{ "entries": [ { "ID": "IA-001", "Slug": "building-blocks-of-interpretability", "Name": "The Building Blocks of Interpretability", "URL": "https://distill.pub/2018/building-blocks/", "Artefact type": "Interactive article", "Visual form": "Spatial attribution; Activation grid", "Makes visible": "Features; Attribution", "Role of the visual": "Finding", "Runnable by a visitor": "Pre-set examples only", "Date published": "2018-03-06", "Year": "2018", "Link status": "Live", "Built by (people)": "Chris Olah; Arvind Satyanarayan; Ian Johnson; Shan Carter; Ludwig Schubert; Katherine Ye; Alexander Mordvintsev", "Organisation": "Google Brain", "Venue / published in": "Distill", "Sector": "Frontier lab + academia", "Country / region": "US", "Open source": "Yes", "Code repo": "https://github.com/tensorflow/lucid", "Method / technique": "Feature visualisation composed with attribution; semantic dictionaries; activation grids; spatial attribution; neuron groups", "Model(s) studied": "GoogLeNet / InceptionV1 (vision)", "Access required": "Open weights", "What it visualises": "Which neurons fire where in an image, what each one detects, and how much each contributes to the classification", "Interaction affordances": "Hover to inspect; toggle between attribution modes; explore layer by layer; composable interface demos", "Pragmatic vs Basic science": "Basic science", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Observational", "Intended audience": "Researchers; Practitioners", "Description (card)": "The founding text for interactive interpretability. Argues that interpretability techniques studied in isolation are far weaker than the interfaces you get by composing them, and demonstrates this with a set of live, hoverable interfaces over an image classifier.", "Why it matters": "Established that the interface IS the contribution — the template every entry in this ledger inherits from.", "Visual / design notes": "Distill house style: generous whitespace, inline hoverable figures, colour-coded attribution overlays. Still the visual benchmark 8 years on.", "Tags": "modality:vision; method:feature-visualisation; method:attribution; phenomenon:feature-interpretation", "Thumbnail URL": "https://distill.pub/2018/building-blocks/thumbnail.jpg", "Media files": "", "Citation": "Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K. and Mordvintsev, A., 2018. The Building Blocks of Interpretability. Distill. DOI 10.23915/distill.00010", "Related entries": "IA-002", "Confidence": "High", "Source of info": "Read in full (article + Distill metadata), 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Lucid repo link is the associated tooling, not the article source — verify before publishing. Classified Observational: the composed interfaces use gradient attribution, but the model is not altered.", "tags_parsed": { "modality": [ "vision" ], "method": [ "feature-visualisation", "attribution" ], "phenomenon": [ "feature-interpretation" ] } }, { "ID": "IA-007", "Slug": "activation-atlas", "Name": "Activation Atlas", "URL": "https://distill.pub/2019/activation-atlas/", "Artefact type": "Interactive article", "Visual form": "Scatter plot; Activation grid", "Makes visible": "Features", "Role of the visual": "Finding", "Runnable by a visitor": "Pre-rendered, view only", "Date published": "2019-03-06", "Year": "2019", "Link status": "Live", "Built by (people)": "Shan Carter; Zan Armstrong; Ludwig Schubert; Ian Johnson; Chris Olah", "Organisation": "Google Brain; OpenAI", "Venue / published in": "Distill", "Sector": "Frontier lab", "Country / region": "US", "Open source": "Unknown", "Code repo": "", "Method / technique": "Feature inversion applied to millions of averaged activations, laid out as a navigable two-dimensional map", "Model(s) studied": "InceptionV1 / GoogLeNet", "Access required": "Open weights", "What it visualises": "A navigable map of what a vision classifier has learned — millions of activations rendered as feature-inversion images and arranged so that nearby regions are semantically related", "Interaction affordances": "Pan and zoom across the atlas; switch layer; filter by class", "Pragmatic vs Basic science": "Basic science", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Observational", "Intended audience": "Researchers; General public", "Description (card)": "Renders millions of activations from an image classifier as feature-inversion images laid out on a single navigable map, so you can pan across the concepts a network has learned the way you would read an atlas.", "Why it matters": "Made a model's whole learned concept space visible at once rather than one neuron at a time — the clearest ancestor of the feature-neighbourhood maps in Scaling Monosemanticity.", "Visual / design notes": "Pan-and-zoom over a dense image grid. Still the most immediately arresting object the field has produced, and the reason 'atlas' is a live metaphor in interpretability at all — including in this project's title.", "Tags": "modality:vision; method:feature-visualisation; method:dimensionality-reduction; phenomenon:feature-interpretation", "Thumbnail URL": "https://distill.pub/2019/activation-atlas/thumbnail.jpg", "Media files": "IA-007-activation-atlas--global-mixed4d--2026-08-12.jpg", "Citation": "Carter, S., Armstrong, Z., Schubert, L., Johnson, I. and Olah, C., 2019. Activation Atlas. Distill. DOI 10.23915/distill.00015", "Related entries": "IA-001; IA-003", "Confidence": "High", "Source of info": "Read the Distill article and its citation metadata, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Open source status not verified — the article references code but this was not confirmed, so the field is left Unknown rather than guessed.", "tags_parsed": { "modality": [ "vision" ], "method": [ "feature-visualisation", "dimensionality-reduction" ], "phenomenon": [ "feature-interpretation" ] } }, { "ID": "IA-006", "Slug": "neuronpedia", "Name": "Neuronpedia", "URL": "https://www.neuronpedia.org/", "Artefact type": "Platform", "Visual form": "Dashboard; Node-link graph", "Makes visible": "Features; Circuits", "Role of the visual": "Instrument", "Runnable by a visitor": "On your own inputs", "Date published": "2023", "Year": "2023", "Link status": "Live", "Built by (people)": "Johnny Lin (creator), with community contributors", "Organisation": "Neuronpedia / Decode Research", "Venue / published in": "Self-published (web)", "Sector": "Non-profit / community", "Country / region": "US", "Open source": "Yes", "Code repo": "https://github.com/hijohnnylin/neuronpedia", "Method / technique": "Hosts SAE and transcoder feature dashboards, attribution graphs (Circuit Tracer), activation steering, probes, custom vectors, and lab-released lenses (Jacobian Lens, Natural Language Autoencoders, Assistant Axis)", "Model(s) studied": "GPT-2 Small, Pythia-70M-deduped, Gemma 2/3/4, Qwen 3/3.5/3.6, Llama 3.1/3.3, Olmo 3, GPT-OSS-20B and others", "Access required": "Open weights", "What it visualises": "Individual latents with top activations, top logits and activation density; attribution graphs on custom prompts; steering effects; over five terabytes of activations, explanations and metadata", "Interaction affordances": "Search 50M+ latents by explanation or by running text through a model; browse per-feature dashboards with permanent URLs; compile shareable lists; embed as iframe; steer with adjustable strength, temperature and seed; trace circuits on your own prompts; full API", "Pragmatic vs Basic science": "Both", "Reverse-eng vs Concept-based": "Both", "Observational vs Interventional": "Both", "Intended audience": "Mixed", "Description (card)": "The field's central public platform. It hosts feature dashboards, attribution graphs, steering and demos across dozens of open models — including the official interactive releases for Anthropic's Jacobian Lens, Natural Language Autoencoders, Assistant Axis and Circuit Tracer, and DeepMind's Gemma Scope.", "Why it matters": "The closest thing interpretability has to a public commons, and the reason a frontier-lab result can now be poked at by an outsider the week it ships.", "Visual / design notes": "One dashboard per feature, each with a permanent URL and iframe embedding. That single decision is what let it become shared infrastructure rather than one lab's internal tool.", "Tags": "modality:language; modality:general; method:sparse-autoencoder; method:transcoder; method:attribution-graph; method:steering; phenomenon:feature-interpretation", "Thumbnail URL": "", "Media files": "IA-006-neuronpedia-circuit-tracer--2026-08-12.jpg", "Citation": "Lin, J., 2023. Neuronpedia: Interactive Reference and Tooling for Analyzing Neural Networks. Software available from neuronpedia.org", "Related entries": "IA-003; IA-004; IA-005", "Confidence": "High", "Source of info": "Read the neuronpedia.org homepage in full, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Supported by Decode Research, Open Philanthropy, the Long Term Future Fund, AISTOF, Anthropic and Manifund. Hosts releases from Anthropic, DeepMind, OpenAI, EleutherAI, Apollo, Fudan OpenMOSS, Goodfire and Stanford NLP. Decision pending: whether individual hosted demos get their own rows.", "tags_parsed": { "modality": [ "language", "general" ], "method": [ "sparse-autoencoder", "transcoder", "attribution-graph", "steering" ], "phenomenon": [ "feature-interpretation" ] } }, { "ID": "IA-010", "Slug": "attentionviz", "Name": "AttentionViz", "URL": "https://attentionviz.com/", "Artefact type": "Interactive tool", "Visual form": "Scatter plot", "Makes visible": "Attention", "Role of the visual": "Instrument", "Runnable by a visitor": "Unknown", "Date published": "2023-05-04", "Year": "2023", "Link status": "Live", "Built by (people)": "Catherine Yeh; Yida Chen; Aoyu Wu; Cynthia Chen; Fernanda Viégas; Martin Wattenberg", "Organisation": "Harvard University", "Venue / published in": "IEEE VIS / TVCG 2023; arXiv", "Sector": "Academia", "Country / region": "US", "Open source": "Unknown", "Code repo": "", "Method / technique": "Joint embedding of the query and key vectors used to compute attention, projected to two dimensions across many input sequences", "Model(s) studied": "Language and vision transformers (specific models not named in the abstract)", "Access required": "Open weights", "What it visualises": "Global attention patterns across many sequences at once — query and key vectors in a shared space, so recurring head behaviours appear as spatial structure rather than per-prompt heatmaps", "Interaction affordances": "Browse heads; switch model and layer; zoom into a single head's query-key space", "Pragmatic vs Basic science": "Basic science", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Observational", "Intended audience": "Researchers", "Description (card)": "Visualises a joint embedding of the query and key vectors a transformer uses to compute attention, which makes it possible to see global patterns across many input sequences rather than one prompt at a time.", "Why it matters": "Shifted attention visualisation from single-example to distribution-level — the exact move Anthropic's HeadVis names as its closest prior work three years later.", "Visual / design notes": "The joint query-key scatter is the whole idea: spatial proximity encodes