Prompt size at turn 12
50× smallerCrystal footprint vs raw-accumulating at the final turn. Crystal: 1,093 tokens. Raw accumulating: 54,324 tokens — 50× larger and growing every call.
Insurance Carriers plan
Carriers need policy memory before claim judgment.
Buy an insurance-carriers library built around shared coverage, claims, and regulatory context before adding private program material.
Multi-turn benchmark
Compact through 12 turns. No degradation detected.
Three arms, same model (gpt-4.1-mini), same question set — no context, raw-accumulating, hosted crystal. At turn 12 the crystal sent 1,093 tokens; the raw arm sent 54,324 — 50× more, compounding every call.
Instruction retention
All arms retainedA formatting rule planted at turn 1 was checked deterministically every subsequent turn. No-context retained it through all checkable turns. Crystal: 1 / 1 checkable turns held.
Answer quality — 12 turns
No degradation detectedAll 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, Insurance Carriers 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
$4,500/mo for Insurance Carriers.
Workspace subscription for carriers that need a shared policy, claims, and regulatory reference layer before private program material is connected.
Monthly allowance
6,000,000 credits/mo
Up to 20M raw-token-equivalent shared insurance reference coverage.
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.
Coverage Language
A policy-language reference layer for separating coverage concepts, exclusions, endorsements, and common interpretive questions.
- Policy-form clause purpose mapping
- Coverage and exclusion language cues
- Endorsement relationship checkpoints
Claims Operations
A claims-context layer for triage language, documentation expectations, and recurring handling patterns.
- Claims intake issue classification
- Reserve and exposure vocabulary alignment
- Documentation sufficiency review cues
Regulatory Change
A public regulatory reference layer for bulletins, market conduct topics, and state-facing compliance language.
- State bulletin topic taxonomy
- Market conduct obligation cues
- Filing-language consistency review