# project: modelmap document: State of the art — technique taxonomy (Phase 2) author: Simon-Pierre Boucher contact: contact@spboucher.ai website: https://modelmap.io created: 2026-08-12 modified: 2026-08-12 status: reviewed

# State of the art — what can currently be measured inside an open-weight LLM

Phase 2 deliverable (charter §5). Built from the Phase 1 sweep (research/notes/4.1–4.10, research/bibliography.md, ~200 sources accessed 2026-08-12). Each technique carries the charter's eleven-field block, condensed. Citations live in the notes and bibliography; this document is the synthesis.

Cost classes. L = laptop-trivial (minutes, fits any 16 GB Mac). M = hours + tens of GB of activations (32–64 GB Mac, careful engineering). H = multi-day / >64 GB working sets / cluster-typical (out of local scope or marginal).

Mac feasibility bands (unified memory; fp16 unless stated): 16 GB → ≤3B fp16 or ≤8B 4-bit; 32 GB → ≤8B fp16 or ≤14B 4-bit; 64 GB → ≤14B fp16 or ~32B 4-bit. Backward passes ≈ 2–3× forward memory. All feasibility judgments are estimated from published numbers, not yet measured — Experiment H converts them to measured.

Epistemic flags. ✅ established · ⚠️ contested · ❌ debunked-in-part.

# Family 1 — Observational probing

# 1.1 Linear probes ✅(as decodability)/⚠️(as "use")

• Reveals: per-layer linear decodability of a property (language, domain, truth, syntax).
• Observational/causal: observational. Access: activations.
• Cost: L. Mac: yes, 0.5B–14B; storage is the binding cost (stream to disk).
• Failure modes: probe learns the task; decodable-but-unused; seed/dataset sensitivity; structure found in random models.
• Reproducibility: good with controls; historically often uncontrolled.
• Implementation: any framework; TransformerLens (MPS opt-in), baukit (device-agnostic); trivial in MLX.
• Main limitation: decodability ≠ causality — ceiling is Level 1 without intervention.
• Extension opportunity: probe maps with selectivity/MDL + FDR control as the atlas's baseline layer; probe-drift across quantization levels (unoccupied).

# 1.2 Logit lens ✅(diagnostic)/❌(quantitative)

• Reveals: layer-by-layer refinement of next-token prediction.
• Observational/causal: observational. Access: activations + unembedding.
• Cost: L. Mac: every runnable size; trivial in MLX.
• Failure modes: family-dependent validity (fails on BLOOM/OPT/some Llamas); early layers unreadable.
• Reproducibility: deterministic; validity varies by family.
• Main limitation: intermediate residual bases differ from the final one.
• Extension: use only as sanity layer; superseded by 1.3/1.4.

# 1.3 Tuned lens ✅

• Reveals: less-biased per-layer prediction trajectories via trained affine translators.
• Observational/causal: observational. Access: activations; gradients w.r.t. lens only.
• Cost: L–M (train lenses once per checkpoint). Mac: yes ≤14B.
• Failure modes: translator capacity objection (mitigated: affine-only).
• Reproducibility: good (validated to 20B).
• Main limitation: must be retrained per checkpoint and per quantization level.
• Extension: tuned-lens drift as a cheap quantization-deformation signal (expF) — not in the literature.

# 1.4 Patchscopes / early decoding (DoLa, Future Lens) ✅(tools)/⚠️(DoLa mechanism)

• Reveals: verbalized readouts of hidden states; multi-token-ahead information.
• Observational/causal: observational readout implemented via patching.
• Cost: L–M (≈2× inference). Mac: yes ≤14B.
• Failure modes: readouts inherit model confabulation; DoLa's layer story unreliable across families.
• Main limitation: qualitative; hard to aggregate into maps.
• Extension: structured Patchscopes corpora as labels for probe maps (combination not systematically tried).

# 1.5 Representation reading (RepE) ⚠️

• Reveals: concept directions via PCA on stimulus-pair activation differences.
• Cost: L. Mac: very friendly (repeng: CPU/MPS, exports GGUF control vectors).
• Failure modes: direction conflates stimulus correlates; single-direction assumption; weak evaluation (AxBench: prompting beats most methods).
• Main limitation: unreliable without baselines.
• Extension: reading-vector stability maps (across seeds of stimulus sets) — fits our replication-first program.

