TimesFM vs Chronos-2 vs Sundial vs Chronos-Bolt

How to read it
Each bar is a count of series sharing that label. `tier` distinguishes leakage-safe synthetic series from real-derived ones; `generator` names the construction mechanism.

Limitations
This renders only the corpus this run actually loaded (`data.path`), never a dev-vs-private side-by-side -- the private split is loaded independently, only by the `confirm` stage, and is not available here to compare against without loading a second corpus this stage was not given.

How to read it
Each panel is one real series from this corpus, not an average or a synthetic composite -- picked per family (`np.random.default_rng` seeded on `run.seed`) rather than by a head slice, since this corpus is written grouped by generator and a plain prefix would silently select a family-skewed sample.

Limitations
Each panel has its own y-axis: series scales are not comparable across panels, and this is illustrative, not a diversity measurement (that is Figure 3 below, computed in feature space).

How to read it
Computed by `benchmark_validation` in **catch22 feature space, never on UMAP coordinates** -- distances here are the real diversity measurement, unlike Figure 4 below. This particular validation report's provenance is UNVERIFIED (item C1.3) -- it is matched to this corpus by configured path only, not by a cross-checked digest, so it may in principle describe a different corpus than the one this run actually used.

Limitations
A high redundancy fraction can be a true finding, not a defect -- e.g. many pure-seasonal series sharing one period differ only in noise and are legitimately near-duplicate in shape. Read this figure alongside effective dimensionality, never alone.

How to read it
Open the linked standalone HTML file directly (`benchmark_validation`'s own `plot_embedding` output) for the interactive 3-D view, colored by generator.

Limitations
Three things this repo insists on stating every time UMAP is mentioned: diversity is measured in catch22 feature space, never on UMAP coordinates (Figure 3 above is the real measurement); UMAP distorts global distances and cluster sizes, so gaps in the 3-D plot are not quantitative; and this is a CORPUS feature-space embedding, a different object entirely from an activation-space embedding of a model's internals.

How to read it
A series below the gate threshold is rejected as too close to a real reference series. Only aggregated quantiles are persisted, not the raw per-candidate distance array, so this is a quantile comparison, not a full histogram.

Limitations
This gate measures INSTANCE-level leakage only (an exact copy of a reference series) -- it says nothing about DISTRIBUTIONAL leakage (a model having seen data *shaped like* this), which is carried by construction for every real-derived-tier sample and cannot be audited by this or any other check.

How to read it
`not_recorded` means no audit ever ran; `not_checked` means the leakage gate ran with zero reference series (`--references none`) so every candidate passed trivially -- NOT the same as a corpus checked and found clean. `inconclusive` and `not_verifiable` rows are claims nothing in this repo can currently resolve, rendered rather than omitted.

Limitations
A `measured`/`clears` verdict is evidence the corpus is sound by the checks that exist here -- it is not proof no other check would find a problem, and rows 4 and 7 are permanently open questions this section states rather than answers.

How to read it
Read the 'observed (headline)' column first. Blocks-captured measures capture_layer_stride loss only and is 1/1 even for a regex that covers just an encoder; params- and FLOPs-captured use whole-model denominators and so do see that. Any model under 90% gets every depth-located finding qualified automatically in the findings list. A headline marked 'or less' is an upper bound taken from the capture-pass/forecast ratio, used when per-block FLOPs did not resolve by name — it still gates correctly, since a model failing the bar even optimistically has certainly failed it.

Limitations
The FLOPs denominators come from the same measured counter as the table above, so a 'suspicious_low' FLOPs check disarms these fractions too. 'FLOPs of full forecast' is measured at this run's decoding settings — for a sampled decoder, num_samples moves the denominator, so it is a property of this run, not of the architecture alone. A blank means not measurable here, which is not the same as full coverage.

How to read it
Each curve rises from 0 toward the fraction of that model's own measured forecast FLOPs its captured blocks account for. Read a depth-located claim about a model whose curve is far below another's at the same x as being about a different amount of computation — a claim at depth 0.5 is only 'halfway through the model' if the curve is near 0.5 there too. Chronos-Bolt, Sundial and TimesFM top out well short of 1.0 here — that matches the coverage table above rather than contradicting it, and is the reason a depth-located claim about those models carries an automatic qualifier.