attention affinity, so a head's behaviour becomes visible geometry.", "Tags": "modality:language; modality:vision; method:attention-visualisation; method:dimensionality-reduction; phenomenon:attention-heads", "Thumbnail URL": "", "Media files": "", "Citation": "Yeh, C., Chen, Y., Wu, A., Chen, C., Viégas, F. and Wattenberg, M., 2023. AttentionViz: A Global View of Transformer Attention. arXiv:2305.03210", "Related entries": "IA-005; IA-011", "Confidence": "High", "Source of info": "Read the arXiv abstract page in full, 2026-08-06. The live demo at attentionviz.com was not inspected — client-rendered, fetch returned an empty page shell.", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "HeadVis (IA-005) explicitly names this as its closest prior work — a documented lineage between two rows in this ledger.", "tags_parsed": { "modality": [ "language", "vision" ], "method": [ "attention-visualisation", "dimensionality-reduction" ], "phenomenon": [ "attention-heads" ] } }, { "ID": "IA-002", "Slug": "towards-monosemanticity-feature-interface", "Name": "Towards Monosemanticity — feature browser interface", "URL": "https://transformer-circuits.pub/2023/monosemantic-features/index.html#setup-interface", "Artefact type": "Interactive article", "Visual form": "Highlighted text", "Makes visible": "Features", "Role of the visual": "Exhibit", "Runnable by a visitor": "Pre-rendered, view only", "Date published": "2023-10-05", "Year": "2023", "Link status": "Live", "Built by (people)": "Trenton Bricken; Adly Templeton; Joshua Batson; Brian Chen; Adam Jermyn; Tom Conerly; Nicholas L. Turner; Cem Anil; Carson Denison; Amanda Askell; et al.", "Organisation": "Anthropic", "Venue / published in": "Transformer Circuits Thread", "Sector": "Frontier lab", "Country / region": "US", "Open source": "No", "Code repo": "", "Method / technique": "Sparse dictionary learning / sparse autoencoders (SAE) over a one-layer transformer's MLP activations", "Model(s) studied": "One-layer transformer (512 neurons); SAE dictionaries ~2k-100k features", "Access required": "Internal / proprietary", "What it visualises": "Individual SAE features: top activating dataset examples, activation histograms, and the logit effects each feature has on the next token", "Interaction affordances": "Browse a feature index; click through to per-feature pages; compare a real neuron against a learned feature side by side", "Pragmatic vs Basic science": "Basic science", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Observational", "Intended audience": "Researchers", "Description (card)": "The paper that made sparse autoencoders the field's dominant method, published with a browsable interface over every extracted feature so readers could check the monosemanticity claim themselves rather than take the authors' word for it.", "Why it matters": "Turned a contested claim (features are more interpretable than neurons) into something a reader could audit by clicking.", "Visual / design notes": "Dense, utilitarian, research-grade. Feature pages are information-rich rather than designed. The contrast with Distill's polish is itself informative.", "Tags": "modality:language; method:sparse-autoencoder; phenomenon:superposition; phenomenon:polysemanticity; phenomenon:feature-interpretation", "Thumbnail URL": "", "Media files": "", "Citation": "Bricken, T., Templeton, A., Batson, J., et al., 2023. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning. Transformer Circuits Thread.", "Related entries": "IA-003", "Confidence": "High", "Source of info": "Read the article and its setup/interface section, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Author list truncated — full list on the paper. Scott Alexander's ACX piece is a good lay companion.", "tags_parsed": { "modality": [ "language" ], "method": [ "sparse-autoencoder" ], "phenomenon": [ "superposition", "polysemanticity", "feature-interpretation" ] } }, { "ID": "IA-003", "Slug": "scaling-monosemanticity", "Name": "Scaling Monosemanticity — Claude 3 Sonnet feature index", "URL": "https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html", "Artefact type": "Interactive article", "Visual form": "Highlighted text", "Makes visible": "Features", "Role of the visual": "Exhibit", "Runnable by a visitor": "Pre-rendered, view only", "Date published": "2024-05-21", "Year": "2024", "Link status": "Live", "Built by (people)": "Adly Templeton; Tom Conerly; Jonathan Marcus; Jack Lindsey; Trenton Bricken; Brian Chen; Adam Jermyn; et al.", "Organisation": "Anthropic", "Venue / published in": "Transformer Circuits Thread", "Sector": "Frontier lab", "Country / region": "US", "Open source": "No", "Code repo": "", "Method / technique": "Sparse autoencoders scaled to a production model; feature steering", "Model(s) studied": "Claude 3 Sonnet", "Access required": "Internal / proprietary", "What it visualises": "Millions of extracted features including safety-relevant ones (deception, bias, sycophancy, dangerous content), with a searchable index and feature-neighbourhood maps", "Interaction affordances": "Search the feature index; browse nearest-neighbour features; read steering examples", "Pragmatic vs Basic science": "Both", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Both", "Intended audience": "Researchers; Policy; General public", "Description (card)": "First demonstration that sparse autoencoders scale from toy models to a frontier production model, published as a browsable index of millions of features — including the Golden Gate Bridge feature that later became a public demo.", "Why it matters": "The moment interpretability stopped being a toy-model science, and the origin of the field's most famous public artefact.", "Visual / design notes": "Feature-neighbourhood visualisations are the standout: proximity in feature space rendered as an explorable map.", "Tags": "modality:language; method:sparse-autoencoder; method:steering; phenomenon:polysemanticity; phenomenon:safety-relevant-features", "Thumbnail URL": "", "Media files": "", "Citation": "Templeton, A., Conerly, T., Marcus, J., et al., 2024. Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread.", "Related entries": "IA-002; IA-004", "Confidence": "High", "Source of info": "Read the article; cross-referenced in BlueDot and ACX pieces, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Golden Gate Claude (the deployed demo) is a separate candidate entry. Classified Both: the feature index is observational; the feature-steering demonstrations are interventional.", "tags_parsed": { "modality": [ "language" ], "method": [ "sparse-autoencoder", "steering" ], "phenomenon": [ "polysemanticity", "safety-relevant-features" ] } }, { "ID": "IA-008", "Slug": "golden-gate-claude", "Name": "Golden Gate Claude", "URL": "https://www.anthropic.com/news/golden-gate-claude", "Artefact type": "Deployed model demo", "Visual form": "None", "Makes visible": "Nothing", "Role of the visual": "Experience", "Runnable by a visitor": "No longer available", "Date published": "2024-05-23", "Year": "2024", "Link status": "Live", "Built by (people)": "Anthropic Interpretability team (individual contributors not named on the post)", "Organisation": "Anthropic", "Venue / published in": "Anthropic news", "Sector": "Frontier lab", "Country / region": "US", "Open source": "No", "Code repo": "", "Method / technique": "Sparse autoencoder feature identification followed by activation steering — clamping the Golden Gate Bridge feature to roughly ten times its maximum activation value", "Model(s) studied": "Claude 3 Sonnet", "Access required": "Internal / proprietary", "What it visualises": "Nothing graphical. It makes a single feature's causal influence perceptible by letting the public converse with a model in which that feature is clamped high", "Interaction affordances": "Talk to the steered model directly, during the 24-hour window it was online", "Pragmatic vs Basic science": "Both", "Reverse-eng vs Concept-based": "Concept-based", "Observational vs Interventional": "Interventional", "Intended audience": "General public", "Description (card)": "For 24 hours Anthropic put a version of Claude 3 Sonnet online with its 'Golden Gate Bridge' feature clamped to roughly ten times maximum activation, so anyone could talk to a model that could not stop mentioning the bridge.", "Why it matters": "Turned an interpretability result into something the public could feel rather than read, and showed that a single identified feature exerts precise causal control over behaviour — not prompting, not fine-tuning.", "Visual / design notes": "Almost no conventional visual design; the artefact is the conversation. Included deliberately as the outer edge of this collection's inclusion rule.", "Tags": "modality:language; method:sparse-autoencoder; method:steering; phenomenon:feature-interpretation", "Thumbnail URL": "", "Media files": "", "Citation": "Anthropic, 2024. Golden Gate Claude. anthropic.com/news/golden-gate-claude, 23 May 2024", "Related entries": "IA-003", "Confidence": "High", "Source of info": "Read the announcement post in full, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "The demo ran for a 24-hour window and is no longer available; the announcement page survives, so Link status is Live but the artefact itself is not. WEAKEST visual element in the ledger — it qualifies on public interactivity rather than on visualisation. Flagged for Ava to decide whether the inclusion rule stretches this far.", "tags_parsed": { "modality": [ "language" ], "method": [ "sparse-autoencoder", "steering" ], "phenomenon": [ "feature-interpretation" ] } }, { "ID": "IA-011", "Slug": "talktuner", "Name": "TalkTuner — dashboard for transparency and control of conversational AI", "URL": "https://arxiv.org/abs/2406.07882", "Artefact type": "Dashboard", "Visual form": "Dashboard", "Makes visible": "User model", "Role of the visual": "Instrument", "Runnable by a visitor": "Unknown", "Date published": "2024-06-12", "Year": "2024", "Link status": "Live", "Built by (people)": "Yida Chen; Aoyu Wu; Trevor DePodesta; Catherine Yeh; Kenneth Li; Nicholas Castillo Marin; Oam Patel; Jan Riecke; Shivam Raval; Olivia Seow; Martin Wattenberg; Fernanda Viégas", "Organisation": "Harvard University", "Venue / published in": "arXiv (v3, October 2024)", "Sector": "Academia", "Country / region": "US", "Open source": "Unknown", "Code repo": "", "Method / technique": "Linear probes over internal activations to extract a 'user model', displayed live and made editable so the user can intervene on it", "Model(s) studied": "An open-source LLM (specific model not named in the abstract)", "Access required": "Open weights", "What it visualises": "What the chatbot has internally inferred about you — age, gender, education level, socioeconomic status — shown in real time beside the conversation, with controls to change it", "Interaction affordances": "Converse normally while watching the inferred user model update; edit those inferences directly and watch the system's behaviour change", "Pragmatic vs Basic science": "Pragmatic", "Reverse-eng vs Concept-based": "Concept-based", "Observational vs Interventional": "Both", "Intended audience": "General public", "Description (card)": "A dashboard that sits beside a chatbot and shows, in real time, what