We use mechanistic interpretability findings to detect framing drift in information environments.
A literary translation paradigm establishes that content representations peak at ~50% of network depth. Subspace activation patching demonstrates that a single direction in 8,192-dimensional space captures content-specific causal effect. Content directions are pair-specific and orthogonal, with mean pairwise cosine similarity of +0.041.
Framing similarity peaks at layer 25 (~31% of depth) while content peaks at layer 40 (~50%). A null baseline using different topics in the same registers produces the opposite pattern, confirming the signal is topic-dependent framing rather than register similarity. Validated across pharmaceutical, financial, and insurance domains.
Cross-lingual comparison replicates the three-phase trajectory from Paper I. Initial results across French, Greek, Russian, and Japanese show divergence gaps exceeding +0.36 between engines that capture semantic intent and those that translate literally.
Pinpoints material divergence from approved disclosures, classifies and measures severity against a cross-domain baseline.
Open-weight models. Deployable on local, air-gapped infrastructure.
Pharmaceutical promotional material compliance. Sentence-level drift detection between marketing copy and approved prescribing information. Validated against FDA OPDP enforcement actions. Live scans produce findings including data vintage mismatches and mechanism-of-action reframing.
Insurance compliance analysis. Detects framing drift between marketing materials and policy language across commercial lines including BOP, professional liability, cyber, and workers' compensation.
Information operation detection via two-layer residual stream extraction. Content-layer topic matching paired with framing-layer comparison identifies narrative amplification networks. Zero false positives across mainstream partisan outlets in live monitoring.
We help organizations move from frontier API dependency to private, domain-specialized AI on infrastructure they own. Open-weight models. On-premise hardware. Data that never leaves the building.
Audit your current AI stack and API spend. Identify which tasks move to open-weight models, which stay on frontier APIs, and where the cost crossover makes sense. Typical outcome: 60–80% reduction in inference costs with equivalent or better performance on domain-specific tasks.
Supervised fine-tuning on your operational data. We build training datasets from your actual documents, fine-tune open-weight models using QLoRA, and validate against your workflows. The resulting model speaks your language, knows your formats, and runs on your hardware.
Structured data extraction from operational documents — bills of lading, contracts, compliance filings, claims, invoices. OCR, retrieval-augmented generation, and fine-tuned extraction models built for your specific document types.
A configured, ready-to-run inference box loaded with your fine-tuned models. Plug it in, process documents, no cloud dependency. We handle hardware selection, model deployment, integration with your existing systems, and ongoing support.
Small models on local hardware for high-volume routine tasks. Larger models via low-cost inference APIs for complex reasoning. Automatic routing between tiers based on task complexity. Predictable costs, maximum capability, data stays on-premise where it matters.
Production-ready hardware preconfigured with your domain model, RAG pipeline, and web interface. We build it, test it, and ship it. Your team plugs it in. Updates delivered as model files — no maintenance required.
Engagements typically run 2–5 weeks from assessment to deployed system. We work on-site or remote depending on data sensitivity requirements.