Limitations
Normalized by predict.flops (the full forecast, incl. every sampled decode pass) when `budget.measure_predict` measured it — the fix (D2) / named, now built — falling back to the capture-pass FLOPs only when no forecast measurement exists, in which case a curve reaching 1.0 means 'covers its own forward pass', not 'covers the forecast'.

How to read it
Lower is better on the y axis; further left is cheaper on the x axis. A model in the lower-left is winning outright. A model that is lower but further right is buying its accuracy with compute, and whether that is a good trade is a deployment decision this chart deliberately does not make for you. Both x axes are logarithmic, because model sizes worth comparing differ by orders of magnitude.

Limitations
FLOPs are a measured proxy for cost, not a measure of wall-clock latency — a model can be FLOP-cheap and slow, or FLOP-heavy and fast, depending on how well its shapes map onto the hardware. Both are measured at this run's single context length and horizon and do not extrapolate to others. Models whose internals could not be measured are absent from this chart rather than plotted at zero.

How to read it
Latency is a median over repeated calls on this machine and is the least portable row here — read FLOPs and parameters for anything meant to transfer. FLOPs/series divides by the measured batch, so it is comparable across models even when their batch sizes differ. In the compute-normalized panel, down and to the left is better: lower MASE for less compute.

Limitations
FLOPs are not latency: a model with fewer FLOPs can be slower if its shape suits the hardware worse, and a fused or custom kernel the counter cannot see is undercounted (the flops_sanity row states whether the measurement clears an analytic lower bound — a suspicious_low verdict means this model's compute-normalized numbers should not be trusted). The parameter role split is a positional heuristic, not an architectural fact, and is reported as such per model. Cost is measured on this run's context length and horizon only; both scale the answer, attention super-linearly.

How to read it
A small 'bin width' fraction means the tokenizer can distinguish fine detail; a large one means real amplitude differences within a series collapse onto the same token. Read 'series with any clipping' alongside it — clipping means a value was extreme enough to fall entirely outside the tokenizer's bin range, not just coarsely binned.

Limitations
Only meaningful for a re-quantizing tokenizer — a continuous embedding has no bin geometry for this probe to measure and is reported as 'not applicable', never a fabricated zero (Chronos-2, Chronos-Bolt, Sundial and TimesFM in this run). The bin geometry is read directly from the checkpoint's own tokenizer, never hardcoded.

How to read it
The residual is the mean absolute difference between the original forecast and the rescaled-then-unscaled one, normalized by each series' own typical step size (the same scale MASE uses) so it's comparable across series. Near zero means the model is effectively scale-equivariant at that factor; a residual that grows with the scale factor means extreme scales genuinely confuse the model's own internal normalization, not just numerical noise. A model flagged as 'not scale-equivariant through this adapter' crossed a threshold set relative to the other models measured in this same run (10x their median worst residual, floored at 1.0 context-scale unit) — read this as 'this adapter likely bypasses the checkpoint's own input normalization', not as an architectural property of the model family.

Limitations
Non-finite series (overflow/underflow at an extreme scale factor) are excluded from the residual average but counted separately — a model that fails outright at 1000× is a different, more severe finding than one that degrades gracefully, and averaging the two together would hide which one happened.

How to read it
'Graceful' means accuracy degrades roughly evenly as more recent context goes missing; 'cliff' means it stays flat until one specific length, then jumps sharply — a hard floor below which the model has almost nothing useful left to work with. 'no_degradation' means withholding recent context didn't measurably hurt this model on this corpus.

Limitations
This is a genuinely different axis from phase-sensitivity (sub-patch-width front shifts) and context-scaling/E20 (which keeps recent history intact and only drops old history) — do not read this section as a duplicate or a replacement for either. A model that cannot forecast far enough to bridge a large staleness gap has that length point skipped rather than crashing the whole diagnostic; fewer than 3 surviving points means the shape could not be characterized at all for that model. Baseline/most-truncated MASE and the shape classification all use the per-length MEDIAN, not the raw mean, over series — at the shortest available-context lengths, a handful of series are locally near-flat over that short a window, which sends MASE's own denominator toward its floor and that series' MASE toward an astronomical, non-representative value regardless of forecast quality. The raw mean is shown alongside specifically so a large gap between the two columns is visible as the tell that this is happening, rather than silently absorbed into one number.