the model has internally inferred about the user's age, gender, education and socioeconomic status — and lets the user edit those inferences and watch the responses change.", "Why it matters": "Almost everything else in this ledger is built for researchers; this is built for the person being modelled, and a user study found it helped participants expose the system's biased behaviour.", "Visual / design notes": "Dashboard-beside-chat is the key move, but making the user model editable rather than merely visible is what turns transparency into control.", "Tags": "modality:language; method:probing; method:steering; phenomenon:bias; phenomenon:user-modelling", "Thumbnail URL": "", "Media files": "", "Citation": "Chen, Y., Wu, A., DePodesta, T., Yeh, C., Li, K., Castillo Marin, N., Patel, O., Riecke, J., Raval, S., Seow, O., Wattenberg, M. and Viégas, F., 2024. Designing a Dashboard for Transparency and Control of Conversational AI. arXiv:2406.07882", "Related entries": "IA-010", "Confidence": "High", "Source of info": "Read the arXiv abstract page in full, 2026-08-06. Project page (bit.ly/talktuner-project-page) not fetched.", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Cited in Sharkey et al. 2025 section 3.7 on human-computer interaction with model internals. phenomenon:bias and phenomenon:user-modelling were proposed for this row and promoted into the controlled vocabulary by Ava on 2026-08-06; both are now in use here.", "tags_parsed": { "modality": [ "language" ], "method": [ "probing", "steering" ], "phenomenon": [ "bias", "user-modelling" ] } }, { "ID": "IA-009", "Slug": "transformer-explainer", "Name": "Transformer Explainer", "URL": "https://poloclub.github.io/transformer-explainer/", "Artefact type": "Interactive tool", "Visual form": "Flow diagram", "Makes visible": "Architecture", "Role of the visual": "Explanation", "Runnable by a visitor": "On your own inputs", "Date published": "2024-08-08", "Year": "2024", "Link status": "Live", "Built by (people)": "Aeree Cho; Grace C. Kim; Alexander Karpekov; Alec Helbling; Zijie J. Wang; Seongmin Lee; Benjamin Hoover; Duen Horng (Polo) Chau", "Organisation": "Georgia Institute of Technology (Polo Club of Data Science)", "Venue / published in": "IEEE VIS 2024; arXiv", "Sector": "Academia", "Country / region": "US", "Open source": "Yes", "Code repo": "https://github.com/poloclub/transformer-explainer", "Method / technique": "Live in-browser execution of a model with every stage of the forward pass visualised, and smooth transitions between abstraction levels", "Model(s) studied": "GPT-2", "Access required": "Open weights", "What it visualises": "The full GPT-2 forward pass — embeddings, attention queries keys and values, MLP, and next-token probabilities — updating live as you type", "Interaction affordances": "Type your own input and watch predictions update in real time; expand any stage down to the underlying matrix operations; adjust temperature; no installation or GPU required", "Pragmatic vs Basic science": "Neither / educational", "Reverse-eng vs Concept-based": "Neither", "Observational vs Interventional": "Observational", "Intended audience": "General public; Students", "Description (card)": "An interactive tool that runs a live GPT-2 instance in your browser and visualises every stage of the forward pass — embeddings, attention, MLP, output probabilities — updating as you type your own text.", "Why it matters": "The most accessible route into transformer internals that exists: no installation, no GPU, no prior knowledge. That is a different kind of contribution from any research result in this ledger, and arguably a wider-reaching one.", "Visual / design notes": "The core design idea is smooth movement between abstraction levels — you can zoom from a block diagram down to the matrix multiplication without losing your place.", "Tags": "modality:language; method:attention-visualisation; phenomenon:attention-heads", "Thumbnail URL": "", "Media files": "", "Citation": "Cho, A., Kim, G.C., Karpekov, A., Helbling, A., Wang, Z.J., Lee, S., Hoover, B. and Chau, D.H., 2024. Transformer Explainer: Interactive Learning of Text-Generative Models. IEEE VIS 2024. arXiv:2408.04619", "Related entries": "IA-010", "Confidence": "High", "Source of info": "Read the arXiv abstract page in full, 2026-08-06. The live tool itself was not inspected — it is client-rendered and the fetch returned an empty page shell.", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Educational rather than a research discovery: the architecture it visualises is known, not uncovered. A boundary case for the inclusion rule, logged because the visualisation IS the contribution and it was peer-reviewed at IEEE VIS. Access classified Open weights: GPT-2 runs locally in the visitor's browser, no server or account needed.", "tags_parsed": { "modality": [ "language" ], "method": [ "attention-visualisation" ], "phenomenon": [ "attention-heads" ] } }, { "ID": "IA-004", "Slug": "biology-of-a-large-language-model", "Name": "On the Biology of a Large Language Model — attribution graphs", "URL": "https://transformer-circuits.pub/2025/attribution-graphs/biology.html", "Artefact type": "Interactive article", "Visual form": "Node-link graph", "Makes visible": "Circuits", "Role of the visual": "Exhibit", "Runnable by a visitor": "Pre-set examples only", "Date published": "2025-03-27", "Year": "2025", "Link status": "Live", "Built by (people)": "Jack Lindsey; Emmanuel Ameisen; Adam Pearce; Joshua Batson; et al.", "Organisation": "Anthropic", "Venue / published in": "Transformer Circuits Thread", "Sector": "Frontier lab", "Country / region": "US", "Open source": "Yes", "Code repo": "https://github.com/safety-research/circuit-tracer", "Method / technique": "Attribution graphs built on cross-layer transcoders; a 'replacement model' with interpretable nodes; causal intervention to validate edges", "Model(s) studied": "Claude 3.5 Haiku", "Access required": "Internal / proprietary", "What it visualises": "Step-by-step computational paths from prompt to response — which features caused which — across ten case studies including planning, multilingual reasoning, hallucination and jailbreaks", "Interaction affordances": "Step through each case study; expand and collapse graph nodes; trace individual paths; follow intervention results", "Pragmatic vs Basic science": "Both", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Interventional", "Intended audience": "Researchers; Practitioners; General public", "Description (card)": "Ten case studies of Claude 3.5 Haiku's internal mechanisms rendered as explorable attribution graphs — showing it plans rhymes ahead, reasons across a shared multilingual concept space, and sometimes fabricates reasoning backwards from a hinted answer.", "Why it matters": "The clearest existing demonstration that a model's actual reasoning can diverge from its stated reasoning, made legible by graph.", "Visual / design notes": "Interactive graphs you expand case by case; the pedagogy is in the pacing. Companion methods paper carries the technical load so the biology paper can stay readable.", "Tags": "modality:language; method:attribution-graph; method:transcoder; method:activation-patching; phenomenon:planning; phenomenon:hallucination; phenomenon:jailbreak; phenomenon:unfaithful-reasoning", "Thumbnail URL": "", "Media files": "", "Citation": "Lindsey, J., Gurnee, W., Ameisen, E., et al., 2025. On the Biology of a Large Language Model. Transformer Circuits Thread.", "Related entries": "IA-005", "Confidence": "High", "Source of info": "Read the Anthropic research write-up in full, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Paired with the Circuit Tracing methods paper. Tools open-sourced May 2025; Neuronpedia hosts a public graph explorer. Access classified Internal / proprietary: Claude 3.5 Haiku is not public, though the circuit-tracing method was open-sourced for use on open models.", "tags_parsed": { "modality": [ "language" ], "method": [ "attribution-graph", "transcoder", "activation-patching" ], "phenomenon": [ "planning", "hallucination", "jailbreak", "unfaithful-reasoning" ] } }, { "ID": "IA-005", "Slug": "headvis", "Name": "HeadVis — interactive tool for investigating attention heads", "URL": "https://transformer-circuits.pub/2026/headvis/index.html", "Artefact type": "Interactive tool", "Visual form": "Scatter plot", "Makes visible": "Attention", "Role of the visual": "Instrument", "Runnable by a visitor": "Pre-set examples only", "Date published": "2026-05-04", "Year": "2026", "Link status": "Live", "Built by (people)": "R. Luger; Harish Kamath; Doug Finkbeiner; Purvi Goel; Adam Jermyn; Sam Zimmerman; Joshua Batson; Tom Conerly", "Organisation": "Anthropic", "Venue / published in": "Transformer Circuits Thread", "Sector": "Frontier lab", "Country / region": "US", "Open source": "Partial", "Code repo": "https://transformer-circuits.pub/2026/headvis/index.html", "Method / technique": "Attention pattern visualisation; per-head distribution metrics (e.g. induction score); PCA of Q/K/O activations; QK and OV feature attributions via a 10M-feature weakly causal crosscoder", "Model(s) studied": "Claude Haiku 3.5; Gemma 3 1B (public demo)", "Access required": "Open weights", "What it visualises": "Attention head behaviour across a full data distribution — scatter plots of heads by metric, token-level attention patterns, PCA clusters, and QK/OV circuit diagrams", "Interaction affordances": "Scatter-select a head by any metric; hover tokens to see attention; browse dataset examples; run custom sequences (server build); rank all heads on a query-key pair", "Pragmatic vs Basic science": "Basic science", "Reverse-eng vs Concept-based": "Reverse-engineering", "Observational vs Interventional": "Both", "Intended audience": "Researchers", "Description (card)": "An interactive tool for interrogating attention heads, released as the research output in its own right. Its central finding is methodological: a head's behaviour on the full data distribution rarely matches what a narrow task suggests.", "Why it matters": "Reframes the tool as the contribution, and supplies the concrete examples the field needs to attack attention decomposition.", "Visual / design notes": "Scatter-plot head selector with configurable axes is the smart move — it makes 'find an interesting head' a two-click operation instead of a research project.", "Tags": "modality:language; method:attention-visualisation; method:crosscoder; method:dimensionality-reduction; phenomenon:induction-heads; phenomenon:polysemanticity; phenomenon:attention-heads", "Thumbnail URL": "", "Media files": "", "Citation": "Luger, R., Kamath, H., Finkbeiner, D., Goel, P., Jermyn, A., Zimmerman, S., Batson, J. and Conerly, T., 2026. HeadVis: An Interactive Tool For Investigating Attention Heads. Transformer Circuits Thread.", "Related entries": "IA-004", "Confidence": "High", "Source of info": "Read the full article, 2026-08-06", "Date added": "2026-08-06", "Added by": "Claude", "Notes": "Frontend open-sourced; backends not. Custom-prompt and logit-lens views excluded from the public release. Hosted demos for Haiku 3.5 subset and full Gemma 3 1B. Access classified Open weights on the strength of the public Gemma 3 1B demo; the full Haiku 