Learning in a World of Notes
Comments on “Learning in a World of Notes”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Engram reports that reinforcement learning on top of fixed, high-quality notes raised agent accuracy by 7 to 20 points across three workspaces. On Harvey C&H, a trained Qwen 27B model matched GPT-6 Sol with notes at 78% for $0.06 per query versus $0.17, and training cost roughly 5-30% of building the notes. A model trained once still beat the base model after 23M tokens of note edits. These are the company's own results; on Hex it did not reach frontier accuracy, and the robustness test covers one workspace.
Why it ranked
No. 18 of 51 with a score of 0.62. The panel rated it highest on new information (0.87) and reader interest (0.77), lowest on significance (0.61). The judges disagreed most on substantiated (spread 0.42), so the story is marked contested. Placed No. 18, No. 16 by score alone: the fifth AI & machine learning story of the day (−0.040). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Vendor-style research post on RL training with text memory for agents. Substantive experiments and cost analysis, but largely self-reported with some promotional framing and unverifiable synthetic benchmarks.
- AI judgeJudge 3gpt-5.6-luna
- A detailed empirical study of RL-trained agents using text memory, with benchmarks, ablations, cost analysis, changing-notes experiments, and behavioral traces. Its synthetic settings and optimistic break-even assumptions limit generality, but the findings are technically useful and substantially evidenced.
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
0.62total under rubric v2.3Contestedjudges differ by 0.42 on substantiated
- Significance0.61
- Reader interest0.77
- Audience breadth0.65
- Practical relevance0.71
- Substantiated0.73
- Technical depth0.74
- New information0.87
- Counts against: Promotional0.21
- Counts against: Overclaim0.39
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? | 0.67 | 0.45 | 0.72 | 0.61 | 0.27 | +0.24 | +0.147 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.69 | 0.70 | 0.91 | 0.77 | 0.22 | +0.26 | +0.199 |
| Audience breadthHow many people who build technology does it concern? | 0.67 | 0.55 | 0.72 | 0.65 | 0.17 | +0.13 | +0.084 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.71 | 0.60 | 0.82 | 0.71 | 0.22 | +0.22 | +0.157 |
| 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. | 0.92 | 0.50 | 0.78 | 0.73 | 0.42 , above the contested threshold | +0.05 | +0.037 |
| Technical depthHow technically substantive is it? | 0.83 | 0.60 | 0.78 | 0.74 | 0.23 | +0.05 | +0.037 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.87 | 0.80 | 0.93 | 0.87 | 0.13 | +0.05 | +0.043 |
| 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. | 0.11 | 0.35 | 0.18 | 0.21 | 0.24 | −0.19 | −0.041 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.37 | 0.45 | 0.34 | 0.39 | 0.11 | −0.11 | −0.043 |
| Sum of contributions | +0.621 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.621 | ||||||
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/2026-11-08/18/"><img src="https://rubric.hrih.si/badge/2026-11-08/18.svg" alt="Ranked #18 on Rubric, 8 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 givenNumber
Training models to use their notes with reinforcement learning yields accuracy improvements of 7 to 20 percentage points across Engram's benchmarks.
The article gives no link or citation for it.
The article says “accuracy improvements of 7 to 20 percentage points across our benchmarks”
The source says No source to compare with.
No source givenNumber
Training costs roughly 5-30% as much as building a good text memory system.
The article gives no link or citation for it.
The article says “training costs roughly 5-30% as much as building a good text memory system”
The source says No source to compare with.
No source givenBenchmark result
On Harvey C&H, the trained model matches GPT-6 Sol (medium effort) with notes at 78% accuracy but costs $0.06 per query versus $0.17.
The article gives no link or citation for it.
The article says “the trained model matches GPT-6 Sol (medium effort) with notes at 78% accuracy, but costs $0.06 per query against $0.17”
The source says No source to compare with.
No source givenNumber
The trained models break even against the cheapest comparable-accuracy frontier model within days in every workspace, from under a day on Engram Internal to about 9 days on Harvey C&H, for 20 users sending 20 queries a day.
The article gives no link or citation for it.
The article says “it breaks even within days in every workspace—from under a day on Engram Internal to about 9 days on Harvey C&H”
The source says No source to compare with.
No source givenBenchmark result
On Hex DataBench, training lifts accuracy by 15 points over notes alone, the largest gain among the workspaces.
The article gives no link or citation for it.
The article says “The gain is largest on Hex, where notes alone help only modestly and training lifts accuracy by 15 points”
The source says No source to compare with.
No source givenNumber
A model trained once on the initial notes still beats an untrained model after 23M tokens of note edits on Harvey C&H.
The article gives no link or citation for it.
The article says “A model trained once on an initial version of its notes still beats an untrained one after 23M tokens of text edits”
The source says No source to compare with.
No source givenGeneralization
The robustness of the reused trained model shows it learns general strategies for using notes rather than overfitting to their contents, so it needs no retraining after every edit.
The article gives no link or citation for it.
The article says “The model learns general strategies for using notes rather than overfitting to their contents, so it doesn’t need to be retrained after every edit”
The source says No source to compare with.
No source givenBenchmark result
On Hex, RL training with notes does not fully match frontier model performance.
The article gives no link or citation for it.
The article says “On Hex, RL training with notes isn't fully able to match frontier performance”
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 8 Nov 2026, 05:46 UTC. It played no part in the ranking.