How to read it
'errors' and 'handled' are both acceptable outcomes (a loud failure is honest; genuine handling is a real capability). 'propagates' is the dangerous case: a non-finite forecast with no error raised looks like an ordinary result until something downstream notices the NaN. 'mixed' means the verdict depends on where in the context the value went missing.

Limitations
Checked at only three hand-placed positions (front/middle/back of context) and one missing-fraction (default 5% of context) — not exhaustive. A model verdict here says nothing about forecast QUALITY on ordinary, well-formed input; it only characterizes what happens at this one specific failure mode.

How to read it
Height is only meaningful relative to the other bars for the same model — it is a ranking device, not a quantity with units. A screen worth trusting picks layers in the middle-to-late range and clearly separates them from the rest; a flat profile means the screen found nothing to distinguish layers by, and the selection is close to arbitrary.

Limitations
This is a cheap proxy for interestingness, not interestingness itself — it was chosen by a bake-off against random and uniform-stride nulls on one corpus and one checkpoint pair. It also scores only the layers this run actually captured, so if the header above warns that not every block was screened, a better layer may exist that was never looked at.

How to read it
Height is only meaningful relative to the other bars for the same model — it is a ranking device, not a quantity with units. A screen worth trusting picks layers in the middle-to-late range and clearly separates them from the rest; a flat profile means the screen found nothing to distinguish layers by, and the selection is close to arbitrary.

Limitations
This is a cheap proxy for interestingness, not interestingness itself — it was chosen by a bake-off against random and uniform-stride nulls on one corpus and one checkpoint pair. It also scores only the layers this run actually captured, so if the header above warns that not every block was screened, a better layer may exist that was never looked at.

How to read it
Height is only meaningful relative to the other bars for the same model — it is a ranking device, not a quantity with units. A screen worth trusting picks layers in the middle-to-late range and clearly separates them from the rest; a flat profile means the screen found nothing to distinguish layers by, and the selection is close to arbitrary.

Limitations
This is a cheap proxy for interestingness, not interestingness itself — it was chosen by a bake-off against random and uniform-stride nulls on one corpus and one checkpoint pair. It also scores only the layers this run actually captured, so if the header above warns that not every block was screened, a better layer may exist that was never looked at.

How to read it
Height is only meaningful relative to the other bars for the same model — it is a ranking device, not a quantity with units. A screen worth trusting picks layers in the middle-to-late range and clearly separates them from the rest; a flat profile means the screen found nothing to distinguish layers by, and the selection is close to arbitrary.

Limitations
This is a cheap proxy for interestingness, not interestingness itself — it was chosen by a bake-off against random and uniform-stride nulls on one corpus and one checkpoint pair. It also scores only the layers this run actually captured, so if the header above warns that not every block was screened, a better layer may exist that was never looked at.

How to read it
Read this as "where this method thinks it's worth spending expensive compute", not as a finished interpretability claim -- nothing else in the report depends on it being right, only on what gets trained downstream if `sae.targets` is left empty.

Limitations
The default method (`work_bend`) won a single-corpus, single-checkpoint-pair bake-off (2026-08-05) — real signal, not a settled cross-architecture rule. On some models/corpora the theoretical best possible selector has little room to beat a free uniform-stride null; a method not beating it there is not necessarily broken. `factor_emergence` has a known, diagnosed weighting flaw and underperformed in that bake-off.

How to read it
Higher means the representation is using more of its available capacity at that layer; a dip means activity is collapsing onto fewer effective directions there. Compare *shape* across depth within a model (expansion then compression is a common pattern) rather than comparing raw values across models with different widths — a 64-dim and a 48-dim layer aren't on the same scale.

Limitations
This is a per-model diagnostic, not a cross-model comparison — two models can have very different effective-dimensionality curves and still solve the forecasting task equally well. It also only sees the *window-pooled* representation, so within-window structure that pooling discards isn't reflected here.

How to read it
Near 1.0 means this layer is still basically a linear function of raw local window stats (typical right after an embedding, before much mixing). Near 0 means the representation encodes something the simple local baseline can't see at all, most often because attention has already mixed information across positions/windows before this depth — that decorrelates a *window's* state from that *same window's* own raw statistics even though the information hasn't vanished, it's just no longer purely local.