3.5 version requires internal access.", "tags_parsed": { "modality": [ "language" ], "method": [ "attention-visualisation", "crosscoder", "dimensionality-reduction" ], "phenomenon": [ "induction-heads", "polysemanticity", "attention-heads" ] } } ], "vocabularies": { "Artefact type": [ "Interactive tool", "Interactive article", "Browsable index / database", "Platform", "Dashboard", "Static figures in a paper", "Video / animation", "Art object / installation", "Notebook / library", "Deployed model demo", "Other" ], "Role of the visual": [ "Instrument", "Finding", "Exhibit", "Explanation", "Experience", "Unknown" ], "Link status": [ "Live", "Archived", "Dead", "Paywalled", "Unknown" ], "Sector": [ "Frontier lab", "Academia", "Startup", "Independent researcher", "Artist", "Non-profit / community", "Frontier lab + academia", "Unknown" ], "Open source": [ "Yes", "Partial", "No", "Unknown" ], "Access required": [ "Open weights", "API", "Internal / proprietary", "None (pre-rendered)", "Unknown" ], "Pragmatic vs Basic science": [ "Pragmatic", "Basic science", "Both", "Neither / educational", "Unknown" ], "Reverse-eng vs Concept-based": [ "Reverse-engineering", "Concept-based", "Both", "Neither", "Unknown" ], "Observational vs Interventional": [ "Observational", "Interventional", "Both", "Unknown" ], "Intended audience": [ "Researchers", "Practitioners", "Policy", "General public", "Students", "Mixed", "Unknown" ], "Confidence": [ "High", "Medium", "Low", "Unknown" ], "Candidate status": [ "Proposed", "Approved", "Rejected", "Logged" ], "TAG modality": [ "general", "language", "vision", "multimodal", "audio", "code", "image-generation", "protein", "genomics", "rl-agent", "tabular" ], "TAG method": [ "activation-patching", "attention-visualisation", "attribution", "attribution-graph", "auto-interp", "causal-tracing", "concept-injection", "crosscoder", "dimensionality-reduction", "feature-visualisation", "logit-lens", "model-diffing", "probing", "sparse-autoencoder", "steering", "transcoder" ], "TAG phenomenon": [ "attention-heads", "bias", "emotion", "evaluation-awareness", "feature-interpretation", "hallucination", "in-context-learning", "induction-heads", "introspection", "jailbreak", "knowledge-editing", "persona", "planning", "polysemanticity", "refusal", "safety-relevant-features", "superposition", "unfaithful-reasoning", "user-modelling", "world-models" ], "Proposed new tags": [ "(add here, tell Ava, do not use in Ledger until approved)" ], "Runnable by a visitor": [ "On your own inputs", "Pre-set examples only", "Pre-rendered, view only", "No longer available", "Unknown" ], "Visual form": [ "Activation grid", "Dashboard", "Flow diagram", "Heatmap", "Highlighted text", "Node-link graph", "Scatter plot", "Spatial attribution", "None", "Unknown" ], "Makes visible": [ "Features", "Circuits", "Attribution", "Attention", "Activations", "Architecture", "Training dynamics", "Model behaviour", "Model differences", "User model", "Nothing", "Unknown" ], "Article term": [ "Activation grid", "Dashboard", "Flow diagram", "Heatmap", "Highlighted text", "Node-link graph", "Scatter plot", "Spatial attribution", "None", "Unknown", "Features", "Circuits", "Attribution", "Attention", "Activations", "Architecture", "Training dynamics", "Model behaviour", "Model differences", "User model", "Nothing" ], "Article": [ "an", "a", "a", "a", "", "a", "a", "", "", "", "", "", "", "", "", "", "", "", "", "a", "" ] }, "definitions": [ { "Term": "HOW TO READ THE THREE PROVENANCE COLUMNS", "Definition": "Origin says who coined the TERM. Coined in gives the citation. Definition drafted from says what Claude actually read to write the sentence — which is not the same thing.", "Plain language": "", "Note / example": "", "Origin": "", "Coined in / attributed to": "", "Definition drafted from": "" }, { "Term": "Field term", "Definition": "Established vocabulary in the published interpretability literature. The citation is the earliest use Claude could verify, not a guaranteed first use.", "Plain language": "", "Note / example": "", "Origin": "", "Coined in / attributed to": "", "Definition drafted from": "" }, { "Term": "Common ML usage", "Definition": "In wide use across machine learning with no single coining source. Safe to use, not attributable.", "Plain language": "", "Note / example": "", "Origin": "", "Coined in / attributed to": "", "Definition drafted from": "" }, { "Term": "Atlas schema", "Definition": "Coined for this index to describe its own structure. Not a claim about the field, and no external source exists.", "Plain language": "", "Note / example": "", "Origin": "", "Coined in / attributed to": "", "Definition drafted from": "" }, { "Term": "Ava's framing", "Definition": "Supplied by Ava in conversation rather than drawn from a source.", "Plain language": "", "Note / example": "", "Origin": "", "Coined in / attributed to": "", "Definition drafted from": "" }, { "Term": "Claude's coinage", "Definition": "The practice is real but the LABEL was invented by Claude and is not established. Rename before publishing where a standard term exists.", "Plain language": "", "Note / example": "", "Origin": "", "Coined in / attributed to": "", "Definition drafted from": "" }, { "Term": "Purpose", "Definition": "Every column in the Ledger and every permitted value is defined here in one sentence, so that two people filling in the same row reach the same answer.", "Plain language": "", "Note / example": "Also intended as published glossary copy for the website.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Controlled value", "Definition": "A value that may only be chosen from the list on the Vocabularies sheet, never typed freely.", "Plain language": "", "Note / example": "Free-typed values fragment the website's filters.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Free text", "Definition": "A field written in prose, with no fixed list of permitted values.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "ID", "Definition": "A permanent identifier for one artefact, never reused, so that a retired entry's number can never resolve to a different record.", "Plain language": "", "Note / example": "IA-007", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Slug", "Definition": "The lowercase hyphenated form of the entry's name, used as its web address.", "Plain language": "", "Note / example": "activation-atlas", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Name", "Definition": "The artefact's own title, as given by its makers.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "URL", "Definition": "The canonical public address of the artefact itself, rather than of a paper describing it.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Artefact type", "Definition": "The form the work takes, independent of what it contains or argues.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Role of the visual", "Definition": "The function the visual element performs within the research, which is the central judgement in each row.", "Plain language": "", "Note / example": "Controlled value. Definitions below.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Runnable by a visitor", "Definition": "Whether, and how far, a person arriving today can operate the artefact themselves.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Date published", "Definition": "The date the artefact first became publicly available, in ISO form where known.", "Plain language": "", "Note / example": "2019-03-06", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Year", "Definition": "The publication year held separately, so that entries sort chronologically without parsing dates.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Link status", "Definition": "Whether the recorded URL currently resolves to the artefact.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Built by (people)", "Definition": "The named individual contributors, in the order the source credits them.", "Plain language": "", "Note / example": "Semicolon separated.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Organisation", "Definition": "The institution that produced the work.", "Plain language": "", "Note / example": "Distinct from where it was published.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Venue / published in", "Definition": "Where the work appeared, which is often not the same as who produced it.", "Plain language": "", "Note / example": "Google Brain produced Activation Atlas; Distill published it.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Sector", "Definition": "The kind of institution that produced the work, recorded so the composition of the field is visible.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Country / region", "Definition": "Where the producing institution is based.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Open source", "Definition": "Whether the artefact's own source code is publicly available.", "Plain language": "", "Note / example": "Controlled value. Not the model's licence.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Code repo", "Definition": "A link to that source code.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Method / technique", "Definition": "The interpretability technique or techniques the artefact rests on.", "Plain language": "", "Note / example": "Free text; mirrored in method: tags.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Model(s) studied", "Definition": "The specific neural networks the work examines.", "Plain language": "", "Note / example": "Name them; avoid 'various'.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Access required", "Definition": "The level of model access a person would need to reproduce the work.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "What it visualises", "Definition": "The object being made visible, stated in plain terms a non-specialist could follow.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Interaction affordances", "Definition": "What a person can actually do with the artefact, verb by verb.", "Plain language": "", "Note / example": "'Hover to inspect; swap layer; run your own prompt'.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Pragmatic vs Basic science", "Definition": "Whether the work targets one specific behaviour or the model's general structure.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Reverse-eng vs Concept-based", "Definition": "Whether the work decomposes the network and then names the parts, or proposes concepts and then locates them.", "Plain language": "", "Note / example": "Controlled value. After Sharkey et al., 2025.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Observational vs Interventional", "Definition": "Whether the work examines the model unchanged or alters it and observes the consequence.", "Plain language": "", "Note / example": "Controlled value. After Kowalska and Kwasnicka, 2026.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Intended audience", "Definition": "Who the artefact was designed to be understood by.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Description (card)", "Definition": "One or two sentences describing the artefact to a stranger, written to be published rather than as a note to self.", "Plain language": "", "Note / example": "This becomes the website card text.