Edition of Sunday 15 November 2026

Building Strong Safety Evaluations

Comments on “Building Strong Safety Evaluations”

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

Anthropic published guidance on building AI safety evals, borrowing preregistration and sample-size practices from psychology's replication reforms. Key points: score appropriate responses rather than refusal rates, pair harmful with benign prompts, use multi-turn simulations, validate LLM graders against human labels, and require at least 100 multi-turn or 1,000 single-turn items. Anthropic reports single-turn harmless rates of 94–98% across recent models, so those tests mostly serve as regression checks. Several figures come from Anthropic's own evals and have no cited source.

Read the original at anthropic.com. 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. 12 of 62 with a score of 0.69. The panel rated it highest on new information (0.84) and practical relevance (0.80), lowest on significance (0.63). Placed No. 12, No. 8 by score alone: the sixth AI & machine learning story of the day (−0.060). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
Detailed, practitioner-oriented guidance on building and validating AI safety evals, with concrete methods (grader calibration, multi-turn simulation, preregistration). Strongly AI-focused; significance moderate for a niche community.
AI judgeJudge 3gpt-5.6-luna
Substantive, actionable guidance for designing and validating AI safety evaluations, with methodological detail, examples, and uncertainty considerations. It is somewhat grounded in the publisher’s own practices but offers broadly useful technical recommendations.

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

0.69total 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 interest0.77
  • Audience breadth0.71
  • Practical relevance0.80
  • Substantiated0.70
  • Technical depth0.79
  • New information0.84
  • Counts against: Promotional0.13
  • Counts against: Overclaim0.15

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?0.720.450.720.630.28+0.24+0.152
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.700.700.900.770.20+0.26+0.199
Audience breadthHow many people who build technology does it concern?0.740.600.780.710.18+0.13+0.092
Practical relevanceHow useful is it to someone who builds software or tech products?0.820.750.820.800.07+0.22+0.175
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.0.650.700.760.700.11+0.05+0.035
Technical depthHow technically substantive is it?0.890.700.780.790.19+0.05+0.040
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.890.800.820.840.09+0.05+0.042
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.0.070.150.180.130.11−0.19−0.025
OverclaimDo the headline or conclusions claim more than the body supports?0.140.200.120.150.08−0.11−0.017
Sum of contributions+0.693
Claim-check penalty0.000
Total, as published0.693

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 #12 on Rubric, 15 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-15/12/"><img src="https://rubric.hrih.si/badge/2026-11-15/12.svg" alt="Ranked #12 on Rubric, 15 Nov 2026" width="245" height="28"></a>

Claim check

Claim check: 1 matches source · 5 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. No source givenNumber

    Anthropic's current aggregate harmless response rates for its single-turn evaluations on the API are between 94% and 98% for every recent model.

    The article gives no link or citation for it.

    The article says

    “On our API, the current aggregate harmless response rates for our single-turn evaluations are between 94% and 98% for every recent model”

    The source says

    No source to compare with.

  2. No source givenNumber

    Anthropic's single-turn safety suites run across 16 policy areas in seven languages and across multiple product surfaces.

    The article gives no link or citation for it.

    The article says

    “Our single-turn safety suites run across 16 policy areas in seven languages and across multiple product surfaces”

    The source says

    No source to compare with.

  3. No source givenBenchmark result

    On a multi-turn romance scams evaluation on claude.ai, Claude Opus 5 and Opus 5.5 had appropriate response rates of 98% versus 94%, a difference that was not statistically significant.

    The article gives no link or citation for it.

    The article says

    “Claude Opus 5 and Opus 5.5 differed by multiple percentage points (98% versus 94%). However, the difference between the two models wasn’t statistically significant”

    The source says

    No source to compare with.

  4. No source givenNumber

    In testing Claude Fable 5.1, the difference in appropriate response rate between the API and claude.ai configurations on multi-turn safety evals was as high as 34 percentage points.

    The article gives no link or citation for it.

    The article says

    “the difference in the appropriate response rate between these two configurations on our multi-turn safety evals was as high as 34 percentage points”

    The source says

    No source to compare with.

  5. No source givenBenchmark result

    After adding a system prompt instruction and other mitigations, the appropriate response rate on the suicide and self-harm eval was 60% on the API versus 94% on claude.ai with the new system prompt.

    The article gives no link or citation for it.

    The article says

    “60% on the API vs 94% on claude.ai with the new system prompt”

    The source says

    No source to compare with.

  6. Couldn’t checkBenchmark result

    An audit of a leading coding benchmark found that a substantial fraction of its tasks were underspecified, lacked necessary context, or were graded by tests that rejected correct solutions.

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

    The article says

    “one audit of a leading coding benchmark found that a substantial fraction of its tasks were underspecified, lacked necessary context, or were graded by tests that rejected correct solutions”

    The source says

    No passage could be compared.

    Source cited by the article: openai.com/index/introducing-swe-bench-verified

  7. Couldn’t checkGeneralization

    Large-scale replication projects in psychology found that many published findings didn't hold up.

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

    The article says

    “large-scale replication projects found that many published findings didn’t hold up”

    The source says

    No passage could be compared.

    Source cited by the article: doi.org/10.1126/science.aac4716

  8. Matches sourceGeneralization

    Model graders have been shown to recognize and favor their own outputs.

    The cited source says the same thing.

    The article says

    “model graders have been shown to recognize and favor their own outputs”

    The source says

    “We discover that, out of the box, LLMs such as GPT-4 and Llama 2 have non-trivial accuracy at distinguishing themselves from other LLMs and humans”

    Source cited by the article: arxiv.org/abs/2404.13076

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 3 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 15 Nov 2026, 20:46 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 “Building Strong Safety Evaluations”

More on Rubric

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