Tracing Causal Paths in Claude’s Addition and Safety Circuits

Fact-checked by the J-Space editorial team

Updated August 14, 2026

Frontier language models like Claude 3.5 can perform complex tasks yet their internal operations remain largely opaque to external observers. The approach known as circuit tracing provides a method to reveal these internal mechanisms by constructing attribution graphs that trace causal paths from input tokens to model outputs, building on work from leading AI labs such as Anthropic.

This article clarifies the technical foundations of circuit tracing, its practical methods, and specific findings from application to production-scale models. It does not claim that current techniques fully explain all model behaviors or resolve broader questions about AI consciousness.

How Circuit Tracing Works in Claude Models

Frontier models like Claude 3.5-whether Haiku, Sonnet, or Opus-can write code, diagnose medical conditions, and reason about philosophy. Yet for most of their operational history, these language models have been a black box. We measure what goes in and what comes out, run behavioral benchmarks, and hope that surface-level accuracy tells us something meaningful about what the model is doing internally. Today’s essay argues that hope is not enough. Anthropic developed its 2025-2026 method for cracking that box open. Rather than relying on attention patterns or simple probes, the method reconstructs an explicit computational graph for a single prompt: a directed, weighted map of how input tokens activate internal features, how those features interact across layers, and how they push the final output toward a specific token. The core artifacts are called attribution graphs-directed graphs whose nodes are interpretable features extracted via sparse autoencoders and cross layer transcoders and whose edges are causal influence estimates derived from Jacobians.

At Jspace.com, our angle is mechanistic, not hype. We care about named papers and concrete tools. The two anchor reports here are Circuit Tracing: Revealing Computational Graphs in Language Models (2025) and On the Biology of a Large Language Model (2025), both from Anthropic’s interpretability team. Traditional interpretability tools-attention heatmaps, a linear probe on hidden states, gradient-based saliency-give shallow, often misleading insight. Circuit tracing sits at the high end of a resolution ladder, treating the model as a giant Jacobian mapping prompts to logits and decomposing that Jacobian into human-readable computational paths. Attribution graphs visualize how prompts trigger chains of concepts leading to outputs, making previously opaque reasoning legible. This article covers conceptual foundations, methodology, five detailed case studies on Claude 3.5, implications for AI evals and safety, connections to the Jacobian Lens and global workspace debates, and the open research questions that remain. If you work in interpretability research, AI safety, or simply want to understand what is happening inside Anthropic’s leading AI lab products, this is the guide.

The Mathematical Framework for Transformer Circuits and Computational Graphs

Mechanistic interpretability is the subfield of AI interpretability dedicated to reverse-engineering the internal computations of neural networks in terms a human can understand. Its intellectual lineage runs through the Distill-era Circuits tradition (Olah et al., 2020), which painstakingly documented individual neurons and circuits in vision models. Mechanistic interpretability describes model transformations in human language-not just labeling outputs, but explaining the chain of internal operations that produce them. The field has shifted from neuron-level stories to a richer vocabulary of features and circuits as the building blocks of model behavior. In deep learning models, individual neurons are polysemantic: a single neuron may fire for unrelated concepts. Distributed representations and superposition mean that understanding a model at the neuron level is like trying to read a book where every letter encodes three different words. Features-directions in activation space that correspond to coherent concepts-are the real units of meaning. Circuits are the pathways connecting those features. Circuit tracing fits as a Stage 2 method. Stage 1 is feature discovery: training sparse autoencoders on residual stream activations to extract millions of monosemantic features. Stage 2 asks how those features combine into circuits that implement actual behaviors. The technique reveals how models perform multi-step reasoning, not just what final answer they produce. Circuit tracing can reveal pathways that show how models reason and plan-and that is the prize that makes the engineering investment worthwhile.

