Understand the failure.
Verify the repair.
Small, reproducible software failures. The broken implementation, the fix that didn’t work, and the one that passed—preserved together.
Explore the cases ↓How results are verified ↗WHAT THE ARCHIVE CONTAINS
100840 executable cases. 20168 are open.
Every case records the implementation that fails, the fix that did not work, and the repair that passed its checks—with recorded outputs and source hashes. This release adds 100840 cases across 20168 failure mechanisms and 254 domains.
The open tier gives you the failure and the unsuccessful fix for one case in every mechanism. The remaining 80672 cases, 5 variants per mechanism, are member-only: the verified repair, its recorded checks, and the full fixture suite are held in the member archive. Read the methodology ↗
A RECORD OF WHAT WENT WRONG
Browse the archive / 100840
Tensor views: diagonal stride sums both source strides · case 01
diagonal stride sums both source strides.
Tensor views: diagonal stride sums both source strides · case 02
diagonal stride sums both source strides.
Tensor views: diagonal stride sums both source strides · case 03
diagonal stride sums both source strides.
Tensor views: diagonal stride sums both source strides · case 04
diagonal stride sums both source strides.
Tensor views: diagonal stride sums both source strides · case 05
diagonal stride sums both source strides.
Tensor views: broadcast aligns dimensions from the right · case 01
broadcast aligns dimensions from the right.
Tensor views: broadcast aligns dimensions from the right · case 02
broadcast aligns dimensions from the right.
Tensor views: broadcast aligns dimensions from the right · case 03
broadcast aligns dimensions from the right.
Tensor views: broadcast aligns dimensions from the right · case 04
broadcast aligns dimensions from the right.
Tensor views: broadcast aligns dimensions from the right · case 05
broadcast aligns dimensions from the right.
Tensor views: storage span includes stride gaps · case 01
storage span includes stride gaps.
Tensor views: storage span includes stride gaps · case 02
storage span includes stride gaps.
Tensor views: storage span includes stride gaps · case 03
storage span includes stride gaps.
Tensor views: storage span includes stride gaps · case 04
storage span includes stride gaps.
Tensor views: storage span includes stride gaps · case 05
storage span includes stride gaps.
Tensor views: contiguity ignores strides on unit axes · case 01
contiguity ignores strides on unit axes.
Tensor views: contiguity ignores strides on unit axes · case 02
contiguity ignores strides on unit axes.
Tensor views: contiguity ignores strides on unit axes · case 03
contiguity ignores strides on unit axes.
Tensor views: contiguity ignores strides on unit axes · case 04
contiguity ignores strides on unit axes.
Tensor views: contiguity ignores strides on unit axes · case 05
contiguity ignores strides on unit axes.
Tensor views: logical index bounds are checked per axis · case 01
logical index bounds are checked per axis.
Tensor views: logical index bounds are checked per axis · case 02
logical index bounds are checked per axis.
Tensor views: logical index bounds are checked per axis · case 03
logical index bounds are checked per axis.
Tensor views: logical index bounds are checked per axis · case 04
logical index bounds are checked per axis.
Tensor views: logical index bounds are checked per axis · case 05
logical index bounds are checked per axis.
Tensor views: stepped slice multiplies stride · case 01
stepped slice multiplies stride.
Tensor views: stepped slice multiplies stride · case 02
stepped slice multiplies stride.
Tensor views: stepped slice multiplies stride · case 03
stepped slice multiplies stride.
Tensor views: stepped slice multiplies stride · case 04
stepped slice multiplies stride.
Tensor views: stepped slice multiplies stride · case 05
stepped slice multiplies stride.
Tensor views: broadcast strides align with leading singleton axes · case 01
broadcast strides align with leading singleton axes.
Tensor views: broadcast strides align with leading singleton axes · case 02
broadcast strides align with leading singleton axes.
Tensor views: broadcast strides align with leading singleton axes · case 03
broadcast strides align with leading singleton axes.
Tensor views: broadcast strides align with leading singleton axes · case 04
broadcast strides align with leading singleton axes.
Tensor views: broadcast strides align with leading singleton axes · case 05
broadcast strides align with leading singleton axes.
Tensor views: row-major linear index uses dimension suffix products · case 01
row-major linear index uses dimension suffix products.
Tensor views: row-major linear index uses dimension suffix products · case 02
row-major linear index uses dimension suffix products.
Tensor views: row-major linear index uses dimension suffix products · case 03
row-major linear index uses dimension suffix products.
Tensor views: row-major linear index uses dimension suffix products · case 04
row-major linear index uses dimension suffix products.
Tensor views: row-major linear index uses dimension suffix products · case 05
row-major linear index uses dimension suffix products.
Unequal example batches receive equal mean-gradient weight · case 01
Unequal example batches receive equal mean-gradient weight.
Unequal example batches receive equal mean-gradient weight · case 02
Unequal example batches receive equal mean-gradient weight.
Unequal example batches receive equal mean-gradient weight · case 03
Unequal example batches receive equal mean-gradient weight.
Unequal example batches receive equal mean-gradient weight · case 04
Unequal example batches receive equal mean-gradient weight.
Unequal example batches receive equal mean-gradient weight · case 05
Unequal example batches receive equal mean-gradient weight.
Unequal valid-token batches receive equal token-mean weight · case 01
Unequal valid-token batches receive equal token-mean weight.
Unequal valid-token batches receive equal token-mean weight · case 02
Unequal valid-token batches receive equal token-mean weight.
Unequal valid-token batches receive equal token-mean weight · case 03
Unequal valid-token batches receive equal token-mean weight.