# 1.6 Concept erasure (INLP → LEACE) ✅(LEACE guarantee)/⚠️(amnesic inference)

• Reveals: behavioral dependence on a property, by provably-minimal linear removal.
• Observational/causal: interventional on representations (bridge to Family 2).
• Cost: L–M (LEACE closed-form). Mac: yes ≤14B.
• Failure modes: nonlinear residue can survive or grow; erasure-operator choice changes conclusions; INLP rank damage.
• Main limitation: linear-only guarantee.
• Extension: erasure-based capability maps (erase concept → measure per-task damage) as a causal upgrade of probe maps — partially tried (amnesic probing), never as a systematic atlas layer.

Meta (1.x): the controls doctrine. Selectivity (Hewitt & Liang), MDL (Voita & Titov), shuffled labels, capacity-matched probes, random-init baselines, ≥2 datasets, ≥3–5 seeds. Non-negotiable; encoded in Experiment A.

# Family 2 — Causal / interventional methods

# 2.1 Activation patching ✅ (with protocol pinned)

• Reveals: components causally mediating a behavior (single-mediator, input-relative).
• Access: activations (read/write hooks). Cost: L–M (2 forwards × components × pairs).
• Mac: yes 0.5B–8B; 14B slow. Memory-light, time-heavy — good Mac fit.
• Failure modes: conclusions flip with corruption type and metric; hydra/backup compensation; multi-mediator interactions missed.
• Reproducibility: good on classic tasks when protocol fixed.
• Implementation: TransformerLens/NNsight/pyvene/baukit; all MPS-workable; MLX = wrap modules.
• Main limitation: O(components) sweeps; single-mediator blindness.
• Extension: pre-registered patching protocols per map type (methodology.md); patched-maps across quantization levels (unoccupied).

# 2.2 Path patching ✅

• Reveals: causal importance of edges/paths (the IOI instrument).
• Cost: M (combinatorial). Mac: ≤3B–7B practical.
• Failure modes: hypothesis-space explosion + all of 2.1's.
• Main limitation: does not scale to full-model edge maps locally.
• Extension: use only to verify circuits proposed by 2.3/4.5 (screen-then-verify).

# 2.3 Attribution patching (AtP*, EAP, EAP-IG) ✅ (as screening)

• Reveals: first-order estimate of every patch effect in ~3 passes.
• Access: activations and gradients. Cost: L–M.
• Mac: ≤3B fp16 @16–32 GB, ≤8B @64 GB; gradients through 4-bit models are the open constraint; MLX autodiff is the native path to benchmark (expH).
• Failure modes: fails near-zero gradients (fixed by IG), saturated attention, residual stream; documented false negatives; second-order lies (2026).
• Reproducibility: good; public reimplementations replicate.
• Main limitation: an estimator — final claims need real patching.
• Extension: the default Mac circuit-scanner; calibration study (AtP error vs true patching) per model family is publishable.

# 2.4 Causal tracing (ROME-style) ⚠️/❌ (locating→editing inference)

• Reveals: where restored states recover corrupted behavior.
• Cost: M per fact corpus. Mac: yes ≤8B.
• Failure modes: Hase et al.: tracing does not predict editability; Gaussian noising is off-distribution and inflates mid-layer MLP localization.
• Main limitation: interpretation, not mechanics.
• Extension: rerun with symmetric-token corruption + resample baselines; treat as one voter in method-agreement maps, never alone.

# 2.5 Ablations (zero / mean / resample) ✅

• Reveals: necessity of components. Cost: L. Mac: all sizes; ideal for expB.
• Failure modes: zero/mean off-distribution; hydra effect; multiple comparisons.
• Norm: resample > mean > zero.
• Main limitation: necessity only; distribution-relative.
• Extension: ablation-scheme curves (all three schemes) as the atlas standard — no one publishes this systematically.

# 2.6 Causal scrubbing ✅(concept)/little-used

• Reveals: quantitative faithfulness of a full hypothesis graph.
• Cost: M–H. Mac: ≤1B realistically.
• Failure modes: unfaithful hypotheses can score well; combinatorial resample structure.
• Implementation: rust_circuit archived — no maintained macOS-friendly implementation exists (gap).
• Extension: a minimal MLX scrubbing kernel for micro-models would be new tooling.

# 2.7 DAS / interchange interventions (pyvene) ⚠️ (subspace-illusion debate unresolved)

• Reveals: whether a high-level causal model lives in learned subspaces (IIA).
• Access: activations + gradients (rotation trained). Cost: M. Mac: ≤7B.
• Failure modes: dormant-pathway illusion (Makelov) vs over-strict criterion (Wu reply); with unconstrained featurizers, causal abstraction is vacuous (Non-Linear Representation Dilemma).
• Main limitation: optimized bases need convergent evidence.
• Extension: confidence.md records featurizer class; DAS results always paired with natural-basis patching.

