~45 KB
Artefact per Lens
Enterprise Compress · Lens-Packs
Domain-Adaptive Steering.
~45 KB. Zero runtime cost.
Lens packs teach your existing embeddings the difference between “shall” and “shall not” — deterministically, in a single matrix multiply.
Legal Lens — ProductionMedical & Financial — PlannedCustom Lens — ServiceDeterministic~45 KB Artefact
Section 01
The liability trap.
“Embedding-based search has a known failure mode: clauses with opposite meaning (shall vs shall not) appear too close in embedding space. A retrieval system returning contradictory clauses as ‘similar’ creates compliance mistakes, hallucination amplification, and costly review.”
Section 02
Lens-Packs: the product catalog.
| Lens | Status | Use Case |
|---|---|---|
| Legal Lens | Production | Contract analysis, negation-safe retrieval, liability clause detection |
| Medical Lens | Planned Q3 2026 | Clinical NLP, drug-interaction safety, diagnosis ambiguity |
| Financial Lens | Planned Q4 2026 | Regulatory filing analysis, risk-clause separation |
| Custom Lens | Professional Service | Bespoke domain-tuning for your specific corpus and policy boundary |
Section 03
Compared with the industry.
LLM verification, cross-encoder reranking and constraint engines all solve adjacent problems — none deterministically, none at zero runtime cost.
| Approach | Cost / Query | Latency | Deterministic |
|---|---|---|---|
| LLM Verification (GPT-4 Reranker) | $0.01–$0.10 | high | no |
| Cross-Encoder Reranking | $$ | 50–100 ms | no |
| Constraint Engines (symbolic rules) | infra-complex | medium | partial |
| AQEA Lens | ~45 KB baked-in | ~0 ms | fully |
Section 04
Deployment: drops in above your embedder.
~0 ms
Runtime overhead per query
CPU-only
No GPU required; trainable in minutes
A single trained Lens artefact contains both Focus Mode (boost relevant signals) and Shield Mode (suppress confusable opposites). Switch at runtime per query type, no re-deployment.
Interested in a Lens-Pack?
Legal Lens is in production today. Custom Lens engagements typically start with a one-page corpus brief.