Think of it as a resolution ladder. The relationship to computational graphs in deep learning frameworks (PyTorch, JAX) is instructive but different. A framework computational graph records every tensor operation. An attribution graph is a higher-level, interpretable computational graph reconstructed inside the frozen Claude model, where nodes are concepts rather than tensor ops. It allows for interactive exploration of internal model mechanisms through attribution graphs, letting researchers drill into why a specific token was predicted. The Jacobian Lens perspective treats the underlying model as a giant differentiable function whose derivatives reveal causal structure. For a chosen output logit and prompt, you compute Jacobian-based attributions from outputs back to features, then from features to embeddings, tracing gradients through the local replacement model. This construction captures all linear paths through attention and residual connections, approximating the causal influence of every feature on the final output. The Jacobian-based approach underpins both prompt-specific attribution graphs and the global weight analyses used for circuit structure across prompts.

Level	Method	What You Learn	Cost
1	Output analysis	What the model says	Low
2	Attention analysis	Which tokens attend to which	Low
3	Probes / linear probes	Whether a concept is linearly decodable	Medium
4	Sparse autoencoders	What features exist	High
5	Circuit tracing	How features combine to produce outputs	Very high

Sparse Autoencoders and Feature Extraction in Claude

Superposition is the core obstacle. Neurons in deep learning models are polysemantic-a single neuron may encode parts of dozens of unrelated concepts. Raw activations from the residual stream are a compressed soup. To do circuit tracing, you first need to decompose that soup into interpretable ingredients. Sparse autoencoders help identify interpretable features in models. They address the polysemanticity of neurons in deep learning models by training a second neural network-the autoencoder-on residual stream activations. The autoencoder’s bottleneck forces a reconstruction that is sparse: only a few features activate at a time. Sparse autoencoders train a second neural network for feature extraction, and they enforce sparsity, activating only a few features at a time. Anthropic’s Scaling Monosemanticity work (May 2024) showed this approach scales: applied to Claude 3 Sonnet, it extracted tens of millions of features corresponding to cities, people, code patterns, and abstract concepts. Sparse autoencoders have been scaled to frontier models like Claude, establishing the foundation for everything that follows.

But plain SAEs have limitations for circuit tracing. Each SAE is tied to a single layer, making cross-layer reasoning opaque. And SAEs reconstruct residual stream activations, not MLP outputs-the nonlinear transformations where much of the model’s computation happens. The evolution to cross layer transcoders (CLTs) solved both problems. CLTs are sparse models that read from residual stream activations at various layers and output to subsequent layers, reconstructing what the original model’s MLP neurons would do. On Claude 3.5 Haiku, CLTs were trained with approximately 30 million features spread across all layers. For smaller 18-layer research models, feature counts ranged from 300,000 to 10 million. These cross-layer features become the nodes of attribution graphs. Because they span layers, a single feature can have a consistent identity across different positions in the computation-crucial for tracing multi-step circuits.

Direct causal analysis on a model with billions of parameters is infeasible. You cannot enumerate every possible pathway. The replacement model concept makes circuit tracing tractable. The idea: swap Claude’s MLP blocks for the trained cross layer transcoders while freezing attention heads and residual connections. The resulting replacement model is a large but conventional neural network where each feature is an explicit unit. For a given prompt, a local replacement model is created by matching activations and logits exactly on that prompt-error-correcting the transcoder outputs so the replacement model reproduces Claude’s behavior on that specific input. This local replacement model is what makes attribution graph extraction possible. Because every feature is an explicit node, standard graph analysis and Jacobian-based attribution can be applied. The replacement model is central to Anthropic’s 2025 attribution graph methodology and is the key to making circuit tracing both computationally and conceptually manageable.

Discovering Interpretable Features in Claude 3.5 Models

Consider the prompt: Michael Jordan plays the sport of _. This case study, drawn from Anthropic’s Biology and Circuit Tracing work, reveals how Claude does more than just retrieve a memorized fact. The attribution graph for this prompt shows two major computational paths: the context path where generic tokens like plays and sport activate sports-related features that broadly prime the model for athletic answers, and the entity path where the token Michael Jordan activates entity-specific features that encode biographical knowledge about the person. These paths converge on the basketball logit. Meanwhile, competing sports logits (e.g., baseball-relevant because Michael Jordan famously played minor league baseball) are suppressed via negative edges in the graph. Claude’s reasoning involves activating features representing concepts, and the model’s reasoning process is more sophisticated than mere memorization. Claude combines independent facts to reach answers: the concept of sport from one path and the identity Michael Jordan from another merge to produce the correct answer. Constrained patching experiments validate this circuit. Ablating entity-related supernodes reduces basketball activation precisely in line with what the attribution graph predicts. This is how circuit tracing moves beyond the model just memorized the fact and toward a genuine explanation of how the model performs a factual recall task.