# 2.8 Steering vectors (ActAdd/CAA/ITI) ✅(existence)/⚠️(reliability)

• Reveals: linear sufficiency of behavior directions.
• Cost: L. Mac: the most Mac-friendly causal technique; works on GGUF via llama.cpp control vectors.
• Failure modes: input-heterogeneous, anti-steering, OOD-brittle, non-identifiable; loses to prompting baselines (AxBench).
• Main limitation: evaluation, not mechanism.
• Extension: steerability maps with variance and baselines per model/quantization — the literature reports means, not distributions.

# Family 3 — Feature decomposition

# 3.1 Superposition theory ✅(toy)/⚠️(real models)

• Reveals: why polysemanticity exists; predicts feature geometry under sparsity.
• Cost: L (toy models in seconds). Mac: trivially; ideal MLX testbed with planted ground truth.
• Main limitation: linear-feature assumption is not a theorem (multi-dimensional features exist).
• Extension: planted-superposition testbeds to validate our SAE pipeline before real models (validation-first rule).

# 3.2 Sparse autoencoders (ReLU → TopK/JumpReLU/BatchTopK/Matryoshka) ✅(method)/⚠️(faithful "true features")

• Reveals: overcomplete sparse dictionaries of activation directions.
• Access: activations at scale. Cost: the cost driver — 10⁸–10⁹ tokens/layer; 500M tokens ≈ 4 TB fp16 if materialized → streaming mandatory.
• Mac: M for 0.5B–2B models at 4k–65k width (SAELens documents an M3 Max run); H for 7B+ at ≥131k width.
• Failure modes: ~30% seed replication (the noise floor); dataset dependence; absorption/splitting; proxy metrics ≠ downstream utility; dark-matter reconstruction error.
• Reproducibility: the weak point — quantified and low.
• Implementation: SAELens (MPS works), dictionary_learning, sparsify (no-cache streaming, CUDA-leaning), mlxterp (MLX, embryonic).
• Main limitation: decompositions are pragmatic, not canonical.
• Extension: seed-replication rate as a published per-SAE metric; buffer-shuffle vs full-shuffle replication study (unpublished); SAE feature drift under quantization (one 2026 paper — extend it).

# 3.3 Transcoders / CLTs / crosscoders ✅(ascendant)

• Reveals: sparse input→output MLP maps (analyzable feature-to-feature weights); cross-layer features; cross-model diffs.
• Cost: M (≤2B) to H (7B+, cross-layer multiplies capture). Mac: ≤2B pairs realistic.
• Failure modes: MLP-only coverage; error nodes dominate on some prompts; diffing latents can be sparsity artifacts (fixed by BatchTopK + latent scaling).
• Main limitation: analyses describe a replacement model.
• Extension: an MPS/MLX transcoder recipe at 0.5–2B with replication reporting.

# 3.4 Automated interpretability (auto-interp) ✅(tooling)/❌(label quality claims)

• Reveals: natural-language labels for latents, at scale.
• Cost: M, dominated by LLM inference — local via the maclustr LLM endpoint.
• Failure modes: low precision/recall under targeted tests; top-activation bias; metrics fail to distinguish trained from random transformers.
• Rule: never above Level 1 without intervention scoring.
• Extension: intervention-scored labels only; label stability across explainer models as an uncertainty channel.

# 3.5 SAE evaluation (SAEBench, RAVEL, MIB causal-variable track) ✅(best practice)/⚠️(metric stability)

• Reveals: which decompositions are worth anything (proxy + downstream + intervention metrics).
• Cost: L–M with public SAEs (Gemma Scope on 2B fits 16 GB). Mac: yes.
• Failure modes: benchmark rankings unstable across tasks/seeds; Goodharting.
• Extension: port a SAEBench subset as modelmap's SAE gate; add our replication-rate metric to it.

# Family 4 — Circuits and computational graphs

# 4.1 Induction heads ✅ (most replicated finding in the field)

• Reveals: the [A][B]…[A]→[B] two-head circuit; cross-architecture universality.
• Cost: L (one forward with attention capture). Mac: to 32B 4-bit via MLX.
• Main limitation: behaviorally defined; heads are multifunctional.
• Extension: the calibration landmark for any cross-model coordinate system (if alignment can't match induction heads, it fails); first atlas map type.

