{"id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","argument":"Eduardo C. Garrido-Merchán’s paper tests how far a large language model can take weight-space programming across games of increasing difficulty. In this approach, every feed-forward network weight is set by hand from an explicit specification, without gradient descent or training data. Garrido-Merchán reports ten networks across a four-rung game ladder, with an independent network using a structurally unrelated game at matched difficulty. Six networks admit complete, non-statistical proofs of correctness against an independent oracle. The chess-endgame networks handle move legality and, in one case, an exact one-ply mate detector, but their success falls from 84.1-100% on one-ply tactics to 0.57-9.13% in full adversarial play. The replicate’s decline is less severe with a shorter required plan and exact rather than heuristic safety checks. For the standard-sized rung, the paper says adding shallow search while holding hand-set evaluation weights fixed leaves a gap that reflects an architectural ceiling, not a lack of engineering effort; it reports the collapse across two structurally unrelated games. The authors release the networks, independent oracles, construction-cost measurements and a cold-prompt protocol for future cross-model comparison. These results concern the tested games and networks; the abstract does not report that comparison.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports performance on the tested games and networks; it does not establish how the results generalize to other tasks or models. The abstract describes the chess-endgame results and the authors’ architectural-ceiling interpretation, but does not provide the detailed methods or evidence needed to assess them here. The cross-model comparison is presented as future work; the abstract does not report its results.","speaker":{"name":"Garrido-Merchán, Eduardo C.","roleAtSource":"Article author"},"source":{"title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.09271","publishedAt":"2026-11-06","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper studies how far a large language model can perform weight-space programming across games of increasing difficulty, without gradient descent or training data in constructing the networks.","excerpt":"Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors built ten networks across a four-rung ladder and used a structurally unrelated game at matched difficulty to help distinguish difficulty-level findings from findings about an individual game.","excerpt":"We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"Six networks admit complete, non-statistical proofs of correctness against an independent oracle.","excerpt":"Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The chess-endgame networks perform much worse in full adversarial play than on one-ply tactics; the replicate’s decline is less severe under a shorter required plan and exact rather than heuristic safety checks.","excerpt":"from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors say adding shallow search while holding hand-set evaluation weights fixed shows the remaining gap is an architectural ceiling, and that the collapse reproduces across two structurally unrelated games.","excerpt":"The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors provide a cold-prompt protocol for a future cross-model comparison.","excerpt":"We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.09271","checkedAt":"2026-11-19T00:34:10.646Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How will the reported pattern compare across models under the cold-prompt protocol?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","summary":"Eduardo C. Garrido-Merchán’s paper tests how far a large language model can take weight-space programming across games of increasing difficulty. In this approach, every feed-forward network weight is set by hand from an explicit specification, without gradient descent or training data. Garrido-Merchán reports ten networks across a four-rung game ladder, with an independent network using a structurally unrelated game at matched difficulty. Six networks admit complete, non-statistical proofs of correctness against an independent oracle. The chess-endgame networks handle move legality and, in one case, an exact one-ply mate detector, but their success falls from 84.1-100% on one-ply tactics to 0.57-9.13% in full adversarial play. The replicate’s decline is less severe with a shorter required plan and exact rather than heuristic safety checks. For the standard-sized rung, the paper says adding shallow search while holding hand-set evaluation weights fixed leaves a gap that reflects an architectural ceiling, not a lack of engineering effort; it reports the collapse across two structurally unrelated games. The authors release the networks, independent oracles, construction-cost measurements and a cold-prompt protocol for future cross-model comparison. These results concern the tested games and networks; the abstract does not report that comparison.","question":"How will the reported pattern compare across models under the cold-prompt protocol?","claims":[{"text":"The paper studies how far a large