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Why it matters", "Definition": "One line stating the artefact's significance in terms a reader could dispute.", "Plain language": "", "Note / example": "'An important paper' fails this test.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Visual / design notes", "Definition": "Observations on the artefact's visual and interaction design, including what makes it work or fail.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Tags", "Definition": "Faceted keywords in facet:value form, drawn only from the controlled facet lists.", "Plain language": "", "Note / example": "modality:vision; method:attribution", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Thumbnail URL", "Definition": "An image representing the entry.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Citation", "Definition": "A full bibliographic reference in a consistent style.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Related entries", "Definition": "The IDs of other Atlas entries with a documented relationship to this one.", "Plain language": "", "Note / example": "IA-005; IA-011", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Confidence", "Definition": "How well evidenced this row is, judged by what the person writing it actually read.", "Plain language": "", "Note / example": "Controlled value.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Source of info", "Definition": "What was read to write this row, and on what date.", "Plain language": "", "Note / example": "'Read the arXiv abstract page, 2026-08-06'", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Date added / Added by", "Definition": "When the row was created and by whom, human or model.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Notes", "Definition": "Anything qualifying the row, including unresolved questions and flagged boundary cases.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Instrument", "Definition": "Others can operate it on their own inputs or models to produce findings that did not previously exist.", "Plain language": "", "Note / example": "Test: could someone publish a result they got from this?", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava. Not a field taxonomy — it does not exist elsewhere", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Finding", "Definition": "The result is constituted by the visual, such that removing the images would remove the contribution.", "Plain language": "", "Note / example": "Test: delete the pictures — is the contribution gone?", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava. Not a field taxonomy — it does not exist elsewhere", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Exhibit", "Definition": "The visual exists so that a reader can inspect and audit a claim which is itself stated in prose.", "Plain language": "", "Note / example": "Test: is it here so you can check them?", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava. Not a field taxonomy — it does not exist elsewhere", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Explanation", "Definition": "The visual communicates a result established elsewhere, serving teaching rather than discovery.", "Plain language": "", "Note / example": "Test: is it teaching something already known?", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava. Not a field taxonomy — it does not exist elsewhere", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Experience", "Definition": "The artefact is undergone rather than read, and its value lies in what it does to the visitor.", "Plain language": "", "Note / example": "Test: is the point what it does to you?", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava. Not a field taxonomy — it does not exist elsewhere", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Unknown", "Definition": "The role has not been determined.", "Plain language": "", "Note / example": "Never use to avoid making the call.", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava. Not a field taxonomy — it does not exist elsewhere", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "On your own inputs", "Definition": "A visitor can supply their own prompt, image or model and see fresh output.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Pre-set examples only", "Definition": "A visitor can explore interactively, but only within cases the authors selected in advance.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Pre-rendered, view only", "Definition": "The visuals are fixed; interaction is limited to navigation such as panning or zooming.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "No longer available", "Definition": "The artefact was operable once but is not now.", "Plain language": "", "Note / example": "The record is kept; the artefact is gone.", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Unknown", "Definition": "Not determined, often because the page could not be loaded.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "Proposed by Claude 2026-08-06, approved by Ava", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Interactive tool", "Definition": "A standalone application whose purpose is to let a user investigate a model.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Interactive article", "Definition": "A written argument in which the figures respond to the reader.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Browsable index / database", "Definition": "A searchable collection of results presented for inspection.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Platform", "Definition": "Infrastructure hosting many artefacts, models or datasets originating from multiple parties.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Dashboard", "Definition": "A live readout of a system's internal state, displayed alongside that system in use.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Static figures in a paper", "Definition": "Fixed images within a conventional publication.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Video / animation", "Definition": "Time-based linear media.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Art object / installation", "Definition": "Work made primarily for aesthetic or critical reception rather than scientific reporting.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Notebook / library", "Definition": "Code published to be run and adapted by other researchers.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Deployed model demo", "Definition": "A model made publicly usable in a deliberately modified state, as a demonstration.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Other", "Definition": "A form not covered above; propose a new term rather than reusing this indefinitely.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Live", "Definition": "The URL resolves and the artefact functions as described.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Archived", "Definition": "Preserved in static or historical form and no longer maintained.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Dead", "Definition": "The URL no longer resolves to the artefact.", "Plain language": "", "Note / example": "Set this rather than deleting the row.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Paywalled", "Definition": "Reachable only behind payment or registration.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Frontier lab", "Definition": "An organisation training state-of-the-art general-purpose models.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Academia", "Definition": "A university or academic research group.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Startup", "Definition": "A commercial company whose product is interpretability or model tooling.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Independent researcher", "Definition": "A person working outside any institution.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Artist", "Definition": "A practitioner working primarily within an art context.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Non-profit / community", "Definition": "An organisation constituted for public benefit rather than profit.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Frontier lab + academia", "Definition": "Joint authorship spanning both.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Yes", "Definition": "The artefact's own source code is publicly available under a licence.", "Plain language": "", "Note / example": "Not the same as the model being open weight.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Partial", "Definition": "Some components are released and others deliberately withheld.", "Plain language": "", "Note / example": "HeadVis: frontend released, backends not.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "No", "Definition": "No source is published.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Open weights", "Definition": "Reproducible using publicly downloadable model weights.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "API", "Definition": "Requires gated or paid inference access, but not weights.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Internal / proprietary", "Definition": "Requires access available only within the producing organisation.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "None (pre-rendered)", "Definition": "Requires no model access at all; the outputs are fixed.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Pragmatic", "Definition": "Work targeting a specific model behaviour, asking which parts of the network produce it.", "Plain language": "", "Note / example": "Diagnosing a symptom.", "Origin": "Ava's framing", "Coined in / attributed to": "Ava, in conversation, 2026-08-06. Consonant with Nanda et al., 2025, 'A Pragmatic Vision for Interpretability'", "Definition drafted from": "Ava's definition, given in conversation 2026-08-06; consonant with Nanda et al., 2025, A Pragmatic Vision for Interpretability (read)" }, { "Term": "Basic science", "Definition": "Work seeking to reverse-engineer the model's general structure, irrespective of any single behaviour.", "Plain language": "", "Note / example": "Mapping the organism.", "Origin": "Ava's framing", "Coined in / attributed to": "Ava, in conversation, 2026-08-06. Consonant with Nanda et al., 2025, 'A Pragmatic Vision for Interpretability'", "Definition drafted from": "Ava's definition, given in conversation 2026-08-06; consonant with Nanda et al., 2025, A Pragmatic Vision for Interpretability (read)" }, { "Term": "Both", "Definition": "Work pursuing a specific behaviour in order to establish general structure, or the reverse.", "Plain language": "", "Note / example": "", "Origin": "Ava's framing", "Coined in / attributed to": "Ava, in conversation, 2026-08-06. Consonant with Nanda et al., 2025, 'A Pragmatic Vision for Interpretability'", "Definition drafted from": "Ava's definition, given in conversation 2026-08-06; consonant with Nanda et al., 2025, A Pragmatic Vision for Interpretability (read)" }, { "Term": "Neither / educational", "Definition": "Work that neither diagnoses a behaviour nor uncovers structure, but conveys existing knowledge.", "Plain