# 4.2 IOI circuit & head taxonomies ✅ (details ⚠️)

• Reveals: name movers, S-inhibition, copy suppression, successor heads; faithfulness criteria.
• Cost: L–M. Mac: GPT-2-scale trivial; 7B head scans M.
• Failure modes: hydra/backup heads destabilize ablation importance; ~87% faithfulness; task-distribution-relative labels.
• Extension: head-taxonomy maps across sizes/quantizations with stability scores (expD substrate).

# 4.3 MLP key–value memory view ✅(lens)/⚠️(literal reading)

• Reveals: input weights as pattern keys, output weights as vocabulary updates.
• Cost: L, near-weight-only. Mac: trivial.
• Extension: pairs with Family 5 as a weight-first map layer; vocabulary-projection maps with early-layer validity flags.

# 4.4 Attribution graphs / circuit tracing (CLT-based) ✅ (candid limits)

• Reveals: per-prompt causal graphs over features (planning, multi-hop, refusal mechanics).
• Cost: M with pretrained CLTs (Gemma-2-2B graphs ≈ 15–40 GB → 64 GB Mac plausible, 16 GB not); H to train CLTs at 7B.
• Failure modes: error nodes unexplained; frozen attention (QK interference); per-prompt graphs need aggregation to become maps; many graphs unreadable.
• Implementation: circuit-tracer (CUDA-first) — MPS/MLX port is a concrete high-value engineering task.
• Extension: graph aggregation across prompt corpora with replication stats — the "map" layer Anthropic hasn't published.

# 4.5 Automated circuit discovery (ACDC → EAP-IG → edge pruning) ✅ (method choice ⚠️)

• Reveals: minimal subgraphs reproducing behavior under faithfulness objectives.
• Cost: ACDC M–H; EAP-IG L–M (the Mac default); edge pruning M–H (scaled to 13B on cluster-class compute — fits M3U96a/b).
• Failure modes: inherits patching baselines; greedy/independence assumptions; metric overfitting.
• Extension: circuit discovery under multiple ablation schemes reported as curves.

# 4.6 Circuit hypothesis testing (circuitry; faithfulness critiques) ✅

• Reveals: pass/fail on equivalence/independence/minimality; robustness of faithfulness itself.
• Cost: L–M ≤2B. Mac: yes.
• Rule adopted: every promoted circuit runs the battery; results in confidence.md.

# 4.7 Sparse feature circuits ✅(method)/⚠️(robustness)

• Reveals: circuits over SAE/transcoder features (more interpretable than heads); SHIFT debiasing.
• Cost: M with public SAEs. Mac: Pythia/Gemma-2-2B scale yes.
• Failure modes: inherits SAE seed-fragility — circuits over non-replicating features are sandcastles.
• Extension: feature-circuit replication rate across SAE seeds — directly quantifies the sandcastle problem; not in the literature.

# 4.8 Weight-sparse transformers (circuits by construction) ✅(tiny models)

• Relevance: planted-ground-truth benchmark for validating discovery pipelines, not for mapping existing checkpoints.

# Family 5 — Weight-space cartography (zero/low forward-pass)

# 5.1 SVD spectra & LASER ✅(phenomenon)/⚠️(interpretation)

• Reveals: rank structure per matrix; behaviorally-loaded high-order components.
• Cost: L (SVD of 4096×11008 = seconds on Accelerate). Mac: fully native.
• Failure modes: (layer, matrix, rank) sweeps invite multiple comparisons; ID-only proxies diverge OOD (Friedman illusion).
• Extension: per-matrix spectral maps as the atlas's cheapest layer; test whether they predict A–D properties (expE — the unoccupied gap).

# 5.2 Heavy-tailed spectral diagnostics (WeightWatcher) ✅(trends)/⚠️(fine grain)

• Reveals: training quality per layer from ESD power-law fits, zero data.
• Cost: L; pip-installable, runs on Apple Silicon today.
• Failure modes: fit-range sensitivity; Simpson's paradox documented.
• Extension: alpha profiles vs probe/localization maps on the same checkpoints — the direct expE experiment.

# 5.3 Outliers: super weights, massive activations, sinks ✅

• Reveals: a handful of weight-visible scalars with catastrophic causal load; the weight→activation bridge.
• Cost: L (scan) + minutes (verify by zeroing). Mac: yes (MLX 4-bit PPL checks).
• Extension: super-weight/outlier maps per model+quantization; outlier-aware controls for all similarity metrics (CKA is dominated by these dims).

