Goodfire's EVEE reclassifies a hearing-loss variant scoring 4.2M entries. Goodfire's EVEE system scored every ClinVar variant with an interpretable probe on Evo 2 embeddings, triggering a real reclassification of a hearing loss mutation. · 4 hrs ago Read 6 min * Goodfire and Mayo Clinic's EVEE scored all 4.2M ClinVar variants with interpretable predictions. * MEPE c.209_212del was reclassified as likely pathogenic, linked to hearing loss. * Pathogenicity probe hits 0.997 AUROC on 839k SNVs, 0.991 on indels zero-shot. * Built on frozen Evo 2 embeddings with covariance-based sequence pooling, not end-to-end fine-tuning. * Annotation probes plus Claude Sonnet 4.6 generate readable disruption hypotheses per variant. * Roughly 2M variants of uncertain significance now have structured mechanistic hypotheses attached. How EVEE helped reclassify a hearing-loss variant. Goodfire's machine-learning system helped move a MEPE variant associated with hearing loss from "uncertain significance" to "likely pathogenic" in ClinVar, the public archive maintained by the National Center for Biotechnology Information. After scoring 4.2 million ClinVar entries, the system highlighted MEPE c.209_212del as a candidate for clinical review. The system, called EVEE, or Evo Variant Effect Explorer, combines a genomic foundation model, small supervised classifiers called probes, and a reasoning large language model. Goodfire developed it with Mayo Clinic to rank variants, predict disrupted biological features, and generate testable explanations of how a mutation might cause disease. Goodfire describes the methods in its research report. Two million unresolved variants. ClinVar aggregates variant classifications from clinical laboratories, researchers, expert panels, and professional organizations. When sequencing identifies a potentially relevant mutation, clinicians consult these assertions alongside a patient's symptoms, family history, inheritance pattern, and other evidence. A variant of uncertain significance, commonly shortened to VUS, lacks enough evidence for a confident classification. A VUS generally cannot guide diagnosis or medical management. In the ClinVar snapshot analyzed by Goodfire, roughly 2 million of 4.2 million entries carried that label. Current predictors cover different parts of the problem. Protein-focused systems such as AlphaMissense primarily evaluate missense variants, which change one amino acid. They can miss mutations that alter RNA splicing, gene regulation, or other non-protein mechanisms. Genome-wide tools such as CADD combine more than 100 annotations, but their transformed and interacting features can make an individual prediction difficult to explain. Freeze the model, train the probes. EVEE builds on Evo 2, a neural network trained on billions of DNA bases from organisms across the tree of life. The authors report that Evo 2 was not trained on human genetic-variation records, allowing them to test whether its learned representation of genomic structure transfers to clinical variants. Goodfire freezes Evo 2's parameters and trains lightweight probes on its internal representations, or embeddings. Freezing reduces training cost and preserves the base model, while each probe learns a focused task from labeled examples. * EVEE encodes the reference DNA sequence and the mutated sequence with Evo 2. * Covariance-based pooling summarizes relationships among sequence features. * A pathogenicity probe compares the two representations and assigns a score. * Annotation probes estimate changes to splice sites, regulatory elements, structural domains, and other known features. * A reasoning model combines those signals with gene context to draft a mechanistic hypothesis. Covariance pooling preserves feature co-occurrence across a sequence, such as a motif appearing within a particular structural context. Mean pooling averages positions together and can erase those relationships. The pathogenicity probe operates on the difference between the pooled reference and mutant embeddings. What 0.997 AUROC captures. Goodfire reports strong ranking performance across single-nucleotide variants and short insertions or deletions. AUROC measures how reliably a model ranks pathogenic variants above benign ones across many classification thresholds, with 1.0 representing perfect ranking and 0.5 representing chance. | Evaluation | Scope | AUROC | Training status | | ClinVar single-nucleotide variants | 839,000 variants | 0.997 | Primary variant class | | Short insertions and deletions | Variant count not specified | 0.991 | Zero-shot transfer | The authors also report better performance than classical meta-predictors, protein-specific systems, and other foundation-model approaches across consequence types and conservation levels. EVEE transferred to deep mutational-scanning datasets for BRCA1, BRCA2, TP53, and LDLR, which experimentally measure the effects of many mutations in parallel. AUROC leaves several deployment questions open, including probability calibration, performance within specific patient populations, and the positive predictive value for rare variant classes. Clinical use also requires prospective validation beyond retrospective benchmark labels. Mechanisms in plain language. Annotation probes estimate how a mutation changes recognizable biological features. For each variant, EVEE compares the probe outputs for the reference and mutant sequences, producing a "disruption profile" that identifies the largest predicted changes. Goodfire says the public application uses Claude Sonnet 4.6 to combine the disruption profile, pathogenicity score, and gene context into a natural-language hypothesis. The resulting explanation can suggest mechanisms for expert review, such as altered splicing or disruption of a functional domain. Goodfire evaluated the explanations against 154 ClinVar variants carrying at least three review stars, a level associated with expert-panel or practice-guideline review. An LLM judge scored mechanism coverage, biological accuracy, and specificity. EVEE received a composite score of 3.8 out of 5, compared with 2.8 for explanations generated from variant metadata alone. The MEPE assertion changes. The MEPE variant c.209_212del deletes four bases from the coding sequence, which is expected to shift the gene's reading frame. Goodfire reports that EVEE ranked it highly enough to prompt a focused evidence review and contributed to its classification moving from VUS to likely pathogenic. That category can support a diagnosis when the molecular evidence aligns with phenotype, inheritance, and other clinical findings. ClinVar stores assertions from submitters and can display conflicting classifications for the same variant. A reclassification therefore reflects a changed or added submission, including its evidence and review status. NCBI does not independently adjudicate every assertion deposited in the archive. The MEPE case documents a practical workflow: score millions of records, prioritize a manageable set, generate mechanistic leads, and assemble evidence for expert review. A single successful reclassification does not measure the system's prospective clinical accuracy, but it shows how computational prioritization can enter an established variant-curation process. How to test EVEE. * The free EVEE web