1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
//! #2015 — column-equilibration primitive test, and the revived planted
//! ~1e4-spread convergence gate now driven by the SOLVER-LEVEL fix.
//!
//! HISTORY: unit-RMS DATA equilibration was wired into the crosscoder fit
//! path (`run_auto_sae_crosscoder_fit`/`run_sae_crosscoder_fit`) and REVERTED
//! the same night: for a homoscedastic reconstruction objective, dividing
//! columns by their RMS is not a reparametrization — it changes the estimand
//! (noise-dominated columns are amplified to unit RMS and the fit spends
//! capacity explaining them). Measured on MSI 13021686 at the wired commit:
//! `zz2015_tiny_inner_crawl_terminates` refused at the co-collapse floor
//! (EV 0.4566 vs null 0.4583) and the planted transport-law verdicts
//! collapsed (phase R² 0.139, smooth R² 0.837). The `equilibrate_crosscoder_columns`
//! primitive below is kept (spec'd + unit-tested); the fit path passes a unit
//! `column_scale` and no longer calls it.
//!
//! The κ≈1e8 conditioning fix instead lives in the inner SOLVER's linear
//! algebra (design: issue 2015 comment 4949898801, landed in
//! `gam_solve::arrow_schur::reduced_solve::factor_dense_reduced_schur`):
//! Jacobi/Van der Sluis diagonal equilibration of the dense reduced-Schur
//! complement BEFORE its Cholesky factorization, undone exactly (`L = D·L̃`)
//! before the factor is returned — a pure numerical-conditioning aid with no
//! signature change and no estimand change (unlike the reverted data-frame
//! attempt). Because that fix is internal to the solver, the test below
//! exercises the UNMODIFIED `run_auto_sae_behavior_fit` front door directly on
//! an ill-conditioned planted fixture — no equilibration call anywhere in
//! this test — and the fit should converge because the solver now factors the
//! ill-conditioned Schur accurately.
use ndarray::Array2;
use crate::manifold::{
SaeBehaviorAutoFitRequest, SaeCrosscoderAutoFitConfig, equilibrate_crosscoder_columns,
run_auto_sae_behavior_fit,
};
/// Direct unit check of the equilibration primitive itself: every non-empty
/// column of the mutated target has unit RMS, and un-scaling by the returned
/// factor exactly reproduces the original (a diagonal, lossless change of
/// variables).
#[test]
fn equilibrate_crosscoder_columns_normalizes_and_round_trips() {
let mut target = Array2::<f64>::zeros((5, 3));
// Column scale ratio 1e4: column 0 ~ O(1e2), column 1 ~ O(1), column 2 ~ O(1e-2).
for i in 0..5 {
let t = i as f64;
target[[i, 0]] = 100.0 * (t + 1.0);
target[[i, 1]] = 1.0 * (t + 1.0);
target[[i, 2]] = 0.01 * (t + 1.0);
}
let original = target.clone();
let scale = equilibrate_crosscoder_columns(&mut target);
assert_eq!(scale.len(), 3);
for col in 0..3 {
let rms = (target.column(col).iter().map(|v| v * v).sum::<f64>() / 5.0).sqrt();
assert!(
(rms - 1.0).abs() < 1e-9,
"column {col} must be unit-RMS after equilibration, got {rms}"
);
}
// Undo: multiplying each column back by its scale must exactly reproduce
// the original (equilibration is a lossless per-column rescale).
for i in 0..5 {
for j in 0..3 {
let restored = target[[i, j]] * scale[j];
assert!(
(restored - original[[i, j]]).abs() <= 1e-9 * original[[i, j]].abs().max(1.0),
"round trip mismatch at ({i},{j}): {restored} vs {}",
original[[i, j]]
);
}
}
// The ratio of the largest to smallest scale reflects the planted 1e4 spread.
let scale_max = scale.iter().cloned().fold(0.0_f64, f64::max);
let scale_min = scale.iter().cloned().fold(f64::INFINITY, f64::min);
assert!(
scale_max / scale_min > 1e3,
"the equilibration scale must reflect the planted column-scale spread, got ratio {}",
scale_max / scale_min
);
}
/// Row-aligned circle activation (`p_x = 4`) and behavior probabilities
/// (`vocab = 4`, so `p_y = 3`) at `n` positions. `residual_amp` sets the
/// amplitude of ONE tangent direction (`e3`) relative to the fixed
/// `radius = 0.25` circular signal carried by the other two tangent
/// directions (`e1`, `e2`) — the column-scale-spread lever. The activation
/// construction does not depend on `residual_amp` at all, so the two fixtures
/// the test below compares share a BYTE-IDENTICAL activation target; only the
/// behavior block's own column-scale spread differs.
fn planted_fixture(n: usize, residual_amp: f64) -> (Array2<f64>, Array2<f64>) {
let inv_sqrt_two = std::f64::consts::FRAC_1_SQRT_2;
let b = [0.5_f64; 4];
let e1 = [inv_sqrt_two, -inv_sqrt_two, 0.0, 0.0];
let e2 = [0.5, 0.5, -0.5, -0.5];
let e3 = [0.0, 0.0, inv_sqrt_two, -inv_sqrt_two];
let radius = 0.25_f64;
let mut z = Array2::<f64>::zeros((n, 4));
let mut probs = Array2::<f64>::zeros((n, 4));
for i in 0..n {
let t = i as f64 / n as f64;
let theta = std::f64::consts::TAU * t;
// A small deterministic high-harmonic wiggle keeps the activation
// residual positive (so the fixed-λ_y fit sees a genuine two-block
// residual) without perturbing the circle the fitted decoder chases.
