Research explorer

Every measurement, per seed

The technical companion to the story. Pick a seed, then inspect any saved analysis. Metric definitions are at the bottom. All values load directly from the repository's committed artifacts.

Training curves

Accuracy

Loss (log-x; slingshot events appear as spikes)

Cross-entropy loss on the same runs: train loss falls to near zero at memorization; test loss falls only at generalization. Sudden spikes are the slingshot instabilities.

Fourier structure

Embedding spectrum

Neuron frequency alignment

Attention patterns

Per-example attention (rows: query · columns: key · stronger colour = more weight)

Sampled held-out examples saved during analysis. The final row of each map is the “=” position — the one whose output becomes the answer.

Direct logit attribution

Mean direct contribution to the correct answer

Exact decomposition (no LayerNorm): component outputs at the “=” position, projected through the unembedding, sum to the model's centered logits. Near-zero bars for layer-0 components reflect indirect action, not unimportance — compare the ablation panel.

Ablation & activation patching

Mean-ablation: accuracy when each component is off

Sorted worst-first: the lower a bar, the more the model depends on that single component (necessity).

Patching: recovery (denoise) per component

How much of the clean answer margin returns when this one component's clean activity is copied into a corrupted run (sufficiency of its information).

Residual-stream patching (layer × position)

Recovery when a single residual-stream vector is patched clean into a corrupted run: the answer information sits at the operand positions entering block 0 and entirely at “=” after block 1.

Path-patching matrix

Sender → receiver edge effects

Fraction of the answer margin destroyed when only the sender→receiver path carries corrupted information (receiver input channel q/k/v resolved; blank cells are non-causal orderings).

Component ranking

Sorted by causal necessity (normalized performance lost under mean ablation). All four lenses shown for comparison.

Neuron-level refinement

Definitions

Faithfulness
Performance of the model with everything outside a component set mean-ablated, linearly normalized so 0 = everything ablated and 1 = intact model. The scale extrapolates beyond [0, 1].
Mean ablation
Replacing a component's activation with its average over the training set, computed separately per sequence position.
Activation patching (denoise)
Run a corrupted prompt, copy in one component's activation from the clean prompt, and measure how much of the clean answer margin returns (0 = corrupt baseline, 1 = fully clean).
Path patching
Corrupt only the direct connection from one sender component into one receiver input (q, k, v, or MLP input), everything else clean; report the fraction of the answer margin destroyed. Exact here because the residual stream is a linear sum of component outputs.
Direct logit attribution (DLA)
A component's output at the “=” position projected through the unembedding onto the correct answer (centered over answers). Direct path only.
Embedding Fourier power
Squared norm of the number-token embedding projected onto the orthonormal cos/sin basis over Z₉₇, summed per frequency (DC excluded).
Rotation-formula R²
Fraction of the centered logit tensor's variance explained by orthogonal projection onto cos/sin(2πk(a+b−c)/97) at the seed's key frequencies.
Slingshot
A transient optimizer instability: training accuracy collapses from ~100% and recovers within ~10³ epochs. Detected in the saved histories as train accuracy < 90% after having reached 99%.
Architecture

Sources: per-seed metrics.json, fourier.json, discovery.json, ranking.csv, patching.json, path_patching.json, dashboard_data.json under results/production/, assembled by visualization_site/build_data.py.