Limitations
A curve that's low and flat across every depth (including layer 0) usually means mixing saturates fast, which is expected for architectures with very few tokens per context (a handful of patches attend to each other almost immediately). Architectures with many more tokens per context typically show a more gradual decline instead. Low here is neither good nor bad on its own: it says the representation isn't simply local, not whether it's useful.

How to read it
The dotted line is chance (majority-class baseline for however many families exist); the dash-dot line per model is a permutation null — the identical probe refit with family labels reshuffled at the series level, showing how high accuracy can get by chance alone on this exact split and architecture. A peak well above *both* lines means that depth linearly encodes which kind of series this is. The peak layer is called out in the findings below as where family information is most accessible.

Limitations
'Decodable' is not the same as 'used by the forecast' — a layer can carry perfect family information the model never actually reads out (cross-check against the tuned lens and behavioral sensitivity to see what's causally load-bearing). Accuracy is also capped by how separable the configured families actually are in the benchmark, not just by the model. The permutation null is the stronger of the two floors — the majority-class chance line ignores the classifier and split entirely, so a probe can clear it while still scoring within the null's own p95, which means its apparent decodability could be search/split artifact rather than real information.

How to read it
Columns come from different stages and mean different things. eff. dim and CKA to input are descriptive geometry (how many directions the layer's representation spreads over, and how much it still resembles raw input statistics). probe acc. is decodability of the data family from that layer, worth reading as probe − chance. screen score is the cheap interestingness heuristic that chose sae.targets: auto, and screened in marks its picks. skip-lens MASE is the forecast quality obtainable from that layer through the model's own head — it descending and then flattening is the crystallization the Forecast lens section measures. mean Δact and best restoration are the layer's average reaction to the corruption battery and the most any single patch at that layer recovered. closest layer in other model is that layer's argmax over the cross-model CKA matrix.

Limitations
A blank cell means that stage did not measure that layer — stages run at independent strides (l3.patching.layer_stride, capture_layer_stride), so absence is a config fact, not a zero. Nothing here is causal except best restoration, and that is causal within its own model only. Relative depth is a fraction of the model's whole stack including surfaces this run never captured, so the last row of a partially-observed model is not that model's output.

How to read it
A curve that drops toward the final-MASE line early in depth means the forecast is already largely formed well before the last layer ('crystallizes early'); a curve that only reaches it at the last point means the model needs its full depth. The 'crystallization depth' finding below is the first relative depth whose MASE lands within a configurable tolerance of final.

Limitations
Like classic logit lens, early layers can be miscalibrated for reasons unrelated to information content (the final block/head wasn't trained to decode them) — a high early MASE doesn't prove the forecast isn't already linearly present, only that this particular readout can't see it yet. That's exactly the gap the tuned lens below (a probe fit specifically for each layer) is designed to close.

How to read it
High solid-line R² early in depth means the forecast is already linearly recoverable from that layer even if the skip lens (constrained to the model's own head) can't show it yet. The dashed 'ground truth' line is usually lower and caps out at whatever the model's own final-layer accuracy allows — it can't exceed how good the forecast itself is.

Limitations
A linear probe only detects *linear* readability; a layer could carry the forecast in a nonlinear form invisible here. Splits are by held-out series (matching the L2 discipline), so R² reflects generalization to new series, not in-sample fit.

How to read it
A column that stays dark (low MASE) across most of depth means that horizon step crystallizes early; a column that only lightens near the bottom row means the model needs its full depth to forecast that far ahead. Comparing near-term columns (small horizon step) against far-term columns answers whether a model commits to its short-horizon forecast earlier in depth than its long-horizon one.

Limitations
Same caveats as the whole-horizon skip lens above (miscalibration at early layers is a readout limitation, not proof the forecast isn't linearly present yet — see the tuned lens for that) — colors are not comparable across models, only within one model's own grid.

How to read it
A rising curve (crystallization depth increasing with horizon step) means the model settles its near-term forecast earlier in depth than its far-term one — plausible if later steps need more integrated context. A flat curve means the model commits to the whole horizon at once, regardless of how far out a given step is.

Limitations
A gap in the curve (missing marker) means that horizon step's skip-lens MASE never came within tolerance at any captured layer — read as 'not resolved', not as 'infinitely deep'.