let wiggle = 0.05 * (9.0 * theta).sin();
z[[i, 0]] = theta.cos();
z[[i, 1]] = theta.sin();
z[[i, 2]] = 0.4 * (2.0 * theta).cos() + wiggle;
z[[i, 3]] = 0.4 * (2.0 * theta).sin();
// The behavior half-density: the circular signal in e1/e2 at fixed
// `radius`, plus a residual in e3 whose amplitude is the spread lever.
let residual = residual_amp * (3.0 * theta).sin();
let radial = (1.0 - radius * radius - residual * residual).sqrt();
let (cos_t, sin_t) = (theta.cos(), theta.sin());
for token in 0..4 {
let q = radial * b[token]
+ radius * (cos_t * e1[token] + sin_t * e2[token])
+ residual * e3[token];
probs[[i, token]] = q * q;
}
}
(z, probs)
}
fn fit_config() -> SaeCrosscoderAutoFitConfig {
let mut config = SaeCrosscoderAutoFitConfig::standard(1, 3);
config.max_iter = 40;
// Fixed-rho (λ_y held at its seed value) isolates the INNER solve's
// conditioning from the outer λ_y selection, matching the fast
// `zz2015_tiny_inner_crawl_terminates` convergence gate this test mirrors.
config.run_outer_rho_search = false;
config
}
/// #2015 — a behavior block whose own tangent-coordinate columns carry a
/// genuine ~1e4 within-block scale spread must still converge through the
/// UNMODIFIED `run_auto_sae_behavior_fit` front door (no equilibration call
/// anywhere in this test — the fix lives entirely inside the solver's dense
/// Schur factorization), and must recover essentially the SAME joint
/// reconstruction as an otherwise-identical well-scaled (spread ~1) fixture —
/// proof that the solver-level Jacobi/Van der Sluis fix neutralizes the
/// spread's effect on the inner solve's conditioning without changing what
/// gets fitted (the estimand is untouched this time: nothing in the fit
/// pipeline's data frame changed, only the solver's internal factorization
/// path — see `factor_dense_reduced_schur_reconstructs_original_illconditioned_matrix_2015`
/// in `gam-solve` for the primitive-level identity this end-to-end gate
/// depends on).
#[test]
fn planted_1e4_column_spread_behavior_block_converges_and_matches_well_scaled() {
let n = 96usize;
// Well-scaled: the residual direction's amplitude is the same order as the
// signal directions'.
let (z_well, probs_well) = planted_fixture(n, 0.05);
// Ill-conditioned: the residual direction's amplitude is 1e-4x the
// well-scaled fixture's — a deliberate ~1e4 within-behavior-block
// column-scale spread. The activation target is untouched (identical to
// `z_well`), so any degradation in ITS reconstruction quality is entirely
// a symptom of the shared joint solve's conditioning, not new activation
// structure.
let (z_ill, probs_ill) = planted_fixture(n, 0.05e-4);
assert_eq!(
z_well, z_ill,
"the activation target must be untouched by the spread lever"
);
let well = run_auto_sae_behavior_fit(SaeBehaviorAutoFitRequest {
activation: z_well,
probabilities: probs_well,
config: fit_config(),
cancel: None,
})
.expect("well-scaled fixture must fit");
let ill = run_auto_sae_behavior_fit(SaeBehaviorAutoFitRequest {
activation: z_ill,
probabilities: probs_ill,
config: fit_config(),
cancel: None,
})
.expect(
"the solver's Jacobi-conditioned dense Schur factorization must let the ~1e4-spread \
behavior block converge with NO data-frame equilibration in the fit path",
);
assert_eq!(ill.crosscoder.layers.len(), 2);
for layer in &ill.crosscoder.layers {
assert!(
layer.reconstruction_r2.is_finite() && layer.reconstruction_r2 > 0.9,
"{}: ill-conditioned fit must still reconstruct well, got {}",
layer.label,
layer.reconstruction_r2
);
}
assert_eq!(
ill.kl.infinite_rows, 0,
"no fitted row may decode off-simplex"
);
// The recovered ACTIVATION reconstruction (gauge-invariant: it is decoded
// output, not an internal coordinate) must match the well-scaled fixture's
// closely — the shared latent/gate the behavior block's spread could have
// destabilized did not, in fact, destabilize the activation block either.
let activation_diff = &ill.crosscoder.layers[0].fitted - &well.crosscoder.layers[0].fitted;
let activation_max_abs_diff = activation_diff
.iter()
.cloned()
.fold(0.0_f64, |m, v| m.max(v.abs()));
assert!(
activation_max_abs_diff < 0.1,
"activation reconstruction must match the well-scaled fixture's to tolerance, got max diff {activation_max_abs_diff}"
);
// The recovered BEHAVIOR (decoded distributions) must likewise match.
let mut behavior_max_abs_diff = 0.0_f64;
for (a, b) in ill
.fitted_probabilities
.iter()
.zip(well.fitted_probabilities.iter())
{
behavior_max_abs_diff = behavior_max_abs_diff.max((a - b).abs());
}
assert!(
behavior_max_abs_diff < 0.1,
"fitted behavior probabilities must match the well-scaled fixture's to tolerance, got max diff {behavior_max_abs_diff}"
);
}