Edition of Tuesday 17 November 2026

BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science

Comments on “BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science”

Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked

Amazon released BenchProofer, an open-source tool that uses a task's reference fix to generate a formal specification, a verified implementation and a proof, and SWE-Proof, which applies it to all 500 SWE-Bench Verified tasks. In Amazon's own tests, 134 of Claude Opus 4.8's 425 test-passing patches were flagged by specification-guided audits, and a human review of 41 found 21 real defects. Supplying the specification raised verified pass rates, but the results are self-reported and the causal link is unconfirmed. Proofs also rely on fuzz-tested assumptions about existing code.

Read the original at amazon.science. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Why it ranked

No. 15 of 122 with a score of 0.66. The panel rated it highest on new information (0.91) and technical depth (0.88), lowest on audience breadth (0.62). Placed No. 15, No. 27 by score alone: the placement terms moved other stories past it.

AI judgeJudge 2claude-haiku-5-5
Open-source release of formally verified coding benchmark with substantive technical detail and audit results on SWE-Bench Verified; new relative to recent coverage, but mainly of interest to agent and benchmark builders.
AI judgeJudge 3gpt-5.6-luna
A technically detailed open-source benchmark and dataset introduce formal specifications, proofs, audits, and measured effects on SWE-bench results. Strong empirical detail and caveats support the claims, though the announcement framing slightly limits certainty.

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

0.66total under rubric v2.3

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.63
  • Reader interest0.81
  • Audience breadth0.62
  • Practical relevance0.70
  • Substantiated0.81
  • Technical depth0.88
  • New information0.91
  • Counts against: Promotional0.19
  • Counts against: Overclaim0.30

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.670.450.780.630.33+0.24+0.152
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.780.720.930.810.21+0.26+0.211
Audience breadthHow many people who build technology does it concern?0.630.500.720.620.22+0.13+0.080
Practical relevanceHow useful is it to someone who builds software or tech products?0.720.550.820.700.27+0.22+0.154
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.0.970.620.840.810.35+0.05+0.041
Technical depthHow technically substantive is it?0.950.800.900.880.15+0.05+0.044
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.840.900.980.910.14+0.05+0.045
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.0.090.300.180.190.21−0.19−0.036
OverclaimDo the headline or conclusions claim more than the body supports?0.280.400.220.300.18−0.11−0.033
Sum of contributions+0.658
Claim-check penalty0.000
Total, as published0.657

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.

Ranked #15 on Rubric, 17 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-17/15/"><img src="https://rubric.hrih.si/badge/2026-11-17/15.svg" alt="Ranked #15 on Rubric, 17 Nov 2026" width="245" height="28"></a>

Claim check

Claim check: 6 no source given · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.

  1. Couldn’t checkNumber

    Amazon is releasing SWE-Proof, which adds specifications, implementations and proofs to all 500 tasks in the SWE-Bench Verified dataset.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “we are releasing SWE-Proof, which adds these specifications, implementations, and proofs to all 500 tasks in the SWE-Bench Verified dataset”

    The source says

    No passage could be compared.

    Source cited by the article: github.com/amazon-science/BenchProofer

  2. No source givenBenchmark result

    Claude Opus 4.8 passes the standard SWE-Bench Verified tests on 425 of the 500 tasks (85.0%).

    The article gives no link or citation for it.

    The article says

    “Claude Opus 4.8 passes the standard SWE-Bench Verified tests on 425 of the 500 tasks (85.0%”

    The source says

    No source to compare with.

  3. No source givenBenchmark result

    BenchProofer's audit tests flagged 134 of Opus 4.8's 425 test-passing patches as failing while the reference patch passes.

    The article gives no link or citation for it.

    The article says

    “On those tests, 134 of the 425 patches, nearly one in three, fail, while the reference patch passes”

    The source says

    No source to compare with.

  4. No source givenNumber

    Human reviewers examined a random sample of 41 of the 134 flagged Opus patches and found 21 defects, while the other 20 were valid fixes.

    The article gives no link or citation for it.

    The article says

    “human reviewers examined a random sample of 41 of the 134 flagged Opus patches. They identified 21 defects”

    The source says

    No source to compare with.

  5. No source givenBenchmark result

    With BenchProofer's specification supplied, Opus 4.8's verified pass rate rises from 84.6% to 93.6% using Lean, and GPT-5.5's rises from 79.6% to 94.4%.

    The article gives no link or citation for it.

    The article says

    “When it uses BenchProofer’s specification, its pass rate rises to 93.6%. GPT-5.5 rises from 79.6% to 94.4%”

    The source says

    No source to compare with.

  6. No source givenNumber

    Only 61.4% of Opus's generated Lean specifications passed automated review when it produced specifications without a repository patch.

    The article gives no link or citation for it.

    The article says

    “Only 61.4% of its generated Lean specifications passed automated review”

    The source says

    No source to compare with.

  7. No source givenCausal conclusion

    Giving agents formal specifications is what makes them achieve higher verified pass rates.

    The article gives no link or citation for it.

    The article says

    “The same agents achieve more verified passes when given the formal specification”

    The source says

    No source to compare with.

  8. Couldn’t checkCausal conclusion

    The matplotlib-24177 bug arose because the plot's bounds were computed from a simplified set of vertices, cutting off parts of the histogram.

    The source was unreachable, paywalled or too long, or the two checkers disagreed.

    The article says

    “The root cause is that the plot's bounds were computed from a simplified set of vertices, so parts of the histogram were cut off”

    The source says

    No passage could be compared.

    Source cited by the article: github.com/matplotlib/matplotlib/issues/24097

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 2 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 17 Nov 2026, 05:47 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition. How ranking works describes rubric v2.3.

Comments

Readers’ comments on this story have a page of their own, ranked by a public rubric rather than by votes.

Comments on “BenchProofer: Formally verified coding benchmarks, on demand – Amazon Science”

More on Rubric

From the last 7 editions, chosen by shared topic, score and date.

  1. Score 0.69

    Inkling 2 Mini