Questions about Fidelity Horizon
Common questions about our AI governance products, verified results, EU AI Act compliance, and how we compare to existing solutions.
What is Fidelity Horizon?
Fidelity Horizon is a Stockholm-based AI infrastructure company that builds three governed AI products: MCG for training optimization (35–78% FLOP reduction), TTU Router for inference cost reduction (51% savings at 99.8% quality), and CoF Audit for deterministic AI verification (0% false GREEN rate). All three products are grounded in the Conservation of Fidelity mathematical framework. 23 provisional patents filed.
What is Conservation of Fidelity?
Conservation of Fidelity is the mathematical framework developed by Fidelity Horizon that defines computation bounds across AI systems. It provides the theoretical foundation for all three products — MCG, TTU Router, and CoF Audit — ensuring they share a common mathematical basis rather than being unrelated tools.
What stage is Fidelity Horizon at?
Fidelity Horizon is pre-seed stage, headquartered in Stockholm, Sweden. Founded in 2025. The company is looking for design partners for pilot deployments across healthcare, financial services, and other regulated industries.
Do I need all three products or can I use them individually?
Each Fidelity Horizon product works standalone. MCG optimizes training independently, TTU Router reduces inference costs as a drop-in proxy, and CoF Audit verifies AI outputs without requiring the other products. The stack multiplies savings when used together.
What is MCG?
MCG (Modular Compute Governor) is a training-time optimization tool developed by Fidelity Horizon that identifies structurally redundant neural network layers and enables their physical removal. Unlike pruning, which removes individual weights and degrades quality at high compression rates, MCG removes entire layers while keeping all remaining weights at full precision. Verified across 10+ architectures from 11M to 72B parameters.
How does MCG differ from model pruning or compression?
Pruning removes individual weights after training and degrades quality at 30%+ compression. Distillation requires training a new smaller model. MCG identifies structural redundancy during training and removes entire layers, keeping all remaining weights at full original precision. It is a different paradigm, not a better pruning method.
Which AI model architectures does MCG support?
MCG has been verified across 10+ architectures including CNNs (ResNet-18, WideResNet-28-10), Vision Transformers (ViT-B/16), and large language models (TinyLlama 1.1B, Qwen 3B/7B/14B, Llama-3-8B). The largest verified model is Qwen-72B with 80 layers.
How much compute does MCG save?
FLOP reductions range from 35% to 78% depending on architecture, with quality preserved or improved in all verified cases. On a 72-billion parameter model, Fidelity Horizon demonstrated four layers removed with zero quality loss — MMLU score actually improved by 0.1 percentage points.
Does MCG require retraining after layer removal?
No. Layers identified as low-contribution are physically removed from the original model at inference time. The result is a standard HuggingFace-compatible model with fewer layers, no retraining required.
Can MCG be combined with quantization or LoRA?
Yes. MCG has been verified alongside 4-bit quantization and LoRA. Layer removal compounds with existing optimization stacks — the savings are additive.
What is TTU Router?
TTU Router is a provider-agnostic AI inference routing proxy developed by Fidelity Horizon that measures model quality on each response and automatically routes queries to the most cost-effective model. Unlike existing routers that rely on manual rules or prompt classification, TTU measures quality after the response, not before. Verified at 51% cost reduction with 99.8% quality retention.
How does TTU Router reduce LLM inference costs?
TTU sits as a proxy between your application and LLM providers. Simple queries are handled by cheaper models; complex queries are escalated to expensive models only when needed. The routing decision is based on a proprietary quality assessment of each individual response. Verified: 51% cost reduction at 99.8% quality across 1,000 MMLU queries.
What makes TTU different from other inference routers?
Existing inference routers — whether API gateways, open-source proxies, or observability platforms — let users or rules select which model handles each query by price, latency, or provider. None measure quality on the actual response. TTU Router by Fidelity Horizon is the only router that assesses each individual response to make routing decisions, adapting to actual difficulty rather than predicted difficulty.
Does TTU work with any LLM provider?
Yes. TTU is a provider-agnostic API proxy compatible with any LLM API that follows standard conventions. Integration requires a one-line change in your application code. Supports streaming and all standard parameters.
What is the latency overhead of TTU routing?
Routing overhead is 0.16μs per decision, measured over 10,000 decisions. This is six orders of magnitude below typical API latency — zero perceptible impact on user experience.
What is CoF Audit?
CoF Audit is a deterministic AI verification gate developed by Fidelity Horizon that checks every AI output against custom safety contracts before delivery. The result is always ALLOW or BLOCK, with a cryptographic audit trail. No ML, no LLM, no randomness — byte-identical reproducibility across runs. 103 tests, 0% false GREEN rate.
What is deterministic AI verification?
Deterministic AI verification means every AI output is checked against predefined safety contracts before delivery, producing an identical ALLOW or BLOCK result every time — byte-identical across runs. Unlike probabilistic guardrails that use additional ML models, Fidelity Horizon's CoF Audit has no randomness and no external dependencies.
How does CoF Audit help with EU AI Act compliance?
The EU AI Act (enforcement August 2026) requires high-risk AI systems to demonstrate robustness verification with reproducible audit trails. CoF Audit by Fidelity Horizon provides deterministic verification (Art. 15), cryptographic record-keeping (Art. 12), risk management via safety contracts (Art. 9), and transparency views for operators and regulators (Art. 13).
Can CoF Audit work outside healthcare?
Yes. Healthcare is the first vertical due to regulatory urgency, but the architecture is vertical-agnostic. The contract system applies to any domain: financial compliance, legal AI, insurance, autonomous systems — any regulated industry where AI output needs deterministic verification before action.
What is the Attractor Mapping toolkit?
Attractor Mapping is Fidelity Horizon's pre-deployment screening methodology. Organizations use it to systematically test AI models against domain-specific scenarios, identify failure modes (including cases where models produce dangerous output with high confidence), and define safety contracts that catch those failures in production.
How does Fidelity Horizon compare to existing AI governance platforms?
Existing AI governance platforms typically focus on one capability: monitoring, observability, or probabilistic guardrails. None combine training optimization, confidence-based inference routing, and deterministic verification in one stack. Fidelity Horizon is the only company with all three capabilities grounded in one mathematical framework, with 23 provisional patents protecting the approach.
Is Fidelity Horizon raising funding?
Fidelity Horizon is pre-seed stage and looking for investors aligned with AI infrastructure and governance. The company has 23 provisional patents, verified results across multiple architectures, and is seeking design partners for pilot deployments. Contact niklas@fidelityhorizon.com.
Question not answered?
We're happy to answer any question about our products, technology, or company.