How to read it
1.0 is identical geometry up to rotation; 0 is unrelated. High values are expected and not by themselves impressive: both models process the exact same structured input, and CKA is well known to be inflated by shared input-driven variance alone, independent of any shared computation — that is exactly why L2's stitching-gain-over-input-baseline exists as the corrected comparison. The dotted reference line on the depth-correspondence chart below is an empirical null: the same statistic recomputed after shuffling which series lines up with which. A peak far above that null line means the number is at least tracking genuine per-sample correspondence (not an artifact of comparing any two reasonable encoders) — it does not by itself mean the models learned the same thing.

Limitations
Correlational only; says nothing about mechanism (see L2/L3 for that). Sensitive to how many series/windows are sampled — the CI widens for family-conditioned values with few series. Layer-pair geometry can also look similar simply because both networks are under-trained or too small relative to the input's effective dimensionality — that inflates CKA the same way shared input structure does, and the two are hard to tell apart from this number alone.

How to read it
Compare each pair's peak against ITS OWN null column, never against another pair's peak: the null absorbs how much similarity the shared input alone produces for that particular pair of architectures, and that quantity is not the same for two pairs.

Limitations
Every value is a maximum over the two models' layer grids, so larger models offer more candidates to maximize over and their peaks are biased upward relative to smaller ones. The relative depths use the 'block' axis, so a model with an uncaptured surface legitimately never reaches 1.0.

How to read it
A roughly monotonic line (early-to-early, late-to-late) suggests comparable processing order despite different depths/patch sizes. A flat, uniformly-high line usually means the input's own structure dominates every layer's geometry about equally — see the null-line caveat above.

Limitations
The 'best match' is a max over the other model's layers, which mechanically biases the curve upward versus any single fixed pairing (more candidates to match against) and can look artificially smooth even when the underlying matrix is noisy — always sanity-check against the heatmap above. The x-axis is the 'block' depth axis: 'block' places a layer by its position over the model's whole stack, including any surface this run never captured, so a model whose captured layers stop partway up it legitimately ends short of 1.0 instead of being stretched to fill the axis, as Chronos-Bolt and TimesFM do in this run.

How to read it
Compare bar heights across families, not to some universal threshold — a lower bar means this family's structure is where those models' representations differ most, which is exactly the kind of finding the exemplar case studies exist to make concrete.

Limitations
Families with fewer than `l1.min_family_series` series are dropped entirely (too few series for a stable per-family CKA), so a family's absence here isn't evidence of anything.

How to read it
Spearman ρ near 1 means the two models rank which series-pairs are similar/dissimilar in the same order at that layer pair; near 0 means the rankings are unrelated even though CKA (a different notion of agreement) may have matched these layers.

Limitations
Computed on series-level pooled embeddings only (no window resolution), over a capped sample (`l1.rsa_max_series`) for cost — a rough second check, not a replacement for the CKA statistics above.

How to read it
A high value here means the two models' SPARSE, human-legible feature dictionaries agree on which series look similar — a stronger, more interpretable echo of shared structure than the raw activation-space CKA above, since a feature basis is (attempted to be) meaningful on its own rather than an arbitrary rotation of the hidden space.

Limitations
Requires `sae.persist_features: true` (off by default) and at least two SAE targets spanning both comparison-pair models — most runs will not have this artifact at all, which is why it renders only when found rather than as an empty placeholder.

How to read it
Only the gain ABOVE the input baseline is evidence of shared learned structure, so a value at or below zero means that direction carries no stitching evidence regardless of how high its raw R² is. The two directions of one pair are fit independently and are not expected to match.

Limitations
A point estimate without its CI: the per-direction headings below carry the interval, and a gain whose CI straddles zero is not evidence no matter what this column shows. Nothing here corrects for having looked at C(n,2) x 2 directions.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Red (positive ΔR²) at a pair means the source layer predicts the target layer better than raw input features alone can — genuine evidence of shared structure beyond input. Blue (negative) means the input baseline actually wins; that pair carries no stitching evidence at all. The title line's gain CI is the one number to trust: if its low end is above zero, the best pair's gain is real; if the CI straddles zero, treat the whole heatmap as noise no matter how saturated it looks.

Limitations
Ridge R² rewards *linear* translatability only — a nonlinear correspondence between two layers would show up as zero gain here even if it exists. Direction matters: A→B and B→A are fit and scored independently and are not expected to agree.