language": "", "Note / example": "", "Origin": "Ava's framing", "Coined in / attributed to": "Ava, in conversation, 2026-08-06. Consonant with Nanda et al., 2025, 'A Pragmatic Vision for Interpretability'", "Definition drafted from": "Ava's definition, given in conversation 2026-08-06; consonant with Nanda et al., 2025, A Pragmatic Vision for Interpretability (read)" }, { "Term": "Reverse-engineering", "Definition": "Decomposing the network into parts and then determining what each part does.", "Plain language": "", "Note / example": "Identifies roles for components.", "Origin": "Field term", "Coined in / attributed to": "Sharkey et al., 2025 §2 — the paper's own framing", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "Concept-based", "Definition": "Proposing a human concept in advance and then locating where the network represents it.", "Plain language": "", "Note / example": "Identifies components for roles.", "Origin": "Field term", "Coined in / attributed to": "Sharkey et al., 2025 §2 — the paper's own framing", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "Both", "Definition": "The work does each in turn, typically proposing concepts to interpret a decomposition.", "Plain language": "", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Sharkey et al., 2025 §2 — the paper's own framing", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "Neither", "Definition": "The work does not attempt either, as with purely pedagogical artefacts.", "Plain language": "", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Sharkey et al., 2025 §2 — the paper's own framing", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "Observational", "Definition": "The model is examined without alteration, so the findings are correlational.", "Plain language": "", "Note / example": "Probing, feature visualisation, lenses.", "Origin": "Field term", "Coined in / attributed to": "Kowalska & Kwasnicka, 2026 §2.2 — the paper's own taxonomy", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06" }, { "Term": "Interventional", "Definition": "The model is altered and the consequences observed, supporting causal claims.", "Plain language": "", "Note / example": "Patching, ablation, steering.", "Origin": "Field term", "Coined in / attributed to": "Kowalska & Kwasnicka, 2026 §2.2 — the paper's own taxonomy", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06" }, { "Term": "Both", "Definition": "The work observes to generate a hypothesis and intervenes to test it.", "Plain language": "", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Kowalska & Kwasnicka, 2026 §2.2 — the paper's own taxonomy", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06" }, { "Term": "Researchers", "Definition": "Specialists in interpretability or machine learning.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Practitioners", "Definition": "People deploying or auditing models professionally.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Policy", "Definition": "Regulators, governments and civil society organisations.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "General public", "Definition": "Non-specialists, with no assumed technical background.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Students", "Definition": "Learners being taught the material for the first time.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Mixed", "Definition": "Deliberately designed to serve more than one of the above.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "High", "Definition": "The primary source was read in full by whoever wrote the row.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Medium", "Definition": "The row rests on an abstract, documentation, or a reliable secondary source.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Low", "Definition": "The row rests on inference and has not been checked against a source.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Unknown", "Definition": "The basis for the row was not recorded.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "modality", "Definition": "The kind of model or data the work concerns.", "Plain language": "", "Note / example": "modality:vision", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "method", "Definition": "The interpretability technique the work employs.", "Plain language": "", "Note / example": "method:probing", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "phenomenon", "Definition": "The property or behaviour of the model that the work is about.", "Plain language": "", "Note / example": "phenomenon:superposition", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Why only three", "Definition": "Organisation, artefact type, sector and open source are already columns; tagging them again would create two sources of truth that drift apart.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "Why no judgements", "Definition": "A tag is a filter, not an opinion; claims about significance belong in Why it matters, where a reader can dispute them.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "general", "Definition": "Not tied to any particular model or data type.", "Plain language": "Not tied to one kind of data. Applies across different models and media.", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "language", "Definition": "Models that generate or process text.", "Plain language": "Text. Models that read and write words.", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "vision", "Definition": "Models that process static images.", "Plain language": "Images. Models that look at pictures.", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "multimodal", "Definition": "Models that process more than one modality jointly.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "audio", "Definition": "Models that process sound or speech.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "code", "Definition": "Models whose subject matter is programming languages.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "image-generation", "Definition": "Models that produce images, such as diffusion models.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "protein", "Definition": "Models of protein sequence or structure.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "genomics", "Definition": "Models of genetic sequence.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "rl-agent", "Definition": "Agents trained by reinforcement learning to act within an environment.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "tabular", "Definition": "Models over structured, columnar data.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "feature-visualisation", "Definition": "Synthesising an input that maximally activates a chosen component.", "Plain language": "Generating an image that shows what a part of the model responds to most strongly.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Olah, Mordvintsev & Schubert, 2017, 'Feature Visualization', Distill", "Definition drafted from": "Olah et al., 2017/2018, via Distill Building Blocks — read 2026-08-06" }, { "Term": "attribution", "Definition": "Assigning responsibility for an output to particular inputs or components.", "Plain language": "Working out which inputs were responsible for an output.", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "No single origin; saliency and attribution literature from Simonyan et al., 2014 and Sundararajan et al., 2017", "Definition drafted from": "Olah et al., 2018, Building Blocks; also Sharkey et al., 2025 — read" }, { "Term": "sparse-autoencoder", "Definition": "Learning an overcomplete, sparsely activating basis for a layer's activations.", "Plain language": "A technique for pulling a model's tangled internals apart into separate, nameable pieces.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Sparse coding: Olshausen & Field, 1997 (as cited in Kowalska & Kwasnicka, 2026). Applied to interpretability: Sharkey, Braun & Millidge, 2022; Bricken et al., 2023; Huben/Cunningham et al., 2024", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "transcoder", "Definition": "A sparse dictionary trained to reproduce a layer's output rather than its input.", "Plain language": "A variant that reconstructs what a layer passes onward, rather than what it received.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Dunefsky, Chlenski & Nanda, 2024 (as cited in Sharkey et al., 2025)", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "crosscoder", "Definition": "A sparse dictionary trained jointly across several layers or several models.", "Plain language": "A variant that works across several layers, or several models, at once.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Lindsey et al., 2024, 'Sparse Crosscoders for Cross-Layer Features and Model Diffing', Anthropic (as cited in Sharkey et al., 2025)", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "attribution-graph", "Definition": "A causal graph of interpretable components explaining a single forward pass.", "Plain language": "A map of which internal pieces caused which, for one specific prompt.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Ameisen et al., 2025, 'Circuit Tracing', Anthropic", "Definition drafted from": "Anthropic, Circuit Tracing / On the Biology of a Large Language Model, 2025 — read 2026-08-06" }, { "Term": "attention-visualisation", "Definition": "Displaying which positions a model attends to, and how strongly.", "Plain language": "Showing which words the model looked at, and how hard.", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "Descriptive; earliest widely used tool is BertViz, Vig, 2019", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "probing", "Definition": "Training a simple classifier on internal activations to test what they encode.", "Plain language": "Training a small, simple classifier to test whether some information is present inside the model.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Kohn, 2015; Gupta et al., 2015; Alain & Bengio, 2017 (as cited in Sharkey et al., 2025 §2.2.1)", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06; also Kowalska & Kwasnicka, 2026 §4.1.1" }, { "Term": "steering", "Definition": "Adding or subtracting a direction in activation space in order to alter behaviour.", "Plain language": "Nudging the model's internals mid-thought to change what it says.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "'Activation steering' / 'activation addition': Turner et al., 2024, credited with introducing it in Sharkey et al., 2025 §3.2.2", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "logit-lens", "Definition": "Projecting an intermediate representation directly into the output vocabulary.", "Plain language": "Peeking at what the model would say if it stopped thinking at this layer.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "nostalgebraist, 2020 (as cited in Sharkey et al., 2025 and Kowalska & Kwasnicka, 2026 §4.1.2)", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06 §4.1.2" }, { "Term": "activation-patching", "Definition": "Substituting one activation for another to measure its causal effect on the output.", "Plain language": "Swapping one internal value for another to see whether it mattered.