When Rounding Breaks Machine Learning
Comments on “When Rounding Breaks Machine Learning”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Robert Joseph George checked all 154 TorchLean operators (162 with variants) against exact results using FloatLib, a Lean library for modelling floating-point arithmetic, and compared them with PyTorch, JAX, TensorFlow, NumPy, CuPy, ONNX Runtime and MLX across roughly 780 million outputs. He reports rounding bugs in several widely used libraries and in TorchLean's own float32 softmax, plus rewrites that fix them. A formula proved correct over the reals can still fail in floating point. The figures are the author's own and come with no independent source.
Why it ranked
No. 3 of 142 with a score of 0.73. The panel rated it highest on technical depth (0.95) and substantiated (0.93), lowest on significance (0.69).
- AI judgeJudge 2claude-haiku-5-5
- Rigorous, technically deep study of floating-point behavior across ML operators and libraries with bounded proofs and cross-library measurements; substantive, though niche in audience and largely research-oriented.
- AI judgeJudge 3gpt-5.6-luna
- A highly technical, machine-checked study of floating-point failures across neural-network operators and major libraries, with reproducible methodology, extensive measurements, and corrective rewrites. It is broadly useful to ML and numerical-software practitioners.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.73total under rubric v2.3
- Significance0.69
- Reader interest0.80
- Audience breadth0.73
- Practical relevance0.73
- Substantiated0.93
- Technical depth0.95
- New information0.92
- Counts against: Promotional0.05
- Counts against: Overclaim0.25
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.73 | 0.55 | 0.78 | 0.69 | 0.23 | +0.24 | +0.165 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.69 | 0.80 | 0.91 | 0.80 | 0.22 | +0.26 | +0.208 |
| Audience breadthHow many people who build technology does it concern? | 0.78 | 0.60 | 0.82 | 0.73 | 0.22 | +0.13 | +0.095 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.72 | 0.70 | 0.78 | 0.73 | 0.08 | +0.22 | +0.161 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.98 | 0.85 | 0.95 | 0.93 | 0.13 | +0.05 | +0.046 |
| Technical depthHow technically substantive is it? | 0.98 | 0.92 | 0.96 | 0.95 | 0.06 | +0.05 | +0.048 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.90 | 0.90 | 0.97 | 0.92 | 0.07 | +0.05 | +0.046 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.06 | 0.05 | 0.04 | 0.05 | 0.02 | −0.19 | −0.010 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.43 | 0.20 | 0.12 | 0.25 | 0.31 | −0.11 | −0.028 |
| Sum of contributions | +0.732 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.732 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-18/3/"><img src="https://rubric.hrih.si/badge/2026-11-18/3.svg" alt="Ranked #3 on Rubric, 18 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenCausal conclusion
TorchLean's own softmax implementation had the same bug as the one Jon Barron described: in float32 the target gradient became zero although the true answer was nonzero and representable.
The article gives no link or citation for it.
The article says “in float32, the target gradient became zero even though the true answer was nonzero and representable”
The source says No source to compare with.
No source givenNumber
The authors found bugs in widely used code, in PyTorch, JAX, TensorFlow Probability, diffusers, PyTorch Geometric and xLSTM, several of them never reported before, and in TorchLean itself.
The article gives no link or citation for it.
The article says “That turned up bugs in widely used code, in PyTorch, JAX, TensorFlow Probability, diffusers, PyTorch Geometric and xLSTM, several of them never reported before”
The source says No source to compare with.
No source givenNumber
The study examined all 154 TorchLean operators, or 162 entries when variants are counted separately.
The article gives no link or citation for it.
The article says “We examined all 154 TorchLean operators, or 162 entries when variants are counted separately”
The source says No source to compare with.
No source givenNumber
Executable Lean rules exist for 159 of the entries; erf, erf-based GELU, and the custom extension are outside this executable sweep.
The article gives no link or citation for it.
The article says “We have executable Lean rules for 159 entries; erf, erf-based GELU, and the custom extension for user-defined programs are outside this executable sweep”
The source says No source to compare with.
No source givenNumber
The main comparison spans roughly 780 million outputs, and selected formulas are later taken through 526 different formats using FloatLib.
The article gives no link or citation for it.
The article says “The main comparison spans several commonly used precisions and roughly 780 million outputs. Later, we take selected original and rewritten formulas through 526 different formats using FloatLib”
The source says No source to compare with.
No source givenBenchmark result
In the Lean sweep, 57,547 outputs (96.8%) matched the exact result rounded once, and another 1,329 matched TorchLean's specified sequence of rounded operations.
The article gives no link or citation for it.
The article says “In the Lean sweep, 57,547 outputs (96.8%) matched the exact result rounded once. Another 1,329 matched TorchLean's specified sequence of rounded operations”
The source says No source to compare with.
No source givenNumber
In float32, adding 2^-25 to 1 rounds back to 1, so subtracting 1 afterwards gives zero even though 2^-25 is exactly representable in float32.
The article gives no link or citation for it.
The article says “The addition rounds back to 1, because the exact sum is closer to 1 than to the next float32 value”
The source says No source to compare with.
No source givenFirst, only or fastest
Prieto and collaborators had already identified Softmax Collapse.
The article gives no link or citation for it.
The article says “Prieto and collaborators had already identified Softmax Collapse”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 18 Nov 2026, 05:47 UTC. It played no part in the ranking.