Company docs didn't help our AI agent. Four shortnotes doubled accuracy
Comments on “Company docs didn't help our AI agent. Four shortnotes doubled accuracy”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Contextflo ran a Claude Haiku 5.5 agent on DABstep, the Adyen and Hugging Face payments benchmark, and found that company docs did not raise accuracy while four short rule notes did. On 42 held-out questions, tries right went from 40% (tables only) to 36% (docs) to 76% (docs plus notes), at roughly a quarter of the cost. One broad note broke a previously correct answer type. The results are self-reported, use one model and one dataset, and mix official and rebuilt scoring.
Why it ranked
No. 18 of 140 with a score of 0.67. The panel rated it highest on substantiated (0.91) and technical depth (0.84), lowest on significance (0.53). Placed No. 18, No. 24 by score alone: the fourth AI & machine learning story of the day (−0.020). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Practical, well-documented agent evaluation on a public payments dataset with held-out testing; some overclaiming in headline and scoring caveats, but substantive and actionable for builders of AI agents.
- AI judgeJudge 3gpt-5.6-luna
- A carefully controlled, reproducible benchmark shows that concise operational rules substantially improve an AI agent’s database-question accuracy and cost, while documenting scoring corrections, held-out testing, failure modes, and limits.
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
0.67total under rubric v2.3
- Significance0.53
- Reader interest0.78
- Audience breadth0.63
- Practical relevance0.81
- Substantiated0.91
- Technical depth0.84
- New information0.84
- Counts against: Promotional0.08
- Counts against: Overclaim0.31
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? | 0.60 | 0.35 | 0.63 | 0.53 | 0.28 | +0.24 | +0.127 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.74 | 0.70 | 0.90 | 0.78 | 0.20 | +0.26 | +0.203 |
| Audience breadthHow many people who build technology does it concern? | 0.73 | 0.50 | 0.67 | 0.63 | 0.23 | +0.13 | +0.082 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.86 | 0.80 | 0.78 | 0.81 | 0.08 | +0.22 | +0.179 |
| 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. | 0.99 | 0.80 | 0.94 | 0.91 | 0.19 | +0.05 | +0.046 |
| Technical depthHow technically substantive is it? | 0.99 | 0.75 | 0.78 | 0.84 | 0.24 | +0.05 | +0.042 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.85 | 0.70 | 0.96 | 0.84 | 0.26 | +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. | 0.07 | 0.15 | 0.03 | 0.08 | 0.12 | −0.19 | −0.016 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.36 | 0.35 | 0.22 | 0.31 | 0.14 | −0.11 | −0.034 |
| Sum of contributions | +0.670 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.670 | ||||||
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/2026-11-20/18/"><img src="https://rubric.hrih.si/badge/2026-11-20/18.svg" alt="Ranked #18 on Rubric, 20 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 1 matches source · 7 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 givenBenchmark result
On the held-out test, the agent got 40% of tries right with only the tables and 36% with the company's docs added.
The article gives no link or citation for it.
The article says “40% of tries right without them, 36% with”
The source says No source to compare with.
No source givenBenchmark result
With the docs and four short notes, the agent reached 76% of tries right on the held-out test, at about a quarter of the cost per answer.
The article gives no link or citation for it.
The article says “They took it to 76% right, at about a quarter of the cost per answer”
The source says No source to compare with.
Matches sourceNumber
DABstep is a public benchmark built by Adyen and Hugging Face that describes a fake payments company with realistic data and 450 questions.
The cited source says the same thing.
The article says “DABstep is a public test built by Adyen, a large payments company, and Hugging Face”
The source says “Adyen and Hugging Face built the Data Agent Benchmark for Multi-step Reasoning (DABstep) together. DABstep consists of over 450 data analysis tasks”
Source cited by the article: huggingface.co/blog/dabstep
No source givenCausal conclusion
In exploration, adding note 2 (add up every price that applies) raised accuracy from 31% to 59%, the biggest jump, and runs got about 40% faster.
The article gives no link or citation for it.
The article says “With it, accuracy went from 31% to 59%, the biggest jump, and runs got about 40% faster because the agent stopped going in circles”
The source says No source to compare with.
No source givenNumber
The team's first scoring gave 77% in exploration, but after checking against the official scoreboard marks the headline is 63%.
The article gives no link or citation for it.
The article says “the headline is 63%, not 77%”
The source says No source to compare with.
No source givenCausal conclusion
Note 2 broke a question the agent previously answered correctly on all 3 tries (most expensive way to pay for a €10 GlobalCard payment), which dropped to 0 of 3 after the note.
The article gives no link or citation for it.
The article says “Before the note, the agent got the "most expensive way to pay for a €10 GlobalCard payment" question right on all 3 tries. After it, 0 out of 3”
The source says No source to compare with.
No source givenCausal conclusion
Adding the docs alone did not change accuracy but cut the cost per question by about two thirds.
The article gives no link or citation for it.
The article says “It did cut the cost per question by about two thirds”
The source says No source to compare with.
No source givenGeneralization
The agent's accuracy gain from the notes comes mostly from familiar question types, indicating notes work as rules that generalize to new questions of the same kind.
The article gives no link or citation for it.
The article says “A note is a rule, so it works on the next question of that kind too”
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 20 Nov 2026, 05:47 UTC. It played no part in the ranking.