline you. An insurer can refuse to cover you. A vendor can only sell to you. Every item is stamped with its tier when it enters, and the tier decides the weight, not how many people repeated it.
| Tier | What it is | Weight |
|---|---|---|
| T1 | Court ruling, statute | 1.00 |
| T2 | Bar opinion, regulatory guidance | 0.80 |
| T3 | Empirical study, institution | 0.60 |
| T4 | Commentary, trade press | 0.35 |
| T5 | Vendor marketing | 0.008 |
A stream and a tier answer different questions: a stream is the topic, a tier is how much power the speaker holds. Most rules-and-rulings items are high authority and most vendor items are low, but the radar keeps the two apart so a low-authority claim cannot borrow a prestigious name.
A fact does not enter the feed because someone found it. It enters because it survived being argued against.
This is the step that caught real mistakes: a court rule only one source mentioned, a "disqualification" that was actually a reprimand, a claim about two vendors no source could hold up.
Every fault line is read on five meters at once. Three watch the supply side, the chain of things that has to exist before a firm can act. Two watch the governance side, whether the law and the market are about to force the issue. They do not all fire in that order, and the board does not pretend they do: convergence and vendor liability both carry a ruling reading above 8 with a capability reading at 2.5, because a statute binds without any capability driver behind it.
Every fault line is read on the same five meters, and colour runs hot to cold by pressure. Hover a fault line name to see what it is, or a column header to see what that meter measures.
| Fault line | What it is | The move | Capability | Ruling | Adoption | Enablement | Software |
|---|---|---|---|---|---|---|---|
| competence required | Knowing the tool well enough to catch it being quietly wrong. | Train every lawyer to catch the tool being quietly wrong, and run that as a firm program, not a policy nobody reads. | 5.8 | 9.3 | 5.5 | 7.9 | 5.0 |
| vendor liability required | Who takes the blame, the seller or the firm, when an AI tool does harm. | Document which tool produced what, so the blame lands on the right party when a tool does harm. | 2.5 | 8.3 | 3.3 | 3.0 | 3.8 |
| confidentiality required | Keeping client secrets out of the AI tool that reads them. | Map where every client secret goes, and get the vendor to attest to what they hold and how. | 4.5 | 9.4 | 6.0 | 6.6 | 5.5 |
| convergence required | Rules landing across courts and states that a firm must follow everywhere at once. | Run one operating model everywhere and follow the strictest rule in force, so no single state's standard catches you out. | 2.5 | 9.3 | 6.0 | 4.0 | 5.5 |
| disclosure required | Saying when AI helped produce the work. | Disclose AI help by default, with an audit trail that shows who touched what. | 5.5 | 9.5 | 6.0 | 7.2 | 5.5 |
| verification required | Checking that what the AI wrote is actually true before it reaches a court or a client. | Check every AI output before it goes out, and keep the trace log that proves it was checked. | 7.8 | 9.7 | 6.5 | 8.4 | 5.6 |
| judge analytics not yet required | Predicting how a judge will rule, allowed in some places and banned in others. | Simulate strategy, and never claim to predict a specific judge. | 4.0 | 7.8 | 3.0 | 3.0 | 3.0 |
| benchmark not yet required | A shared test that certifies an AI tool is accurate enough to trust. | Route work only through tools that have passed a shared accuracy test. | 8.4 | 7.6 | 6.4 | 8.3 | 7.0 |
| agentic not yet required | AI that acts on its own, and how much a human stays in charge. | Keep a human on every decision an agent would make, and limit what it may do on its own. | 6.4 | 7.2 | 6.3 | 6.0 | 7.5 |
| fees not yet required | Billing for work AI does in minutes that used to take a lawyer hours. | Bill on the value delivered, not the hours, because AI has already collapsed the hours. | 2.5 | 6.7 | 5.3 | 3.5 | 3.0 |
| insurance not yet required | Insurers deciding whether they will still cover a firm that uses AI. | Build governance the way an insurer would price it, so their questionnaire comes back yes. | 2.5 | 6.5 | 8.1 | 6.6 | 4.0 |
Everything above is what the record says, and it is true of every firm. This part depends on who you are, and it is where the radar stops describing and starts advising. Two questions are asked of every control, and they are never blended into one answer.
A norm only counts if someone with something to withhold requires it. An insurer who can drop your coverage. A client who can move the engagement. A platform that can withhold the docket. A high adoption score on its own is not enough, and that is deliberate: it is why the model does not use one.
Leverage is firm-relative. A carrier you name yourself clears the evidence bar by itself, because one carrier requiring something of you is a stronger signal than two requiring it of the market in general. Naming your carrier is what turns a thin norm into a supported one.
Carriers have started asking. CNA added AI-governance questions to malpractice renewals and underwriting in June 2026, which is a questionnaire rather than a condition, and no bar has yet adopted a named benchmark.
The event to watch is the step past asking: an insurer naming a certified tool as a condition of coverage. That is what flips the method from voluntary to enforceable.
The most valuable signal in the whole system is the thing that has not happened yet.