How to read it
Clean, well-separated color blobs mean that model's representation carves the benchmark into distinct groups at this layer; a single smeared blob means it doesn't separate much at all there. Hover any point for its series id, true family, and the cluster's approximate label — compare cluster shapes side by side, not exact colors (cluster indices aren't matched across models).

Limitations
The 2D layout is a projection for visualization only — apparent distances between clusters aren't meaningful, only which points share a cluster. Cluster *labels* are approximate (majority family + salient signal stats), so treat them as a reading aid, not ground truth; the partition-overlap statistic below is the quantitative version of what this plot shows visually.

How to read it
A heatmap concentrated on (close to) a diagonal-like permutation means the two partitions correspond well, even if cluster numbers don't literally match; a spread-out heatmap means little correspondence. AMI=1.0 is identical partitions, 0 is what you'd expect from independent random partitions of the same size.

Limitations
AMI corrects for chance agreement from cluster count/size alone, but its bootstrap CI (when shown) resamples label pairs, not the clustering itself — it reflects assignment stability, not k-means initialization variance. `k` is often 'auto' (number of benchmark families), so overlap partly reflects how family-like each model's natural clusters are, not just agreement between the two models.

How to read it
AMI is chance-corrected, so 0 is what independent partitions of the same sizes would give and the values are directly comparable across rows even when the pairs have different cluster counts.

Limitations
Each row is an independent statistic with its own CI; nothing here corrects for the fact that C(n,2) of them were looked at, so read the spread across rows rather than singling out the largest.

How to read it
Whiskers are the metric's own bootstrap CI where one exists. The dotted vertical line is that metric's own reference (a null, a chance level, or the input-feature baseline) -- read a pair's distance from ITS metric's own reference, never compare raw positions across rows. Use the dropdown to highlight one pair across every metric at once.

Limitations
Metrics disagree by design (this run's own measured fact): a pair close on one axis can be far on another. A missing row means that metric's own stage artifact was not present for this run.

How to read it
Brighter cells mean that layer's representation shifted more under that corruption. Reading down a column shows where in depth a given property (say, seasonality) gets encoded; reading across a row shows what a given layer is currently sensitive to. Two models with similar column shapes react to the same properties at similar relative depths, even if their absolute layer counts differ. The depth axis (named per panel above) places each layer within the model's *whole* stack, not just its captured layers, so a model with a real but uncaptured surface legitimately caps out partway up the axis rather than reaching 1.0, as Chronos-Bolt and TimesFM do in this run.

Limitations
This is a magnitude-of-change measure, not a causal one — a layer can shift a lot without that shift affecting the final forecast at all (see the behavioral-sensitivity bars and the activation-patching curve below for the causal follow-through). Different architectures normalize activations differently, so raw magnitudes are not directly comparable across models — only the column *shape* (where the peak is) should be compared.

How to read it
+1 means both models' sensitivity peaks at the same relative depth for that corruption; -1 means they peak at opposite ends; 0 means unrelated depth profiles. The lowest bar is called out in the findings as the most divergent corruption for each pair — the structural property that pair's architectures seem to handle at meaningfully different points in depth.

Limitations
Rank correlation over a coarse 33-point depth grid can be noisy for very shallow models (few layers to interpolate between), and says nothing about whether the property matters to the forecast at all — cross-reference with behavioral sensitivity below.

How to read it
Taller bars mean that corruption matters more to this model's output. A corruption with a tall activation fingerprint but a short bar here is being represented internally without much consequence for the forecast — an interesting mismatch worth checking against the patching curve below, which is the causal version of this same question. Read the printed value, not just bar height, before concluding a corruption 'does nothing.'

Limitations
Scale is in MASE-like units (MAE over each series' own naive scale), so it's comparable across families but reflects each corruption's configured strength as much as the model's intrinsic sensitivity — a fair cross-model comparison, not a fair cross-corruption one unless strengths were tuned to match. How unfair a given cross-corruption comparison is, is measured rather than described here: the breakdown table below gives each corruption's own input footprint and perturbation energy, which is what a bar's height should be read against. A corruption altering well under 1% of timesteps produces a small series-averaged value arithmetically, whatever the model does at the points it touched — the per-window patching heatmap further down is where such localized damage shows whether it is still causally recoverable. Dotted reference lines mark each sampling model's own repeat-run MASE noise floor (see the Overall metrics section above) -- a bar below its model's line is not distinguishable from repeat-run noise.