# 5.4 Representation/weight similarity (CKA/SVCCA/Procrustes) ✅(tool)/❌(standalone evidence)

• Reveals: cross-layer/cross-model geometry alignment.
• Cost: L on cached activations. Mac: yes.
• Failure modes: outlier domination; manipulability; metric disagreement (ReSi); saturation on LLM pairs; estimator drift.
• Rule: ≥2 metrics + statistical tests + outlier-stripped variants + random/width-matched nulls.

# 5.5 Layer redundancy / depth maps ✅

• Reveals: deletable contiguous deep blocks (angular similarity).
• Cost: L (small calibration set). Failure modes: perplexity survives while reasoning dies.
• Extension: redundancy maps as a target for weight-only predictors; feeds localvm working-set questions (Q5).

# 5.6 Stitching ⚠️ | 5.7 Git re-basin ⚠️(transformers) | 5.8 Mode connectivity ✅/⚠️

• Stitching: protocol-sensitive; LLM literature thin; tokenizer collision cross-family.
• Re-basin: fails on transformers; the LLM-relevant analogue (cross-size alignment) is open = Experiment G.
• Mode connectivity: cheap base↔instruct interpolation profiles are the Mac-friendly artifact.

# 5.9 Task arithmetic / delta maps ✅(engineering)/⚠️(interpretation)

• Reveals: θ_ft − θ_base as sparse, composable task vectors (90%+ droppable).
• Cost: L (state-dict math). Mac: 7–14B fine.
• Extension: systematic base-vs-instruct delta atlas (per-matrix norms, delta rank, delta spectra) — no published atlas exists; ideal cheap modelmap product.

# 5.10 Crosscoder model diffing ⚠️ + weight-side monitoring (2025) — see 3.3

• The field is converging on weight-first diffing (Watch the Weights; weight-diff interpretation) — adjacent to, not overlapping, our pre-screen program.

# 5.11 MoE router structure ✅(negative result known)

• "Expert = topic" debunked; router-weight geometry is an open, cheap target. Mixtral-class needs 4-bit on 64 GB.

# 5.12 Weight-space learning ✅(field)/nascent for LLMs

• Model-zoo regime breaks at LLM scale; borrow representations (per-matrix statistics vectors), not the zoo methodology.

# Family 6 — Knowledge localization and editing

# 6.1 ROME/MEMIT/PMET/EMMET/AlphaEdit ✅(mechanics)/❌(localization inference)

• Reveals: facts are rewritable via mid-layer MLP updates; AlphaEdit's null-space projection tames sequential collapse.
• Cost: M (covariance stats = hours of forwards; edits cheap). Mac: feasible ≤8B, but all repos are CUDA-first — an MLX editing port is an open tooling contribution.
• Failure modes: localization ≠ editability (Hase); sequential collapse; ripple-effect failures (in-context editing beats parametric); general-ability damage.
• Extension: editing as perturbation probe (measure what else moves), never as localization evidence.

# 6.2 Knowledge neurons ❌ (one-fact-one-neuron)

• Degenerate/redundant neurons; token-pattern confound. Keep only as a historical baseline for expB nulls.

# 6.3 Factual recall pipeline (Geva) + attribute lens ✅(scaffold)/⚠️(details)

• Reveals: subject enrichment (MLPs) → relation propagation → attribute extraction (attention heads); ~48% of relations linearly decodable.
• Cost: L–M. Mac: ≤7B yes.
• Extension: the concrete map types for Q1: enrichment maps (MLP) and extraction maps (heads), with the additivity caveat (multiple mechanisms sum).

# 6.4 Multilingual knowledge sharing ✅(qualitative)/⚠️(mechanism)

• Reveals: partially language-agnostic substrate, English-hub behavior, poor cross-lingual edit transfer.
• Mac: ≤8B multilingual models fine. Extension: FR/EN localization atlas (expB) has clean prior art to extend and a personal-fit corpus advantage.

# Cross-cutting instruments

# 7.1 Cross-model comparison

Universal neurons (1–5% across seeds — methodology template: correlate → taxonomize → ablate); induction-head landmark; Tigges consistency across training/scale (the expG band is 70M–2.8B); SAE feature matching (moderate overlap, seed-noise-limited); platonic convergence deflated to local/topological alignment (2026); Rosetta-neuron mining untested on LLMs (candidate coordinate system); tokenizer alignment is unsolved — within-family comparison is the safe default; cross-family maps must declare their convention in the schema.

