Law Firms plan

Legal teams need shared authority before matter judgment.

Buy a law-firms library built around the stable legal corpus firms repeatedly consult before applying private client facts.

Hard-retrieval benchmark

The raw context hallucinated. The crystal was right.

Three-arm selective waiver test — SEC voluntary cooperation, private class action following, Second Circuit focus. A specific question with a known correct answer. Two arms got it wrong.

Raw context hallucinated

Wrong case. Wrong circuit.

The raw arm cited Upjohn as a Second Circuit selective waiver case. Wrong case, wrong circuit, wrong doctrine. 414,000 tokens of source material didn't prevent the hallucination. Score: 12 / 15.

Crystal was right without retrieval

Circuit split, correctly mapped

Without any retrieved case law, the crystal correctly identified which circuits accept selective waiver (8th Cir.) and which reject it (2d, 3d, 6th). Conditioned grounding alone. Score: 14 / 15.

Crystal + RAG is the best architecture

Matched the ceiling

With case law retrieved and injected across all three arms, crystal + RAG matched the highest quality score in the test. Crystal for grounding, RAG for authority. Neither bloats the prompt. Score: 12 / 15; all arms improved.

How this was run. Single-question, three arms: no retrieval (raw and crystal) and full retrieval (all arms). 15-point rubric. No-context: 13 / 15. Raw: 12 / 15. Crystal: 14 / 15. With RAG injected across all arms, scores increased; crystal + RAG matched the ceiling.

Multi-turn benchmark

One configuration forgot its instructions. One didn't.

Three arms, same model (gpt-4.1-mini), same question set — no context, raw-accumulating, hosted crystal. At turn 12 the crystal sent 2,684 tokens; the raw arm sent 59,689 — 22× more, compounding every call.

Prompt size at turn 12

22× smaller

Crystal footprint vs raw-accumulating at the final turn. Crystal: 2,684 tokens. Raw accumulating: 59,689 tokens — 22× larger and growing every call.

Instruction retention

No-context forgot

A formatting rule planted at turn 1 was checked deterministically every subsequent turn. No-context broke it 2 times (8 / 10 held). Crystal: 9 / 9 checkable turns held.

Answer quality — 12 turns

No degradation detected

All three arms produced substantive answers through turn 12 with no meaningful quality decline. The separating factors are prompt size and instruction retention, not raw answer quality.

How this was run. 12 turns, Law Firms vertical, gpt-4.1-mini held constant across all three arms. Raw-accumulating arm: source corpus injected at turn 1, then full conversation history re-sent on every subsequent turn — the window grows every call.

Subscription

$6,000/mo for Law Firms.

Workspace subscription for firms that want a shared legal reference layer before connecting private matter material.

Monthly allowance

8,000,000 credits/mo

Up to 25M raw-token-equivalent shared legal reference coverage.

Start 7-day trial

Why this can cost less

Raw RAG can inflate every prompt with retrieved text. The crystal keeps shared operating judgment compact and reusable, so teams buy a monthly memory layer instead of repeatedly paying for larger, noisier context windows.

Context crystals

Representative coverage.

Each plan includes protected access to prepared industry memory across the recurring work areas below. Your team sees the benefit in chat while the library's source construction remains private.

Authority Orientation

A legal-authority map for separating binding authority, persuasive authority, procedural posture, and background context.

  • Jurisdiction-aware authority orientation
  • Citation-grounded holding and rule treatment
  • Procedural posture and standard-of-review patterns

Regulatory Guidance

A public regulatory reference layer for agency positions, obligation types, and compliance language used before matter-specific advice.

  • Agency guidance taxonomy by obligation type
  • Regulatory update interpretation checkpoints
  • Compliance-language consistency review

Practice Drafting

Drafting reference support for common legal work products without exposing templates, prompts, or private firm playbooks.

  • Brief and motion argument structure cues
  • Contract clause purpose and risk notes
  • Public filing and form-language comparison