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Vig et al., 2020; Geiger et al., 2020; named 'activation patching' by Nanda, 2023 (as cited in Sharkey et al., 2025 §2.1.3b)", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06 §4.2.1" }, { "Term": "concept-injection", "Definition": "Inserting a known representation into an unrelated context to test detection or effect.", "Plain language": "Inserting a known idea into the model's internals to see whether it notices.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Lindsey, 2025, 'Emergent Introspective Awareness', Anthropic — names the technique", "Definition drafted from": "Anthropic, Emergent Introspective Awareness, 2025 — read 2026-08-06" }, { "Term": "dimensionality-reduction", "Definition": "Projecting high-dimensional activations into two or three dimensions so their structure can be seen.", "Plain language": "Flattening very high-dimensional data down to two or three dimensions so it can be drawn.", "Note / example": "Merged from Claude's coinages 'pca-geometry' and 'embedding-projection', 2026-08-16.", "Origin": "Common ML usage", "Coined in / attributed to": "Standard technique (PCA, t-SNE, UMAP); no interpretability-specific origin.", "Definition drafted from": "Renamed from the coinages pca-geometry and embedding-projection on 2026-08-16. Standard technique name" }, { "Term": "model-diffing", "Definition": "Comparing two models in order to isolate what differs between them.", "Plain language": "Comparing two models to isolate exactly what is different between them.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Bricken et al., 2024, 'Stage-Wise Model Diffing'; Lindsey et al., 2024, Anthropic", "Definition drafted from": "Anthropic, A diff tool for AI, 2026 — read 2026-08-06" }, { "Term": "causal-tracing", "Definition": "Locating where within a network a specific piece of knowledge is stored.", "Plain language": "Finding where in the model a particular fact is stored.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Meng et al., 2022, 'Locating and Editing Factual Associations in GPT' (ROME) (as cited in Kowalska & Kwasnicka, 2026)", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06 §4.2" }, { "Term": "auto-interp", "Definition": "Using a model to generate natural-language explanations of another model's internals.", "Plain language": "Using one model to write explanations of what is happening inside another.", "Note / example": "Renamed from Claude's coinage 'automated-explanation', 2026-08-16.", "Origin": "Field term", "Coined in / attributed to": "Practice introduced by Bills et al., 2023, OpenAI; 'auto-interp' is the field's established shorthand.", "Definition drafted from": "CLAUDE COINAGE — the practice is real, the label is not a field term. Rename before publishing if a standard term exists." }, { "Term": "superposition", "Definition": "Representing more features than there are dimensions, by encoding them sparsely and near-orthogonally.", "Plain language": "A model packing more ideas into its wiring than it has room for, by letting them overlap.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Elhage et al., 2022, 'Toy Models of Superposition', Anthropic (as cited in Sharkey et al., 2025)", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06 (after Elhage et al., 2022)" }, { "Term": "polysemanticity", "Definition": "A single component responding to several unrelated concepts.", "Plain language": "One part of the model doing several unrelated jobs at once.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Olah et al., 2020, 'Zoom In', Distill; Sharkey et al., 2025 traces the observation to Olah et al., 2017 and earlier", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "feature-interpretation", "Definition": "What an individual learned feature means.", "Plain language": "Working out what one small piece of the model has learned to recognise.", "Note / example": "Renamed from Claude's coinage 'feature-semantics', 2026-08-16 — the field already had this.", "Origin": "Field term", "Coined in / attributed to": "In use across Bricken et al., 2023 and Templeton et al., 2024; Sharkey et al., 2025 §2.1.3 frames it as 'describing the functional role of components'.", "Definition drafted from": "Renamed from the coinage feature-semantics on 2026-08-16. Bricken et al. 2023 and Templeton et al. 2024 use feature interpretation" }, { "Term": "planning", "Definition": "Selecting a future output before generating the text that leads to it.", "Plain language": "The model deciding where a sentence is going before it writes it.", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "General term; as an interpretability finding, Lindsey et al., 2025, 'On the Biology of a Large Language Model'", "Definition drafted from": "Anthropic, On the Biology of a Large Language Model, 2025 — read 2026-08-06" }, { "Term": "hallucination", "Definition": "Producing confident content for which the model has no basis.", "Plain language": "The model stating something confidently when it has nothing to base it on.", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "Widespread in NLP; no single coining source", "Definition drafted from": "Anthropic, On the Biology of a Large Language Model, 2025 — read 2026-08-06" }, { "Term": "jailbreak", "Definition": "Circumventing a model's trained refusals.", "Plain language": "Getting a model to do something it was trained to refuse.", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "Widespread; borrowed from software security usage", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "refusal", "Definition": "The mechanism by which a model declines a request.", "Plain language": "How a model decides to say no.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "'Refusal direction': Arditi et al., 2024, 'Refusal in Language Models Is Mediated by a Single Direction' (as cited in Sharkey et al., 2025)", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06 (after Arditi et al., 2024)" }, { "Term": "induction-heads", "Definition": "Attention heads that continue a pattern seen earlier in the context.", "Plain language": "Parts of the model that spot a pattern earlier in the text and continue it.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Elhage et al., 2021, 'A Mathematical Framework for Transformer Circuits'; developed in Olsson et al., 2022, Anthropic", "Definition drafted from": "Anthropic, HeadVis, 2026 — read 2026-08-06 — paraphrase of their definition" }, { "Term": "world-models", "Definition": "Internal representations of an external state the model was never given directly.", "Plain language": "Signs that a model has built an internal picture of something nobody told it directly.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Li et al., 2023, Othello-GPT (as cited in Sharkey et al., 2025 and Kowalska & Kwasnicka, 2026)", "Definition drafted from": "Kowalska & Kwasnicka, 2026, Unboxing the Black Box — read in full 2026-08-06" }, { "Term": "persona", "Definition": "The character a model adopts, and the stability of that adoption.", "Plain language": "The character a model plays, and how easily it slips out of it.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "'Persona vectors': Anthropic, 2025, arXiv:2507.21509", "Definition drafted from": "Anthropic, Persona Vectors, 2025 — read 2026-08-06" }, { "Term": "emotion", "Definition": "Representations of emotional states and their causal effect on behaviour.", "Plain language": "Internal states that work like emotions and change what the model does.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "'Functional emotions' / 'emotion vectors': Sofroniew et al., 2026, Anthropic", "Definition drafted from": "Anthropic, Emotion Concepts in a Large Language Model, 2026 — read 2026-08-06" }, { "Term": "introspection", "Definition": "A model's capacity to report on its own internal states.", "Plain language": "Whether a model can accurately report on its own thinking.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Lindsey, 2025, 'Emergent Introspective Awareness', Anthropic", "Definition drafted from": "Anthropic, Emergent Introspective Awareness, 2025 — read 2026-08-06" }, { "Term": "evaluation-awareness", "Definition": "A model recognising that it is being tested, and behaving differently as a result.", "Plain language": "A model noticing it is being tested, and behaving differently because of it.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "In use across Anthropic system cards and Apollo Research, 2025. NO SINGLE COINING PAPER IDENTIFIED — verify before citing", "Definition drafted from": "Anthropic, Natural Language Autoencoders, 2026 — read 2026-08-06; also Nanda et al., 2025" }, { "Term": "unfaithful-reasoning", "Definition": "Stated reasoning that does not correspond to the computation actually performed.", "Plain language": "The explanation a model gives not matching the reasoning it actually did.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "'Chain-of-thought unfaithfulness': Turpin et al., 2023; Arcuschin, Conmy et al., 2025 (as cited in Sharkey et al., 2025)", "Definition drafted from": "Anthropic, On the Biology of a Large Language Model, 2025 — read 2026-08-06" }, { "Term": "knowledge-editing", "Definition": "Deliberately altering a specific fact the model holds.", "Plain language": "Deliberately changing one specific fact a model holds.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Meng et al., 2022 (ROME); survey Wang et al., 2024 (as cited in Sharkey et al., 2025 §3.2.2)", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06" }, { "Term": "safety-relevant-features", "Definition": "Representations bearing on harm, deception, bias or misuse.", "Plain language": "Parts of the model to do with harm, deception, bias or misuse.", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "Descriptive. The phrase appears loosely in Templeton et al., 2024 but is not a defined term of art", "Definition drafted from": "Anthropic, Scaling Monosemanticity, 2024 — via secondary reading, not the paper itself — read 2026-08-06" }, { "Term": "in-context-learning", "Definition": "Acquiring a capability from the prompt rather than from training.", "Plain language": "The model picking up a new skill from the prompt alone, without being retrained.