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.
Buy a law-firms library built around the stable legal corpus firms repeatedly consult before applying private client facts.
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.
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.
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.
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.
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.
Crystal footprint vs raw-accumulating at the final turn. Crystal: 2,684 tokens. Raw accumulating: 59,689 tokens — 22× larger and growing every call.
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.
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.
Workspace subscription for firms that want a shared legal reference layer before connecting private matter material.
8,000,000 credits/mo
Up to 25M raw-token-equivalent shared legal reference coverage.
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.
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.
A legal-authority map for separating binding authority, persuasive authority, procedural posture, and background context.
A public regulatory reference layer for agency positions, obligation types, and compliance language used before matter-specific advice.
Drafting reference support for common legal work products without exposing templates, prompts, or private firm playbooks.