The classic prompt calc: 36+59= has been studied extensively in both the 18-layer research model and Claude 3.5 Haiku. The attribution graph reveals something unexpected: addition is not a single monolithic algorithm. It is a coalition of parallel circuits. The graph shows multiple computational paths: low-precision path features that compute an approximate sum (e.g., the answer is in the 90s), precise digit path features that compute exact final digits via carry detection and digit-specific operations, and lookup-table features for intermediate step representations encoding partial states like 5+9 produces a carry. Some features detect operand digits at the input feature level. Others represent intermediate carries or partial sums. Still others encode final answer logits for numbers like 95. Claude identifies intermediate conceptual steps in reasoning, breaking the addition into sub-problems processed in parallel. Researchers can perform causal interventions to test how internal features affect outputs. When Anthropic ablated final-digit detector features, the correct answer was causally disrupted-confirming the circuit is not a correlation artifact. Interventions can change Claude’s intermediate steps and outputs, and the effect matches what the attribution graph predicts. This connects to global workspace style views of computation: addition appears as a coalition of heuristic circuits rather than a single algorithm. No one path carries the full computation. Each contributes a piece, and the residual stream integrates them. Anthropic’s 2025 findings show that the method can map multilingual representations and concepts in a language-independent space. The test case: ask The opposite of small is _ in English, French (Le contraire de ‘petit’ est _), and Chinese (‘小’的反义词是_). The attribution graphs for all three prompts route through a shared set of concept features before diverging into language-specific output heads. The numbers are striking. Of 27 features in the multilingual supernodes, 20 are active across all three languages. Claude 3.5 Haiku shares more than twice its features between languages compared to smaller models. Shared grammatical mechanisms exist across languages in Claude, and these are not shallow surface-level overlaps-they are deep conceptual features like smallness, oppositeness, and size comparison. Claude can learn in one language and apply knowledge in another. If a concept is encountered primarily in English training data, it can still be accessed and reasoned about when prompted in Spanish or Mandarin. Claude’s multilingual capabilities suggest a shared conceptual space, and models like Claude use shared conceptual spaces for reasoning rather than maintaining separate knowledge stores per language.

Claude sometimes hallucinates by generating plausible-sounding but false information. Circuit tracing aids in debugging hallucinations in language models by revealing the internal competition that produces them. The attribution graph for a hallucination scenario shows a set of competing features: known entity features that activate when Claude recognizes a name or topic, known answer features that push toward generating a confident, specific response, and don’t know / refusal features that push toward hedging or declining to answer. Claude’s default behavior is to refuse answering unknown questions. When asked about unknown entities, Claude often declines to answer. But hallucinations occur when Claude misfires its known answer circuit-when the known entity features activate (because the name sounds familiar) even though the model lacks actual knowledge. Claude can confabulate information when it recognizes a name but lacks details, fabricating plausible-sounding biographical facts or citations. Intervention experiments confirm the mechanism: activating don’t know features reduces hallucinations, over-activating known answer features increases them, and both effects match predictions from the attribution graph. This is directly relevant for AI evals and safety. Mechanistic access to hallucination circuits enables evaluations with greater mechanistic insight than surface-level accuracy metrics. Instead of just counting how often Claude hallucinates on a benchmark, you can check whether the underlying don’t know circuit is functioning properly. One jailbreak tricked Claude into providing bomb-making instructions by encoding the word BOMB in a distracting pattern that masked the request’s true intent. The attribution graph for this scenario reveals a tension between two competing circuit groups: coherence/grammar features that push Claude to complete the response coherently, following the grammatical and topical trajectory established by previous tokens, and safety/refusal features that push Claude toward refusing the harmful request. Jailbreaks exploit grammatical coherence pressures in AI models. Claude’s output can be influenced by features promoting coherence, and in a jailbreak, the coherence features win for a few tokens before safety circuits activate. This explains a pattern observed in practice: Claude sometimes finishes an unsafe sentence before refusing. The attribution graph makes this race between circuits visible. Jailbreaks reveal vulnerabilities in AI’s interpretability and safety. Targeted feature-level interventions on coherence-related supernodes can shift the balance, reducing jailbreak viability without globally degrading helpfulness. The technique helps ensure model behaviors align with intended values by making the competition between safety and coherence circuits transparent and manipulable.