Unequal valid-token batches receive equal token-mean weight · case 04
Unequal valid-token batches receive equal token-mean weight.
Unequal valid-token batches receive equal token-mean weight · case 05
Unequal valid-token batches receive equal token-mean weight.
Sample weights are normalized by raw item count · case 01
Sample weights are normalized by raw item count.
Sample weights are normalized by raw item count · case 02
Sample weights are normalized by raw item count.
Sample weights are normalized by raw item count · case 03
Sample weights are normalized by raw item count.
Sample weights are normalized by raw item count · case 04
Sample weights are normalized by raw item count.
Sample weights are normalized by raw item count · case 05
Sample weights are normalized by raw item count.
An empty microbatch dilutes the equal-microbatch mean · case 01
An empty microbatch dilutes the equal-microbatch mean.
An empty microbatch dilutes the equal-microbatch mean · case 02
An empty microbatch dilutes the equal-microbatch mean.
An empty microbatch dilutes the equal-microbatch mean · case 03
An empty microbatch dilutes the equal-microbatch mean.
An empty microbatch dilutes the equal-microbatch mean · case 04
An empty microbatch dilutes the equal-microbatch mean.
An empty microbatch dilutes the equal-microbatch mean · case 05
An empty microbatch dilutes the equal-microbatch mean.
A flushed partial window is divided by configured capacity · case 01
A flushed partial window is divided by configured capacity.
A flushed partial window is divided by configured capacity · case 02
A flushed partial window is divided by configured capacity.
A flushed partial window is divided by configured capacity · case 03
A flushed partial window is divided by configured capacity.
A flushed partial window is divided by configured capacity · case 04
A flushed partial window is divided by configured capacity.
A flushed partial window is divided by configured capacity · case 05
A flushed partial window is divided by configured capacity.
Mean-reduced microbatch losses are summed · case 01
Mean-reduced microbatch losses are summed.
Mean-reduced microbatch losses are summed · case 02
Mean-reduced microbatch losses are summed.
Mean-reduced microbatch losses are summed · case 03
Mean-reduced microbatch losses are summed.
Mean-reduced microbatch losses are summed · case 04
Mean-reduced microbatch losses are summed.
Mean-reduced microbatch losses are summed · case 05
Mean-reduced microbatch losses are summed.
Long sequences dominate a sequence-mean objective · case 01
Long sequences dominate a sequence-mean objective.
Long sequences dominate a sequence-mean objective · case 02
Long sequences dominate a sequence-mean objective.
Long sequences dominate a sequence-mean objective · case 03
Long sequences dominate a sequence-mean objective.
Long sequences dominate a sequence-mean objective · case 04
Long sequences dominate a sequence-mean objective.
Long sequences dominate a sequence-mean objective · case 05
Long sequences dominate a sequence-mean objective.
Sequence means are averaged for a token-mean objective · case 01
Sequence means are averaged for a token-mean objective.
Sequence means are averaged for a token-mean objective · case 02
Sequence means are averaged for a token-mean objective.
Sequence means are averaged for a token-mean objective · case 03
Sequence means are averaged for a token-mean objective.
Sequence means are averaged for a token-mean objective · case 04
Sequence means are averaged for a token-mean objective.
Sequence means are averaged for a token-mean objective · case 05
Sequence means are averaged for a token-mean objective.
Vector coordinates use the number of microbatches as denominator · case 01
Vector coordinates use the number of microbatches as denominator.
Vector coordinates use the number of microbatches as denominator · case 02
Vector coordinates use the number of microbatches as denominator.
Vector coordinates use the number of microbatches as denominator · case 03
Vector coordinates use the number of microbatches as denominator.
Vector coordinates use the number of microbatches as denominator · case 04
Vector coordinates use the number of microbatches as denominator.
Vector coordinates use the number of microbatches as denominator · case 05
Vector coordinates use the number of microbatches as denominator.
A parameter absent from a microbatch is counted as a zero update · case 01
A parameter absent from a microbatch is counted as a zero update.
A parameter absent from a microbatch is counted as a zero update · case 02
A parameter absent from a microbatch is counted as a zero update.
A parameter absent from a microbatch is counted as a zero update · case 03
A parameter absent from a microbatch is counted as a zero update.
A parameter absent from a microbatch is counted as a zero update · case 04
A parameter absent from a microbatch is counted as a zero update.
A parameter absent from a microbatch is counted as a zero update · case 05
A parameter absent from a microbatch is counted as a zero update.
Scaled gradients are averaged before unscaling · case 01
Scaled gradients are averaged before unscaling.
Scaled gradients are averaged before unscaling · case 02
Scaled gradients are averaged before unscaling.
Scaled gradients are averaged before unscaling · case 03
Scaled gradients are averaged before unscaling.
Scaled gradients are averaged before unscaling · case 04
Scaled gradients are averaged before unscaling.
Scaled gradients are averaged before unscaling · case 05
Scaled gradients are averaged before unscaling.
A changing loss scale is applied after gradient summation · case 01
A changing loss scale is applied after gradient summation.
A changing loss scale is applied after gradient summation · case 02
A changing loss scale is applied after gradient summation.
A changing loss scale is applied after gradient summation · case 03
A changing loss scale is applied after gradient summation.
A changing loss scale is applied after gradient summation · case 04
A changing loss scale is applied after gradient summation.
A changing loss scale is applied after gradient summation · case 05
A changing loss scale is applied after gradient summation.
INSPECTABLE BY DESIGN
Every result has a runnable source.
Runnable implementations with recorded outputs, source hashes, and explicit contracts. Related variants share a failure mechanism and belong together in evaluation splits.
Read the methodology ↗