What it does, and what it doesn't — with its limit on every line.

Everything in the first section is something the software does today, inside a running deployment. What it doesn't do yet lives lower down, plainly labeled, and is never dressed up as “coming soon.”

The portfolio, ordered by what's due

real today

Your cases sorted so the matter due first sits on top — the writ due Friday ahead of the brief due next month.

Limit getting onto a deployment is a conversation, not a signup. There is no self-serve cloud signup today, and that's a deliberate cut.

Shipped 2026-05-29.

Conflicts between documents, surfaced

real today

Where two documents in the record say different things — a CT minute order one way, an RT transcript another — Amicus OS flags the contradiction instead of quietly smoothing it over.

Limit it surfaces the conflict for you to resolve. It does not decide which document controls — that judgment stays yours.

Shipped and grounded in this case's own record.

Co-reading the record

real today

Read the record side by side: highlight a passage, attach a note, mark it for review, and tie it to its place in the timeline. When Amicus OS draws a summary, every sentence shows the page:line it came from — or it is held back, not shown. A private note stays out of the signed record and is never sent to any model.

Limit mapping a claim across documents and regenerating a summary depend on an AI assistant being connected, and aren't always live.

Shipped 2026-05-29.

A drafting outline with bar-numbered structure

real today

A skeleton for the brief that keeps each cited passage attached to its record source, organized by issue, carrying your bar number through to the moment you accept it.

Limit it outlines and it checks — it does not write your brief. A sentence lands only if a verified record or authority cite backs it; case law is refused inside the outline and handed off as a research prompt instead. A four-sentence cap and a three-part acknowledgment gate every acceptance. A watch-it-build-live drafting window is design intent, not shipped.

Drafting-outline and bar-numbered structure shipped, 2026 release R16.

The California Style Manual tools

real today

Table of Authorities, Table of Contents, Certificate of Compliance, and the CSM formatter — built by rule, not by guessing.

Limit these are mechanical and rule-based. No AI touches them, so there is nothing here to hallucinate.

California Style Manual tooling — mechanical and rule-based.

Research prompts you take to Westlaw or Lexis+

real today

Plain-language research prompts drawn from your case issues, ready to run on Westlaw or Lexis+.

Limit Amicus OS does not fetch authority or read a holding for you. The prompts are built from redacted excerpts of your own record — no client identifiers and no sealed material ever reach any model. Refining a prompt and saving a result aren't persisted yet; those surfaces say “not yet saved” rather than fake a save.

Shipped. The composer refuses to fetch a holding; surfaces that aren't persisted yet say “not yet saved.”

Getting the record in

real today / limited

Today the working way in is the command-line ingest tool. You point it at the folder of PDFs and it brings the record in, runs OCR, and pulls the citations.

Limit a drag-and-drop browser upload is not built. That route returns an explicit “not implemented” — never a fake success. The command line is the only live path in today.

The browser upload route returns an explicit “not implemented” — never a fake success. The command line is the live path in.

It refuses rather than guesses

When Amicus OS can't verify a citation against a real document in your record, it does not write the citation. It tells you it can't, and it shows you why. It would rather refuse than guess.

What “verify” means here, exactly A citation is checked against the documents in your record and the authority on file in your case. A case authority “matches” only when it is on file in this case — that is a match on file, not a good-law or Shepardize ruling, and it is not a misquote check. We do not check case law today.

Limit, in the same breath Full citation checking is partial today. On our 30-citation test fixture — a curated set we use to hold a regression floor, not the ceiling of what it reads — about 70% of the citations resolve to a treatise or statute we have indexed. Full statute and case-law resolution is the work we're doing now and have not finished.

The citation check returns the refusal and a signed record of every attempt, never a fake pass. The 70% is a measured result above a 60% floor, achieved at the 2026-05-03 release — not a guarantee.

It won't leak

The limit first, because it's the one that matters In the default hosted setup, we hold the encryption key for your organization. That keeps every other firm out of your case — but it does not mean you are the only one who can ever decrypt it. If you need keys that only you hold, that's the on-premises setup, which we have specified but not yet built.

Then the promise: confidential material is encrypted while it sits at rest, and decrypted under your organization's key only to serve your case. Sealed records and flagged personal information stay closed until someone deliberately opens them in the app, and that opening is itself logged. Identifiers — names, dates of birth, case numbers, addresses — are stored only as a one-way keyed digest; the readable value is not kept.

When an AI assistant is involved — for example, regenerating a summary — the prompt is built inside your deployment from redacted excerpts of your own record and public legal authority. Client identifiers and sealed material never leave the boundary, and never reach any model. The model can be a public frontier model or one you run yourself; either way the redaction happens before anything leaves.

A note on older records Any older record that still held an identifier in readable form is refused on open until it is re-keyed — the system blocks rather than expose it. So we don't claim “never written to disk” as a flat absolute. We claim the value is stored only as a one-way keyed digest, and an older readable record is blocked on open until it's re-keyed.

Encryption-at-rest verifies the sealed copy before the plaintext is deleted; identifiers are kept only as a one-way keyed digest; cross-firm access is refused at the boundary.

It won't obscure

Amicus OS keeps an append-only, signed record of every step it completes. If anyone tampers with that record, the verifier inside your deployment refuses to show the rows and reports the break. The whole chain can be re-checked, and an exported record can be re-checked independently.

What the signed record covers Every step it completes. Some surfaces aren't persisted yet and say “not yet saved” rather than fake a signed row. That's why the honest claim is “a signed record of every step it completes,” not “signs everything.”

Limit, in the same breath We call this tamper-evident, not a court-certified chain of custody. Software tamper-evidence is not a notarized, admissible custody record, and we make no admissibility claim.

The verifier refuses to expose any rows on a tampered, truncated, or broken-linkage record; surfaces that aren't persisted yet say “not yet saved” rather than fake a signed row.

Doesn't yet — and we won't call any of it “coming soon.”

Never claimed

A note for engineers evaluating the architecture

The engine underneath is practice-agnostic: reading a record, holding it as a cited chronology, checking a brief against it, and refusing to invent are the same job in any appellate practice. What is built and proven for one practice is California juvenile dependency — the hearing checklists, the issue templates, and the indexed authority are all keyed to it. Another practice area would need its own checklists and authority, which we have not built. The general shape is an architecture fact, not a claim that Amicus OS serves a wider docket today; when it does, we'll say so.

See it in full: the citation check refusing to fabricate a cite → /demo