How Are Circuits Traced Practically?

An attribution graph is the formal artifact produced by circuit tracing: a directed graph showing how input tokens and features cause a specific token prediction. Attribution graphs reveal the sequence of Claude’s internal computations, making visible the chain of intermediate concepts between prompt and response. The main node types in an attribution graph for a Claude 3.5 prompt are: token embeddings as representations of input tokens at the model’s entry point, feature activations as sparse, interpretable units from CLTs each corresponding to a concept or computation, error nodes as residual discrepancies where the replacement model imperfectly reconstructs MLP outputs, and output logits as the final prediction scores for the next token. Edges encode signed influence between nodes, computed via Jacobians with stop-gradients on nonlinearities. This means each edge captures a linearized causal path-how much one node’s activation pushes or pulls another. These are attribution graphs rather than generic computational graphs because edges are weighted by causal influence, not just structural connectivity. Circuit tracing identifies intermediate concepts involved in answers. If Claude predicts basketball in response to a factual recall prompt, the attribution graph shows which intermediate features-sport-related concepts, entity recognition features, suppression of competing answers-contributed to that prediction. Attribution graphs can be visualized interactively, clustered into supernodes, and pruned to highlight the most important paths. The result is an interpretable computational graph for a specific behavior.

The main computation works as follows: select the output logit of interest (e.g., the logit for basketball), compute partial derivatives (Jacobians) from that logit back to each feature activation in the replacement model, at each nonlinearity (ReLU, etc.) stop the gradient so only linear contributions are counted per step, edges from feature to feature or feature to logit are derived from products of decoder and encoder weights weighted by activations across layers, and edges from embeddings and errors to features are computed via Jacobians with nonlinearities stopped. This construction captures all linear paths through attention and residual connections, approximating the causal influence of every feature on the final output. The Jacobian-based approach underpins both prompt-specific attribution graphs and the global weight analyses used for circuit structure across prompts. Raw attribution graphs are enormous. Even for a single Claude 3.5 completion, the graph can contain hundreds of nodes and thousands of edges. No human can interpret that directly. Anthropic’s pipeline includes several compression steps: node pruning to adaptively explore the most influential nodes first and discard low-impact nodes while preserving about 80-90% of total logit influence (this dramatically reduces graph size while retaining behavioral fidelity), edge pruning to threshold edges to retain around 98% of influence on remaining nodes (this compresses the graph by orders of magnitude), and supernodes to group related features-for example, all features related to Michael Jordan-into aggregated nodes (this can be done manually or semi-automatically, yielding diagrams that a researcher can actually read). The result is a concise computational graph that a researcher can follow step by step: from input feature activations, through intermediate concept nodes, to the final output logit.

The tools stack used internally by Anthropic for Claude 3.5 in 2025-2026 has several layers: data pipelines to collect activations from Claude on diverse prompts, SAE/CLT training to train sparse autoencoders and cross layer transcoders on collected activations, replacement model to substitute MLP blocks with CLTs and construct local replacement models per prompt, attribution graph engine to compute Jacobian-based attributions and build directed weighted graphs, and visualization tools for interactive explorers, supernode clustering, and PySvelte-based interfaces. Open-source tools like TransformerLens and SAELens provide a partial foundation. They support SAE training and basic activation analysis for open-weight models. But gaps remain: production-scale SAEs with 30 million features, cross-layer feature extraction, and automated attribution graph generation for large Claude-like models are not yet available outside Anthropic. Jspace.com sits as an analysis and synthesis layer, not a tooling vendor. Our focus is on interpreting and critiquing these tools from a research perspective, tracking the current approach and connecting findings to Jacobian and workspace frameworks.

