Edition of Monday 16 November 2026

I let LLMs play a 1988 detective gamebook 646 times

Comments on “I let LLMs play a 1988 detective gamebook 646 times”

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

A developer ran 95 runs and 646 games of LLMs playing a 1988 detective gamebook using the GitHub Copilot SDK. Every first attempt lost, but capable models learned from retries, with more attempts outperforming a bigger model.

Read the original at blog.mastykarz.nl. 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. 24 of 60 with a score of 0.66. The panel rated it highest on new information (0.90) and substantiated (0.83), lowest on significance (0.52). The judges disagreed most on audience breadth (spread 0.43), so the story is marked contested. Placed No. 24, No. 16 by score alone: the seventh AI & machine learning story of the day (−0.080). The placement terms move stories; they are not part of the score.

AI judgeJudge 2claude-haiku-5-5
A blinded, well-substantiated experiment on LLM agents learning from retries, with clear practical lessons for agent builders, though the subject is a toy gamebook and the findings are narrow.
AI judgeJudge 3gpt-5.6-luna
A well-supported original experiment on agent learning, retries, feedback loops, state visibility, and misleading root-cause explanations, with concrete methodology and actionable implications, though generalization beyond gamebooks remains partly inferential.

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 v2.3Contestedjudges differ by 0.43 on audience breadth

Judges’ mean answer per question, on a scale from 0 to 1. Questions marked minus count against the score.
  • Significance0.52
  • Reader interest0.79
  • Audience breadth0.65
  • Practical relevance0.73
  • Substantiated0.83
  • Technical depth0.70
  • New information0.90
  • 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).

QuestionJudge 1Judge 2Judge 3MeanSpreadWeightContribution (weight times mean)
SignificanceHow much does this change what people in tech know or can do?0.640.300.620.520.34+0.24+0.124
Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on?0.790.680.900.790.22+0.26+0.205
Audience breadthHow many people who build technology does it concern?0.830.400.720.650.43 , above the contested threshold+0.13+0.084
Practical relevanceHow useful is it to someone who builds software or tech products?0.860.550.780.730.31+0.22+0.160
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.0.940.720.830.830.22+0.05+0.042
Technical depthHow technically substantive is it?0.890.550.670.700.34+0.05+0.035
New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)?0.820.900.980.900.16+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.0.040.080.030.050.05−0.19−0.010
OverclaimDo the headline or conclusions claim more than the body supports?0.260.300.180.250.12−0.11−0.027
Sum of contributions+0.660
Claim-check penalty0.000
Total, as published0.659

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 #24 on Rubric, 16 Nov 2026
<a href="https://rubric.hrih.si/editions/2026-11-16/24/"><img src="https://rubric.hrih.si/badge/2026-11-16/24.svg" alt="Ranked #24 on Rubric, 16 Nov 2026" width="245" 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.

  1. No source givenNumber

    All 95 first attempts across every model and setting lost.

    The article gives no link or citation for it.

    The article says

    “Across the whole project, models played 95 first attempts, and all 95 lost”

    The source says

    No source to compare with.

  2. No source givenBenchmark result

    GPT‑6 Astra at maximum effort won 7 of 15 runs with up to 10 attempts, and all 6 runs with up to 15 attempts.

    The article gives no link or citation for it.

    The article says

    “GPT‑6 Astra at maximum effort won 7 of 15 runs with 10 attempts, and 6 of 6 with 15”

    The source says

    No source to compare with.

  3. No source givenNumber

    Cost per solve for Astra dropped from 789 premium requests to 396 when attempts rose from 10 to 15.

    The article gives no link or citation for it.

    The article says

    “Cost per solve dropped from 789 premium requests to 396”

    The source says

    No source to compare with.

  4. No source givenNumber

    Of 97 times a capable model reached the final check, it was missing something 80 times, and in 74 of those the missing piece was marker 60.

    The article gives no link or citation for it.

    The article says

    “Of the 97 times a capable model reached that check, it was missing something 80 times. In 74 of those, the missing piece was marker 60”

    The source says

    No source to compare with.

  5. No source givenBenchmark result

    On the sequel The Water Spider, Astra won all six runs, on attempts 5, 5, 5, 6, 6 and 7.

    The article gives no link or citation for it.

    The article says

    “Astra won all six runs, on attempts 5, 5, 5, 6, 6 and 7”

    The source says

    No source to compare with.

  6. No source givenNumber

    The author's search-based solver found a winning route through The Money Spider in 0.1 seconds but gave up on The Water Spider after 10 million states.

    The article gives no link or citation for it.

    The article says

    “found a winning route through The Money Spider in 0.1 seconds. On The Water Spider, it gave up after 10 million states without finding one”

    The source says

    No source to compare with.

  7. No source givenNumber

    None of the six runs that used self-written notes instead of transcripts won.

    The article gives no link or citation for it.

    The article says

    “None of the six runs won”

    The source says

    No source to compare with.

  8. No source givenCausal conclusion

    Failures that point back quickly to their cause are learned from, while failures far from their cause are not, explaining the difference between the two books.

    The article gives no link or citation for it.

    The article says

    “the difference comes down to whether the failure points back at its cause”

    The source says

    No source to compare with.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 4 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 16 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 “I let LLMs play a 1988 detective gamebook 646 times”

More on Rubric

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