Weight explorer — every stored matrix as a cloud, a heatmap, and bytes on disk

weights — help

Every learned matrix of the committed checkpoint, one mark per stored cell. The 3-D cloud spreads a matrix flat: x = column, z = row, height and color = the selected quantity. The heatmap shows the same cells as a flat grid. The 4-D shadows view projects the full per-cell 4-tuple four ways at once. The checkpoint bytes strip below is the actual file: header, config, index, and each tensor’s span at byte scale.

quantity

stored weight — the committed Q16.48 integer itself; no sample, no step, no derivation.
exact contribution — fp_mul(w[row,col], input[col]): the exact local Q16.48 product of one stored weight with one recorded input coordinate at the chosen trace sample and step. It is not attribution, not a gradient, and not the neuron’s total output — those would need more than one number. What makes it honest: each row’s products are re-summed here, in exact BigInt arithmetic, and asserted equal to the recorded output vector of the C oracle — on load for the default step, and again on every step you select.
wte / wpe are lookup tables: a row is selected, never multiplied, so no input vector exists and contribution mode is disabled.

4-D shadows

With a sample and step chosen, every cell of a multiplied matrix is a 4-tuple: (column, row, stored weight, exact contribution). Four dimensions do not fit one screen, so the view draws four 3-D projections — each panel drops one coordinate and its caption names the drop. All four panels share one camera: drag, wheel, or arrow keys turn them together, because they are four shadows of the same object, never four objects. Hovering a mark rings that same cell in all four panels. Marks are colored by each panel’s own height axis (weight or contribution, as captioned). Index axes are spread over the panel; value axes are scaled by that quantity’s own exact max |value| (display float; readouts exact). Undefined for wte/wpe: a lookup row has no input, so no contribution coordinate exists.

which input feeds which matrix

wq, wk, wv	xn_attn (attention-side RMSNorm output)
wo	ao (concatenated head mixes, before Wo); row sums + x0 = x_mid
fc1	xn_mlp (MLP-side RMSNorm output)
fc2	ReLU(h_pre); row sums + x_mid = x_out
lm_head	x_out (final residual); row sums = logits

encodings

#4dabf7 → #ffa94d	diverging ramp: blue negative, warm positive, panel-dark zero — symmetric about zero (display float; readouts exact). Heat is linear in value; the cloud and 4-D shadows use saturated poles (#339aff / #ffd43b) with a √|value| curve so faint points stay visible.
height	cloud only: same quantity as color, scaled by the tensor’s max |value| (display float)
dashed outline	cloud: the zero plane — marks above are positive, below negative
gold ring / cell	hovered or keyboard-focused cell, exact readout below
mark size / brightness	3-D depth (nearer = larger, brighter) — display geometry

scene links, motion, tour

The address bar always carries the current scene: ?tensor=…&view=…&sample=…&step=…&yaw=…&pitch=…&zoom=…. Copy the URL and the same page state — tensor, view, quantity, sample, step and camera pose — is restored exactly on load, so identical links render identical frames. auto-rotate orbits the cloud or 4-D camera gently (one revolution per 48 s); any drag, wheel or key on the stage pauses it, the button (or P) resumes it, and it never starts by itself when your system asks for reduced motion. The heatmap and the byte strip are never animated. auto-step walks every (sample, step) pair at one per second, wrapping at the end — combine it with auto-rotate for a moving view of the contribution sequence; it never starts by itself. tour steps through captioned preset scenes by hand (N/B) — nothing advances on a timer. For scripted capture the page exposes window.VIZ: ready, getScene(), setScene(s), sceneURL() and renderAt(ms) — the last draws the frame a paused scene would show after ms of rotation, synchronously and as a pure function of (scene, ms), so frame N of any capture is reproducible.

exactness contract

Exact (BigInt, never floats): stored weights, recorded trace vectors, every fp_mul product, every row-sum identity, all hex and decimal readouts, the FNV-1a-64 tensor hashes checked against the C harness’s record, and the byte offsets in the file strip. Display-only floats (labeled in place): screen projections, heights, colors, mark sizes. The startup self-check re-parses the embedded checkpoint, re-hashes all 9 tensors, and re-multiplies every applicable matrix row against recorded inputs at two trace steps.

keyboard

1 2 3	view: 3-D cloud / heatmap / 4-D shadows
W C	quantity: stored weight / exact contribution (3-D and heatmap; the 4-D view always shows both)
←↑→↓	cloud or 4-D focused: orbit (4-D: all four panels together) · heatmap focused: move the cell cursor
+ −	zoom the cloud / the shared 4-D camera
R	reset the camera
P	play / pause the gentle auto-rotation (cloud and 4-D only)
T	enter / leave the captioned tour
N B	tour: next / back (manual — scenes never auto-advance)
Esc	close this overlay (or leave the tour)
?	toggle this overlay

close (Esc)