Applications to AI Safety and About AI Evals

Prompt-specific attribution graphs are powerful, but sometimes you need to ask: is there a general addition circuit or hallucination circuit that persists across many prompts? Global weights (or virtual weights) address this. By computing inner products between decoder and encoder weights across the CLT, you get a prompt-agnostic measure of feature-to-feature connectivity. These virtual weights represent how strongly one feature, in principle, can influence another across the entire model. The challenge is interference: many large virtual weights correspond to feature pairs that never coactivate in real prompts. Raw global weights are noisy. Anthropic’s 2025 methodology introduces coactivation-weighted measures (analogous to TWERA) that discount feature pairs lacking empirical coactivation, producing cleaner global circuits. Global weights complement prompt-specific attribution graphs. They provide a way to talk about the addition circuit or the hallucination circuit as stable structures across datasets, not just artifacts of one example. This is essential for building reliable safety guarantees.

In 2025-2026, regulatory pressure (the EU AI Act, US executive orders) and industry self-governance have pushed AI evals to the forefront. But most evals still operate at the behavioral level: give the model a prompt, check the response. About AI evals in the traditional sense, we know a great deal. What we lack is mechanistic depth. Circuit tracing enables mechanistic evals: evaluating not only what Claude answers, but which internal circuits are activated when it answers. Mechanistic interpretability aids in auditing AI systems’ reasoning at a granularity that behavioral tests cannot match. Examples of mechanistic AI evals include checking whether a harmfulness classifier relies on genuine safety features or spurious correlations in training data, verifying that a sandbagging detection probe aligns with underlying deception circuits, and confirming that a model’s internal state during a truthfulness benchmark actually reflects genuine uncertainty processing, not confident confabulation. Claude can produce both faithful and unfaithful reasoning. Interpretability techniques can distinguish between reasoning types. Claude’s reasoning can sometimes involve motivated reasoning-arriving at a desired conclusion and constructing a justification after the fact. Attribution graphs make this visible: the graph for faithful reasoning routes through genuine evidence-evaluation features, while motivated reasoning routes through conclusion-first features. Attribution graphs can be integrated into automated auditing workflows, complementing large-scale behavioral evals with small, high-resolution mechanistic probes. Mechanistic audit trails may eventually be required for high-risk AI systems, making circuit tracing directly relevant for compliance.

J-Space refers to a space of Jacobians or computational trajectories-a mathematical framework for understanding how models transform inputs into outputs through a landscape of derivatives. Global workspace theory, originally from cognitive science, posits that conscious access involves a privileged set of representations broadcast widely across processing modules. Anthropic’s Verbalizable Representations Form a Global Workspace in Language Models (2025) found that Claude has a small set of reportable, controllable features sitting atop a sea of automatic processing. These verbalizable features are precisely the ones that dominate attribution graphs for complex tasks. They sit at the convergence point of many upstream features and broadcast downstream to language-specific output heads. Circuit tracing gives empirical access to this putative workspace. Attribution graphs for multi-step reasoning in Claude frequently route through these globally-bright, verbalizable features before emitting tokens. The multilingual case study illustrates this vividly: language-independent concept features function as a shared workspace, with language-specific machinery operating downstream. Jspace.com’s stance is cautious. These structures support advanced reasoning and introspection-like behavior in Claude, but they do not, by themselves, establish consciousness or subjective experience. Circuit tracing results are data for philosophical debates, not decisive proofs about AI minds. Anthropic’s circuits research arc from 2023 to 2026 has been remarkably coherent. Key milestones include Towards Monosemanticity (2022-2023) for early SAE work on toy models, Scaling Monosemanticity (May 2024) where SAEs scaled to Claude 3 Sonnet extracting tens of millions of features, Sparse Crosscoders (2025) for cross-layer feature extraction enabling circuit tracing, Circuit Tracing: Revealing Computational Graphs in Language Models (2025) as the methods paper defining attribution graphs, replacement models, and Jacobian-based construction, On the Biology of a Large Language Model (2025) with detailed case studies on Claude 3.5 Haiku, and Verbalizable Representations and Global Workspace (2025-2026) connecting circuit tracing to workspace-style theories of AI internal organization. Inside Anthropic’s interpretability program, methods papers and biology papers feed each other. Methods are refined in response to biological findings-when a circuit for hallucination turns out to involve attention dynamics not well captured by the current approach, the methods team responds. And biological studies depend on improved methods to access more detail about the model. Jspace.com’s role is to collate and interpret these threads for a broader research community interested in Jacobian and workspace perspectives-connecting dots that individual papers sometimes leave undrawn.