language model can perform weight-space programming across games of increasing difficulty, without gradient descent or training data in constructing the networks.","excerpt":"Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction.","support":"source-stated"},{"text":"The authors built ten networks across a four-rung ladder and used a structurally unrelated game at matched difficulty to help distinguish difficulty-level findings from findings about an individual game.","excerpt":"We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty","support":"source-stated"},{"text":"Six networks admit complete, non-statistical proofs of correctness against an independent oracle.","excerpt":"Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle.","support":"source-stated"},{"text":"The chess-endgame networks perform much worse in full adversarial play than on one-ply tactics; the replicate’s decline is less severe under a shorter required plan and exact rather than heuristic safety checks.","excerpt":"from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks.","support":"source-stated"},{"text":"The authors say adding shallow search while holding hand-set evaluation weights fixed shows the remaining gap is an architectural ceiling, and that the collapse reproduces across two structurally unrelated games.","excerpt":"The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games.","support":"source-stated"},{"text":"The authors provide a cold-prompt protocol for a future cross-model comparison.","excerpt":"We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison","support":"source-stated"}],"limitations":["The abstract reports performance on the tested games and networks; it does not establish how the results generalize to other tasks or models.","The abstract describes the chess-endgame results and the authors’ architectural-ceiling interpretation, but does not provide the detailed methods or evidence needed to assess them here.","The cross-model comparison is presented as future work; the abstract does not report its results."],"mediaId":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","correctionReason":"Revised the summary to make the paper’s central contrast between exact performance on simpler checks and weaker full adversarial play clearer, while retaining its qualifications and attribution.","id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","publicationId":"470445f4-fb56-47f5-8fd1-ea2da4fc9473","version":2,"sourceUrl":"https://arxiv.org/abs/2611.09271","sourceTitle":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-06","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:45.966Z","createdAt":"2026-11-19T00:03:08.260Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","kind":"article","url":"https://arxiv.org/abs/2611.09271","sourceUrl":"https://arxiv.org/abs/2611.09271","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-19T00:01:59.351Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.09271","checkedAt":"2026-11-19T00:34:10.646Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","confidence":0.62,"provider":"typesafe","model":"jev-1.13.0","inputHash":"29a95a296cb445f5a7208ec935ac52e86958cece8be2283a4cf6cfbcf91e2c05","decidedAt":"2026-11-19T00:02:01.222Z"},"provenance":{"textHash":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","observationId":"5febd2995d1874cd0a285902af08fdf0649f1f04d0b1b20ef84d7412f2a2d1b0","retrievedAt":"2026-11-19T00:01:59.112Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"c5223aca-f352-4ee1-8ecd-07d34253d93c"},"reviewer":{"model":"gpt-6-luna","runId":"4e40d232-e5a5-4df9-866a-ac19fa99916f"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-19T00:03:08.260Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Garrido-Merchán, Eduardo C.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Garrido-Merchán, Eduardo C.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Garrido-Merchán, Eduardo C.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.09271","locator":"head/meta[name=author|citation_author]"},"assertionId":"35cb326700e7f65d535304ff927550dc47ae62c25cbe55504b90c63a256a8b86:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-19T00:34:10.646Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"35cb326700e7f65d535304ff927550dc47ae62c25cbe55504b90c63a256a8b86","sourceUrl":"https://arxiv.org/abs/2611.09271","contentSha256":"41e54e55481fc172695ffcead91f34dfe01502d79b6ea0814eefdcafbbd237d3","textHash":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","observedAt":"2026-11-19T00:34:10.646Z"}}],"participationCoverage":{"total":1,"returned":1,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.09271","snapshotId":"35cb326700e7f65d535304ff927550dc47ae62c25cbe55504b90c63a256a8b86","contentSha256":"41e54e55481fc172695ffcead91f34dfe01502d79b6ea0814eefdcafbbd237d3","textHash":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","observedAt":"2026-11-19T00:34:10.646Z","basis":"publisher-page-metadata","sourceLinks":42,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf/source"}},"canonicalUrl":"https://escapemeridian.com/readings/zero-shot-weight-space-programming-across-a-difficulty-graded-daab0cae","formats":{"html":"/readings/zero-shot-weight-space-programming-across-a-difficulty-graded-daab0cae","markdown":"/readings/zero-shot-weight-space-programming-across-a-difficulty-graded-daab0cae.md","json":"/readings/zero-shot-weight-space-programming-across-a-difficulty-graded-daab0cae.