The ANCHOR Engine
Models listen. Code decides. An attorney signs.
The engine behind Counsel Assist turns a client's phone call into a verified family-law case file and a first draft of the papers. This is how it is built to be trusted, and how we prove it.
0
wrong facts on the call bench
132 full scripted calls: 22 case files × 6 caller personas, scored against answer keys.
99%
of facts right with model-played callers
198 full calls across 9 caller behaviors, from stressed to off-script. Every miss adjudicated; 96.4% before setting aside the caller's own contradictions.
5,605
automated checks, all passing
Across 19 QA suites, from adversarial callers to every paper in all four states, plus the conversation and unit tests. They run on every pull request.
1844
automated tests on every change
Across 181 test files and 19 QA suites, from parsers to state papers.
How it is built
Our code has the final say on every fact.
Language models help ANCHOR understand people. Our own code keeps the record, and the firm's attorney makes every legal call. We walk firms through the details in a technical briefing.
Then a licensed attorney. The engine drafts; the firm's attorney reviews every page before anything is filed or served.
Design guarantees
What holds on every call.
Traceable to the client
Every fact a model contributes carries the client's own words, checked against the transcript before it is kept.
Legal logic in code
Which papers, which statutes, which deadlines: decided by rules that are tested, versioned, and the same every time.
Caution only increases
Danger is heard by code on every turn as well as by the model. A safety fact is added to, never replaced. Review can raise caution, never lower it.
Never stuck on a model
Every model step has a code path behind it. A slow or wrong model answer is dropped, and the call goes on.
Evaluation
Measured five ways, every release.
No single benchmark captures a phone call. So the engine is tested from the smallest parser to a full conversation with a difficult caller, and every layer has to hold.
| Layer | What it tests | Scale | This release |
|---|---|---|---|
| Unit and regression tests | Parsers, the rules engine, state papers, and every past bug, pinned so it cannot return. | 1844 tests in 181 files, plus 19 QA suites | All passing on this release |
| Adversarial and edge-case suites | Spoken openings, adversarial callers, matters, states, and postures, each with the papers that must and must not appear, and every paper read for pleading language in all four states. | 5,605 checks in 19 suites | All passing on this release |
| Scripted call bench | 22 complete case files, each called by 6 caller personas: clean, terse, rambling, noisy, unsure, and over-sharing. Every fact is scored against an answer key. | 132 full calls | 132 of 132 finished, 0 wrong facts |
| Model-played callers | A separate model plays the client in 9 behaviors (stressed, vague, off-script, self-correcting, confused, emotional, and more). An independent adjudicator separates engine errors from the caller contradicting their own case. | 198 full calls | 99% (98.9% measured) right after adjudication (96.4% raw); 56 engine errors in 198 calls |
| Live calls | Our team calls the production line with scripted clients, and every finding becomes a permanent test before the fix ships. | Every release | Findings closed with a test |
- vague99.8%
- off-script99.3%
- terse99.2%
- self-correcting99.1%
- confused99.1%
- plain99.0%
- stressed98.8%
- chatty98.5%
- emotional97.6%
The read-back, written by a model and checked by code
Before the file goes to counsel, the client hears it summed up in plain language. A model writes it from the verified file; code rejects any number, name, or place that is not on the file, any advice, and any dropped safety fact.
19 of 22 test files passed every check · median 4.4s · the rest fall back to the code's own read-back, 100% of the time.
A safety floor that does not depend on a model
Recent violence, a threat to return or to take a child, a weapon used to threaten: code hears these on every turn, marks the file urgent for counsel, and says the safety line first. The model's judgment adds to this floor; it can never lower it.
Coverage
Four states, one engine, the law in code.
Each state's petition, affidavits, worksheets, and service papers are drafted from the same verified file, with the state's own statutes and forms.
40 family-law matters, mapped in the open
Arkansas · Missouri · Oklahoma · Kansas
- 30 handled end to end
- 4 partial
- 6 on the roadmap
A matter counts as handled only when the engine drafts its lead paper in every state and it has its own test cases. We publish the gaps because a buyer should know exactly where the edge is.
Security and data governance
Built for a firm's IT review.
Consent before the microphone
The client sees who processes their words and agrees first. The exact text and its version are kept with the file.
Firm isolation
Row-level security policies in the database scope every matter to its own firm's members, and every case page checks membership too.
Sensitive numbers redacted
Social Security, bank account, routing, and card numbers are masked in stored transcripts.
Keys stay on the server
Model and voice credentials never reach the browser. Model and voice services are reached only through our server, with limits on every route.
Who handles client data
Speech to text: xAI. The assistant's voice: ElevenLabs. Language understanding, review, and the read-back: Anthropic (Claude). Each handles the client's words only to run the call.
No training on client data
CaseForge does not train or fine-tune any model on client calls or files.
What we do not claim
The edges, stated plainly.
A serious evaluation says where it stops. These are ours.
- The bench and model-played callers use fictional people and fictional case files. They measure the engine, not a particular firm's clients.
- Speech recognition can mishear an unusual name. The engine shows every name on screen during the read-back so the client can correct it.
- Every paper is a first draft for the firm's attorney. Nothing is filed, served, or sent to a court by the engine.
- Coverage is four states and the matters marked handled below. The rest is listed as roadmap, not implied.
For your technical team
The questions we expect.
Is this a model you trained?
No. ANCHOR uses frontier language models to understand people, and our own code has the final say on every fact, date, and paper.
What stops a model from inventing a fact?
A fact is kept only when it traces back to what the client said, and the client hears the file read back before the call ends. Counsel reviews every page. We show firms exactly how in a technical briefing.
What happens when a model is slow or down?
The call continues. The questions, the safety response, and the read-back keep working.
How do you know a change did not make it worse?
Every pull request runs the unit tests, the QA suites, and a full production build automatically. Those checks include the adversarial suites and an audit of every generated paper. The call bench is run against its saved baseline before a release, and a wrong fact on the file is treated as a failure.
Can we see the evaluation?
Yes. We walk your team through the method, the answer keys, and the current numbers in a technical briefing, under NDA where needed. The engine's internals stay ours; the evidence is yours to inspect.
Technical briefing
Bring your hardest questions. We bring the evidence.
We walk your IT and data teams through the method, the answer keys, and the current numbers, and run a live call with your scenario.
Request a technical briefingMeasured November 4, 2026 on release 7e22514. Test callers and case files are fictional. Results describe the engine on this release and are re-measured with every change.