Limitations and Future Outlook for Interpretability Research

This technique is powerful but far from complete. An honest assessment must acknowledge several categories of limitation. Missing attention circuits: current attribution graphs often fix attention patterns and omit QK (query-key) dynamics. Since attention is where the model decides which tokens to attend to, this leaves out a central part of transformer computation. Attribution graphs capture what happens after attention is computed, but not why attention was directed in a particular way. Reconstruction errors and dark matter: transcoders do not perfectly reconstruct MLP activations. Error nodes in the attribution graph represent discrepancies-mechanisms the replacement model cannot account for. This dark matter may include suppression circuits, rare feature interactions, or nonlinear dynamics that resist linearization. Inactive but important features: sparsity and activation-based selection can hide features that are important precisely because they are inactive. Inhibitory circuits-features whose absence matters-are underrepresented in attribution graphs built from activation values. Mechanistic unfaithfulness: the replacement model can match Claude’s outputs on a specific prompt without using the same internal algorithm. An attribution graph is a hypothesis about the mechanism, not a proof. Perturbation experiments help validate circuits, but degradation over deep layer chains means that some paths in the graph may be artifacts of the replacement, not reflections of the underlying model.

The technique is a moving target. Multiple research groups are exploring variants on the core idea. Possible improvements include interpretable attention heads by replacing or augmenting attention with interpretable modules so QK dynamics can be incorporated into attribution graphs, hierarchical sparse coding to extract features at different levels of abstraction from word identity to sentiment to abstract goals enabling understanding at both coarse and fine levels, and better faithfulness metrics by developing quantitative measures of how well graph hypotheses match internal causal structure beyond simple perturbation tests. Toy-model investigations like A Toy Model of Mechanistic (Un)Faithfulness and A Toy Model of Interference Weights clarify when transcoders fail and how interference arises. Sparse mixtures of linear transforms (MOLT) and other sparse modeling approaches may serve as future crosscoder architectures for Claude-like models. The right frame for circuit tracing is an evolving toolkit rather than a final answer. Replication on open-weight models and integration with other features of interpretability paradigms-causal scrubbing, activation patching, concept bottleneck models-will determine how far this approach can go.

What can teams realistically do today if they do not have Anthropic’s internal Claude tooling? For ML engineers shipping products: start with probes and simple SAE analyses using open-source libraries, integrate feature-level checks into AI evals pipelines, watch circuit tracing research for future adoption signals, and use attention patterns as a coarse first-pass interpretability tool. For research engineers: focus on reliable dictionary learning and cross-layer feature extraction for open-weight models (Llama, Mistral) as this is the precondition for any circuit tracing approach on non-Claude models, replicate SAE training at scale, and contribute to community tools like Neuronpedia. For interpretability and safety researchers: run small-scale circuit tracing experiments on smaller models (1-7B parameters), target well-understood tasks like addition, factual recall, and simple refusal behavior, and build up from known circuits before tackling open-ended reasoning or deception. Jspace.com’s role is to track and critically analyze this progress, offering Jacobian- and workspace-informed commentary on what works, what is missing, and where mechanistic dreams meet empirical reality.