How to read it
Read input touched (%) and input energy first — they are properties of the corruption, identical for every model, and they set the scale a response should be read against. A corruption altering under 1% of timesteps produces a small series-averaged forecast Δ arithmetically, whatever the model does at the points it touched. Δ in floor units divides the response by that model's own repeat-run variation, so a value near 1 is not distinguishable from the model's own noise; a model with no sampling in its forecast path has a zero floor and is labelled rather than divided by. Rows are ordered by the model-averaged forecast Δ, so the top row is this battery's strongest corruption as measured on these models, not as expected.

Limitations
strength matched is False for every row unless l3.calibrate: input_energy was set, and it is the column that decides whether comparing two rows is fair. Comparing two models within a row is fair either way — they received the same corrupted input. Activation magnitudes are not comparable across architectures (different normalization), so peak Δact/mean Δact should be compared down a model's own column and by peak location across models, not by value. best restoration is present only for the corruptions l3.patching.corruptions selected; every other row says so in the cell rather than leaving it blank, because a corruption nobody patched and a corruption whose patch recovered nothing are different findings.

How to read it
1.0 means patching that layer (on average, one window at a time) fully restores the clean forecast; 0 means no effect; negative means patching that window actively made the corrupted forecast worse (a real and informative outcome, not an error — it means that window's clean state is actively misleading once the rest of the context is still corrupted). A curve that rises with depth suggests the corruption's effect on the output is increasingly concentrated in later layers' local token states; a flat curve near 0 suggests the damage isn't carried locally at all (check the per-window heatmaps below for exactly where it lives instead).

Limitations
Deliberately NOT a whole-context (every position at once) patch: overwriting an entire layer's complete state is mathematically guaranteed to reach exactly 100% restoration at every layer in any purely sequential residual architecture (nothing about the input survives past a fully-overwritten layer), which tests wiring, not depth. Window-local patching avoids that ceiling, but values from different corruptions aren't on a shared physical scale (each is normalized by its own clean-vs-corrupted damage) — compare shapes/crossovers within a corruption, not raw levels across corruptions. Unlike the ΔMASE figures elsewhere in this report, these values are NOT expressed in repeat-run-noise-floor units: restoration is already normalized by each corruption's own damage, and that denominator is not stored in MASE units, so the conversion would need a change to the L3 stage itself rather than to this chart. Read a near-zero restoration as "not localized here", not as "below the noise floor".

How to read it
A bright cell means restoring just that (layer, window) recovered most of the clean forecast — the corrupted property's effect on the output is concentrated there. A hot column at a late window (near forecast start) usually just reflects recency — the model naturally weights recent context more — rather than anything specific to the corruption.

Limitations
Every window is scored against the *same* full-context damage denominator (so cells are additive/comparable within one corruption's grid), but that also means a model with many small-effect windows and one with a single dominant window can show similar curve-level averages above for very different reasons — always check the grid, not just the averaged curve.

How to read it
The patched forecast at each corruption uses the single (layer, window) cell that restored the most on average across the whole sampled batch — named in the table as patch_layer / patch_window — not this series' own best cell, which the table gives separately as own_best_layer / own_best_window. Where the two differ, the plotted patch is not the best available for this series, and recovered_frac is correspondingly lower than own_best_restoration. A negative recovered_frac means patching moved the forecast further from the clean one than the corruption did. The context traces are plotted on the same time axis as the forecasts but before step 0 — the corrupted context is what the model actually read before producing the forecast from corrupted input trace; the clean context is what it read for the other two forecasts. Each panel supports the usual Plotly affordances (drag to zoom into a region, double-click to reset, camera icon to save an image) via the toolbar in its top-right corner — the same detail-on-demand the SAE exemplar panels give through their click-to-enlarge modal, adapted to a multi-panel figure that already draws every corruption at once.

Limitations
These series are a family-stratified sample of the patched batch, not a typical or a worst case: useful for making the aggregate curves concrete and for checking that the patch mechanism does what the aggregate says, not for estimating how often any pattern here occurs. All quantities are scaled by each series' own naive-forecast error, so they are comparable across series of different amplitude.