# 7.2 Ground-truth validation & statistics

Tracr / InterpBench / MIB as pipeline validators (all Mac-runnable); circuitry hypothesis tests; causal abstraction as the Level-3 formal spine plus featurizer-class disclosure; Dead-Salmons doctrine (random-init nulls, FDR control, uncertainty quantification) — adopted into src/modelmap/stats requirements.

# 7.3 Quantization × internals (the thin shelf)

Five shallow papers; effects "inconsistent in direction"; one SAE-drift study (2026). Nothing on probe drift, circuit survival, Q2/Q3, GGUF-runtime capture, or weight-geometry-predicts-quantization-sensitivity. The single clearest open field this survey found.

# 7.4 Systems substrate (macOS)

Split regime: probing/patching corpora (≤10⁶ tokens) → cache everything; SAE corpora → stream, never store. mmap likely wins on Apple NVMe (unmeasured); APFS clones = free store snapshots for provenance; MLX = only Python path to quantized-model capture; llama.cpp eval-callback = only path into actually-quantized compute (unwrapped); MPS hooks work but sync/fallback costs unmeasured. All quantitative claims here are expH's to measure.

# Overlaps and redundancies (what is secretly the same thing)

1. Patching ≡ interchange intervention ≡ causal mediation — one method, three vocabularies (mech-interp, causal-abstraction, NLP-fairness). DAS adds a learned basis; Patchscopes adds a verbal readout; causal tracing adds Gaussian noise (its weakest link).
2. Tuned lens ≈ per-layer linear probes onto vocabulary — a probe family, not a separate epistemic category; inherits all probe pitfalls.
3. LEACE ≈ steering's inverse — erasure removes a direction, steering adds one; both test linear-direction causality and should share controls (random directions).
4. SAE / transcoder / crosscoder / CLT = one estimator family (sparse dictionary on activations) differing in domain (layer, MLP map, model pair, layer stack); they share the seed-replication pathology, so replication findings transfer.
5. Faithfulness metrics ≈ causal scrubbing lite — circuit faithfulness under resample ablation is scrubbing's score without the hypothesis-graph machinery.
6. Layer-redundancy maps ≈ cross-layer CKA (angular similarity is a similarity metric) — inherits CKA's outlier pathology; strip sink dimensions first.

# Combinations already tried (do not claim as new)

• Probes + causal verification: amnesic probing; RAVEL; MIB causal-variable track.
• SAE features + circuits: sparse feature circuits; attribution graphs.
• SAE + steering: AxBench (negative); feature steering in Scaling Monosemanticity.
• Weight diffs + interpretation: Watch the Weights; Learning to Interpret Weight Differences (2025).
• Patching + automated search: ACDC/EAP/EAP-IG/edge pruning.
• Editing + localization: ROME→Hase (the canonical failed combination).
• Atlas + API hosting: Neuronpedia (without provenance contracts).

# Ideas that seem novel but are known (novelty traps)

• "Label every neuron/feature with an LLM" → Transluce, OpenAI 2023, Delphi. Known, and known-unreliable.
• "Compare models with CKA to find shared structure" → large literature and a debunking literature.
• "Find the layers where facts live and edit them" → ROME line + Hase refutation.
• "A browsable atlas of features" → Neuronpedia/Neuroscope. The provenance/confidence layer is the gap, not the browser.
• "Predict quality from weight spectra" → WeightWatcher (2019–2021).
• "Mine correlated units across models" → Rosetta Neurons (vision) — LLM transfer is open but the method is not new.
• "MLX interpretability library" → mlxterp exists (embryonic) — extend or differentiate, cite regardless.

# Cost frontier summary (estimated; expH converts to measured)

# Conclusions feeding Phase 3

1. The unoccupied high-value squares: quantization × internal maps, weight-only pre-screens validated against activation-level ground truth, replication-rate reporting as a first-class artifact metric, quantized-runtime capture (llama.cpp/MLX), provenance-first atlas schema.
2. The methodological spine is buyable off the shelf: Tracr/InterpBench validation, circuitry tests, SAEBench subset, controls doctrine — integrate, don't reinvent.
3. Everything promoted to the atlas needs: pinned patching protocol, ablation-scheme curves, ≥2 similarity metrics with nulls, featurizer-class disclosure, seed replication rate, per-dataset agreement. These become the map-card schema fields.