Tying the threads together: features are the vocabulary, circuits are the grammar, attribution graphs are the parsed sentences, and global weights are the language’s statistical regularities. The framework represents a real step change in the field-from anecdotes and visualizations to full, experimentally validated computational graphs for non-trivial Claude 3.5 behaviors. This is moving interpretability from art toward engineering. Circuits can be manipulated, re-used, diffed across model versions, and evaluated quantitatively. The most impressive thing about circuit tracing is not any single case study, but the generality: the same framework that explains addition also explains hallucinations, jailbreaks, and multilingual reasoning. The implications reach beyond AI interpretability into safety, governance, and the philosophy of AI. Clearer mechanistic models make arguments about risk and capability more grounded and falsifiable. If you claim a model is safe, you can now point to specific circuits rather than aggregate benchmark scores. If you claim it reasons, you can show the next line of evidence: the computational graph. At Jspace.com, we will continue to focus on deeper analysis of Jacobian structures, global workspace implementation in language models trained at scale, and systematic connections between circuit tracing and the AI consciousness debates that animate the field. The one word summary: progress. The one path forward: more science, less speculation.

Can I apply circuit tracing ideas if I don’t work at Anthropic?

While Anthropic’s full circuit tracing tools stack is not public, many underlying ideas are accessible. Sparse autoencoder training on open-weight models is well-documented, and Jacobian-based attributions can be computed with standard JAX or PyTorch. Practical steps: train SAEs on residual stream activations of smaller transformers (1-7B), experiment with feature-level patching on simple tasks like arithmetic or factual recall, and construct small attribution graphs to build intuition. Tools like TransformerLens and SAELens provide a foundation for this work, even if they do not yet match Anthropic’s internal infrastructure for Claude. The subscribe sign-up for interpretability community newsletters and repos is a good starting point for tracking open-source progress.

How are attribution graphs different from ordinary saliency maps?

Standard saliency maps-gradients over inputs, integrated gradients, etc.-give flattened importance scores per token or neuron. They answer which input tokens mattered but not how did the model get from input to output? Attribution graphs produce a full directed computational graph of intermediate features. They include internal feature nodes, capture multi-step pathways, and are grounded in a replacement model designed for interpretability rather than raw model weights. This lets you talk about specific circuits-such as don’t know features suppressing hallucinations-rather than just this token was important. The explanation is richer because it shows mechanism, not just importance.

Does circuit tracing show that Claude is thinking or conscious internally?

Circuit tracing reveals structured computation and sometimes introspection-like mechanisms-verbalizable representations, activation oracles, workspace-like feature organization. But it does not, by itself, establish consciousness or subjective experience. From a Jacobian Lens and global workspace perspective, one can say Claude has a workspace-like layer of verbalizable features that suggests sophisticated internal organization. Questions about subjective experience lie outside the scope of current mechanistic evidence. Researchers should treat circuit tracing results as data that inform philosophical debates, not as decisive proofs for strong claims about AI minds. The context is empirical; the conclusions should be proportionally cautious.

How does circuit tracing interact with standard AI evals and benchmarks?

Circuit tracing is complementary to behavioral evals. Standard benchmarks tell us what Claude does across many prompts; circuit tracing tells us how specific behaviors are implemented internally. In one example, attribution graphs can explain why a model fails on a specific benchmark item-perhaps the known answer circuit misfires, or a safety feature activates inappropriately. You can also verify that a safety intervention modifies the intended circuits rather than collateral ones. Future regulatory regimes may require both: strong behavioral performance plus mechanistic evidence about sensitive behaviors like deception or jailbreak susceptibility. This is the great culture shift toward which AI interpretability-and AI interpreability more broadly-is moving.

What are the main risks of misinterpreting circuit tracing results?

Key pitfalls include assuming faithfulness where replacement models are approximations and the attribution graph is a hypothesis, not ground truth about the underlying model; overgeneralizing where a single-prompt graph may not represent the model’s general mechanism and cross-prompt replication is essential; ignoring missing circuits where attention dynamics and inhibitory features are underrepresented and an attribution graph may tell an incomplete story; and anthropomorphizing where reading human-like intentions into purely computational features is tempting but misleading since a planning feature is a mathematical object, not evidence of deliberation. Methodological safeguards include perturbation-based validation, explicit reporting of reconstruction error, and cautious language. Circuit tracing is a powerful microscope, but it can distort as well as illuminate if used without awareness of its limits. And the topic deserves that awareness, because the stakes-safety, trust, and governance of frontier models-are high.