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Brown et al., 2020 (GPT-3); mechanism in Olsson et al., 2022", "Definition drafted from": "Sharkey et al., 2025, Open Problems in Mechanistic Interpretability — read in full 2026-08-06 (after Olsson et al., 2022)" }, { "Term": "attention-heads", "Definition": "What an attention head does, on a chosen task or across the full data distribution.", "Plain language": "The parts that decide which earlier words matter for the next one.", "Note / example": "Renamed from Claude's coinage 'attention-behaviour', 2026-08-16.", "Origin": "Field term", "Coined in / attributed to": "Elhage et al., 2021, A Mathematical Framework for Transformer Circuits, Anthropic.", "Definition drafted from": "Renamed from the coinage attention-behaviour on 2026-08-16. attention-heads is the standard field term" }, { "Term": "bias", "Definition": "Systematic differential treatment of people or groups by the model.", "Plain language": "The model treating people or groups differently in ways it should not.", "Note / example": "Promoted from staging 2026-08-06.", "Origin": "Common ML usage", "Coined in / attributed to": "Widespread across ML fairness literature; no single origin", "Definition drafted from": "Atlas convention — defined for this project, no external source" }, { "Term": "user-modelling", "Definition": "The model's internal inferences about the person it is talking to.", "Plain language": "What the model has quietly worked out about the person talking to it.", "Note / example": "Promoted from staging 2026-08-06.", "Origin": "Field term", "Coined in / attributed to": "'User model': Chen et al., 2024, TalkTuner, arXiv:2406.07882", "Definition drafted from": "Chen et al., 2024, TalkTuner (arXiv:2406.07882) — abstract read 2026-08-06" }, { "Term": "Visual form", "Definition": "The shape the picture takes: what you would call the graphic if describing it to someone who could not see it.", "Plain language": "", "Note / example": "Controlled, may hold several values. The pair Visual form and Makes visible is what a visual thinker navigates by.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" }, { "Term": "Makes visible", "Definition": "The thing rendered on screen. Not what the work is about, which is the phenomenon tag, but what you are actually looking at.", "Plain language": "", "Note / example": "Towards Monosemanticity: Visual form = annotated text, Makes visible = features, phenomenon = superposition.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" }, { "Term": "Media files", "Definition": "The image filenames belonging to this entry, semicolon separated.", "Plain language": "", "Note / example": "This is what links a Ledger row to its pictures. Without it the website cannot show thumbnails.", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" }, { "Term": "Activation grid", "Definition": "Many small images laid out in a grid, each one standing for what a part of the model responded to at that position.", "Plain language": "Lots of tiny pictures in a grid, each showing what the model saw at that spot.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K. and Mordvintsev, A., 2018. The Building Blocks of Interpretability. Distill. Verified first-hand 2026-08-22", "Definition drafted from": "Fetched and searched the Distill article, 2026-08-22" }, { "Term": "Dashboard", "Definition": "Several visual forms arranged together as one interface, read as a unit.", "Plain language": "A single screen putting several different readouts side by side.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Chen, Y. et al., 2024. Designing a Dashboard for Transparency and Control of Conversational AI. Also standard field usage: 'feature dashboard'", "Definition drafted from": "Field usage; TalkTuner paper title" }, { "Term": "Flow diagram", "Definition": "Boxes and arrows showing a process or structure in sequence.", "Plain language": "Boxes joined by arrows, showing what happens in what order.", "Note / example": "", "Origin": "Information visualisation", "Coined in / attributed to": "Deng, D., Wu, Y., Shu, X., Wu, J., Fu, S., Cui, W. and Wu, Y., 2022. VisImages: A Fine-Grained Expert-Annotated Visualization Dataset. IEEE TVCG. NOT READ FIRST-HAND — taxonomy identified via search summary, verify before publication", "Definition drafted from": "VisImages category: Diagram" }, { "Term": "Heatmap", "Definition": "A grid in which colour intensity encodes magnitude.", "Plain language": "A grid where stronger colour means a bigger number.", "Note / example": "", "Origin": "Information visualisation", "Coined in / attributed to": "Deng, D., Wu, Y., Shu, X., Wu, J., Fu, S., Cui, W. and Wu, Y., 2022. VisImages: A Fine-Grained Expert-Annotated Visualization Dataset. IEEE TVCG. NOT READ FIRST-HAND — taxonomy identified via search summary, verify before publication", "Definition drafted from": "VisImages category: Grid & Matrix" }, { "Term": "Highlighted text", "Definition": "Running text with individual words or tokens shaded to encode a per-token value, most often how strongly a feature activated there.", "Plain language": "Ordinary text with words shaded in, where the shading shows how strongly the model reacted to each word.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Anthropic, 2023-2024, Towards Monosemanticity and Scaling Monosemanticity: 'text examples where the feature activates most strongly, with activating tokens highlighted'. Also 'token highlighting' in the visual analytics literature", "Definition drafted from": "Anthropic feature browsers; token highlighting in visual analytics" }, { "Term": "Node-link graph", "Definition": "Nodes joined by edges, showing what connects to or causes what.", "Plain language": "Dots joined by lines, showing what connects to what.", "Note / example": "", "Origin": "Information visualisation", "Coined in / attributed to": "Deng, D., Wu, Y., Shu, X., Wu, J., Fu, S., Cui, W. and Wu, Y., 2022. VisImages: A Fine-Grained Expert-Annotated Visualization Dataset. IEEE TVCG. NOT READ FIRST-HAND — taxonomy identified via search summary, verify before publication", "Definition drafted from": "VisImages category: Graph & Tree" }, { "Term": "Scatter plot", "Definition": "A graph of paired numerical values, with one variable on the horizontal axis and the corresponding value of a second variable on the vertical axis, used to reveal relationships or association between the variables.", "Plain language": "A graph where each point represents a pair of values, showing the relationship between two sets of numbers on a coordinate plane.", "Note / example": "", "Origin": "Information visualisation", "Coined in / attributed to": "NIST/SEMATECH, Engineering Statistics Handbook, section 1.3.3.26, Scatter Plot. https://itl.nist.gov/div898/handbook/eda/section3/eda33q.htm", "Definition drafted from": "NIST Engineering Statistics Handbook, section 1.3.3.26, read first-hand 2026-08-25" }, { "Term": "Spatial attribution", "Definition": "Attribution values rendered over the positions of an input image, showing which regions drove the output. A saliency map is the most common instance.", "Plain language": "Colour laid over a picture showing which parts of it the model was reacting to.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K. and Mordvintsev, A., 2018. The Building Blocks of Interpretability. Distill. Verified first-hand 2026-08-22", "Definition drafted from": "Distill section heading: Spatial Attribution with Saliency Maps" }, { "Term": "Attribution", "Definition": "How much each part of an input pushed the model towards its output. Distinct from a feature, which is what the model represents, and from an activation, which is how hard a component is firing.", "Plain language": "How much each bit of the input pushed the model towards its answer.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K. and Mordvintsev, A., 2018. The Building Blocks of Interpretability. Distill. Verified first-hand 2026-08-22", "Definition drafted from": "Used throughout the Distill article; Atlas already carries method:attribution" }, { "Term": "None", "Definition": "No graphic form. The artefact qualifies some other way.", "Plain language": "This entry produces no picture of its own.", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" }, { "Term": "Unknown", "Definition": "Not yet determined.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" }, { "Term": "Features", "Definition": "Individual learned directions or concepts inside the model.", "Plain language": "A single thing the model has learned to recognise, like a place, a tone of voice, or a kind of mistake.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Bricken, T. et al., 2023, Towards Monosemanticity; Olah, C. et al., 2020, Zoom In: An Introduction to Circuits, Distill", "Definition drafted from": "Attribution held from the project context brief; NOT re-verified first-hand this session" }, { "Term": "Circuits", "Definition": "Connected components and the paths between them.", "Plain language": "A chain of parts inside the model that work together to do one job.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Zoom In: An Introduction to Circuits. Distill", "Definition drafted from": "Attribution held from the project context brief; NOT re-verified first-hand this session" }, { "Term": "Attention", "Definition": "Which positions a model attends to, and how strongly.", "Plain language": "How the model decides which earlier words matter when it is working on the current one.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Vaswani, A. et al., 2017. Attention Is All You Need", "Definition drafted from": "Attribution held from the project context brief; NOT re-verified first-hand this session" }, { "Term": "Activations", "Definition": "Raw internal values, shown without further decomposition.", "Plain language": "", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Standard throughout the interpretability literature; no single coining paper", "Definition drafted from": "Attribution held from the project context brief; NOT re-verified first-hand this session" }, { "Term": "Architecture", "Definition": "The model's structure rather than anything it learned.", "Plain language": "The shape of the model itself: what parts it has, and the order they run in.", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "General machine-learning usage; no single source", "Definition drafted from": "" }, { "Term": "Training dynamics", "Definition": "How the model changes over the course of training.", "Plain language": "", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "General machine-learning usage; no single source", "Definition drafted from": "" }, { "Term": "Model behaviour", "Definition": "What the model does, observed from the outside.", "Plain language": "", "Note / example": "", "Origin": "Common ML usage", "Coined in / attributed to": "General machine-learning usage; no single source", "Definition drafted from": "" }, { "Term": "Model differences", "Definition": "What separates one model from another.", "Plain language": "", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Model diffing / crosscoder literature; no single coining paper identified", "Definition drafted from": "Attribution held from the project context brief; NOT re-verified first-hand this session" }, { "Term": "User model", "Definition": "What the model has inferred about the person using it.", "Plain language": "The picture a model builds of who it is talking to, such as age, gender or mood.", "Note / example": "", "Origin": "Field term", "Coined in / attributed to": "Chen, Y. et al., 2024. Designing a Dashboard for Transparency and Control of Conversational AI", "Definition drafted from": "Attribution held from the project context brief; NOT re-verified first-hand this session" }, { "Term": "Nothing", "Definition": "Nothing is rendered. The artefact is encountered rather than viewed.", "Plain language": "There is nothing visual to look at.", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" }, { "Term": "Unknown", "Definition": "Not yet determined.", "Plain language": "", "Note / example": "", "Origin": "Atlas schema", "Coined in / attributed to": "The Interpretability Atlas, 2026 — Ava and Claude", "Definition drafted from": "" } ] }