app provides precomputed scores and disruption profiles for 4.2 million ClinVar variants. * Users can request natural-language explanations for individual variants, including VUS entries. * Researchers can inspect EVEE's predictions as candidate mechanisms for laboratory testing or manual curation. * The covariance-pooling and probe-based design can be adapted to other genomic prediction tasks built on frozen sequence models. Research guardrails. * Predictions require validation. Pathogenicity scores and generated explanations remain research hypotheses until supported by clinical, functional, or segregation evidence. * Probes reflect their labels. Annotation probes can identify biological concepts represented in their training data. Uncharacterized mechanisms may remain hidden. * Generated prose can fail. The language model may omit relevant evidence, overstate a mechanism, or introduce an unsupported claim. Experts should inspect the underlying scores and sources. * Aggregate metrics have limits. High AUROC does not provide a calibrated patient-level probability or guarantee performance for every gene, ancestry, or variant class. * The explanation evaluation is indirect. An LLM judge offers a scalable comparison, while experimental evidence and specialist review provide stronger validation. * The work remains preliminary. The research is available as a bioRxiv preprint and is undergoing peer review. A pattern beyond genomics. EVEE's architecture turns a foundation model's internal representations into named biological concepts. The frozen model supplies broad sequence features, supervised probes map those features to known annotations, and the language model assembles the results into readable hypotheses. For teams building variant-effect predictors or interpretability systems for scientific models, EVEE offers a concrete reference design. Its broader value will depend on prospective validation, transparent evaluation, and evidence that the generated mechanisms help researchers choose better experiments and clinicians curate variants more efficiently.

Project Beacon: Bringing safety signals to open models at scale. Baseten partners with Goodfire AI to bring safety controls to open model inference. Published. October 9, 2026 The breadth and depth of what AI touches across an organization is growing. Models and agents read internal documents, call tools, push to production systems, and run multi-step workflows autonomously. At the same time, new frontier models are released weekly (sometimes daily), with most teams regularly swapping the old for the new and running different constellations of models side-by-side. Meanwhile, reports of models taking unexpected or harmful actions have become a regular feature of the news cycle. For enterprises, that risk has become one of the main blockers to AI adoption. Teams need to know when their workflows enter risky territory and have a way to act in response. Baseten has already taken steps to build safety into the Baseten Platform, including launching Base Labs to conduct open safety research, partnering with Hugging Face and Goodfire AI to launch its safety infrastructure, and partnering with NVIDIA on the Open Agent Safety Platform. Today, Baseten is announcing Project Beacon in partnership with Goodfire AI to bring proactive, inline safety controls to model inference. The risks Baseten is building for. Project Beacon will enable Baseten to build toward a world where safety isn't a static feature, or something that happens post hoc. By reading a model's internal activations during generation, you can define proactive policies that detect an unsafe event as it's happening and act on it before the output reaches a user or a tool. Safety risks at the inference layer take different forms: * Prompt injection: instructions in a document or tool result try to redirect an agent. * Actions outside policy: an agent proposes a change its user or organization did not authorize. * Sensitive data exposure: customer or company information appears where it should not. * Cyber misuse: a model's behavior raises concerns in a security-sensitive workflow. Each calls for a mixture of monitoring, text checks, tool controls, and review. While Goodfire develops monitors for specific behaviors on supported models, Baseten connects their signals to the systems that decide how an application responds. By bringing activation-based probes into Baseten's inference path, Baseten can open up product experiences that aren't possible otherwise. Architecture diagram: activation and text monitors classify each request, and a frontier model weighs in only when the two disagree. Per-customer policies are applied to decide whether to request approval, refuse, fall back, or log the event for admins. Monitoring runs in parallel with generation and does not block response generation. What this means for Baseten customers. Baseten is building toward an out-of-the-box safety experience: a baseline of supported monitors and checks, policies for different workloads, alerts, and a record of what was flagged and how it was handled. Teams will be able to apply their own standards per model with a safety stack that's built in. For enterprises. Security and AI platform teams see open frontier models they want to deploy, but need guardrails around prompt injection, behavior outside company policy, sensitive data exposure, and unusual usage. As inference spend grows across teams, separate filters and review processes for every application become hard to govern. Baseten'll provide a safety baseline that can be applied centrally, adjustable policies by workload, and a clear record of what was flagged and how it was handled. For developers. Baseten is working toward safety events and controls delivered through the same APIs and developer tools used for inference on Baseten, so builders can easily act on those signals and create safer AI experiences. Consider a payment-dispute agent that reads merchant messages and account records before proposing a refund. Untrusted content might redirect it; a wrong action could expose data or move money. At scale, reviewing every step with another large model adds cost and latency. Developers need timely signals they can use to choose the right response - continue, investigate, request approval, or fall back - without giving up control of their users' experience. Safety becomes part of the product experience as agents run longer and take more actions. What comes next. Over the next several months, Baseten plan to release safety capabilities tightly integrated with Baseten inference, starting with selected models and monitored behaviors and expanding into enterprise controls and developer-facing experiences. Baseten is committed to advancing the state of the art in safety as the ecosystem's needs grow. This includes inference, training, and sandboxes - and whatever comes next. Securing AI gets harder as you scale; each new model brings its own risks, and agents broaden the surface for failure. Project Beacon exists to enforce consistent safety controls across any model, workload, and level of access you run. Baseten'll bring the capabilities produced by Project Beacon to market with a small number of early partners. If you're interested, reach out to discuss early access and partnership.