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-zero-shot","name":"Zero-shot","explanation":"Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).","referenceUrl":"https://developers.google.com/machine-learning/glossary#zero-shot-prompting"},{"id":"term-llm","name":"Large language model","explanation":"A language model with many learned numerical settings, used to process and generate language.","referenceUrl":"https://developers.google.com/machine-learning/glossary#large-language-model"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","revision":"470445f4-fb56-47f5-8fd1-ea2da4fc9473","revisionKind":"publication-id","title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","url":"https://escapemeridian.com/readings/zero-shot-weight-space-programming-across-a-difficulty-graded-daab0cae"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.09271","revision":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","revisionKind":"sha256-source-text","title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","url":"https://arxiv.org/abs/2611.09271"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","argument":"Eduardo C. Garrido-Merchán’s paper tests how far a large language model can take weight-space programming across games of increasing difficulty. In this approach, every feed-forward network weight is set by hand from an explicit specification, without gradient descent or training data. Garrido-Merchán reports ten networks across a four-rung game ladder, with an independent network using a structurally unrelated game at matched difficulty. Six networks admit complete, non-statistical proofs of correctness against an independent oracle. The chess-endgame networks handle move legality and, in one case, an exact one-ply mate detector, but their success falls from 84.1-100% on one-ply tactics to 0.57-9.13% in full adversarial play. The replicate’s decline is less severe with a shorter required plan and exact rather than heuristic safety checks. For the standard-sized rung, the paper says adding shallow search while holding hand-set evaluation weights fixed leaves a gap that reflects an architectural ceiling, not a lack of engineering effort; it reports the collapse across two structurally unrelated games. The authors release the networks, independent oracles, construction-cost measurements and a cold-prompt protocol for future cross-model comparison. These results concern the tested games and networks; the abstract does not report that comparison.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports performance on the tested games and networks; it does not establish how the results generalize to other tasks or models. The abstract describes the chess-endgame results and the authors’ architectural-ceiling interpretation, but does not provide the detailed methods or evidence needed to assess them here. The cross-model comparison is presented as future work; the abstract does not report its results.","speaker":{"name":"Garrido-Merchán, Eduardo C.","roleAtSource":"Article author"},"source":{"title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.09271","publishedAt":"2026-11-06","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper studies how far a large language model can perform weight-space programming across games of increasing difficulty, without gradient descent or training data in constructing the networks.","excerpt":"Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors built ten networks across a four-rung ladder and used a structurally unrelated game at matched difficulty to help distinguish difficulty-level findings from findings about an individual game.","excerpt":"We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"Six networks admit complete, non-statistical proofs of correctness against an independent oracle.","excerpt":"Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The chess-endgame networks perform much worse in full adversarial play than on one-ply tactics; the replicate’s decline is less severe under a shorter required plan and exact rather than heuristic safety checks.","excerpt":"from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors say adding shallow search while holding hand-set evaluation weights fixed shows the remaining gap is an architectural ceiling, and that the collapse reproduces across two structurally unrelated games.","excerpt":"The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors provide a cold-prompt protocol for a future cross-model comparison.","excerpt":"We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.09271","checkedAt":"2026-11-19T00:34:10.646Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How will the reported pattern compare across models under the cold-prompt protocol?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","summary":"Eduardo C. Garrido-Merchán’s paper tests how far a large language model can take weight-space programming across games of increasing difficulty. In this approach, every feed-forward network weight is set by hand from an explicit specification, without gradient descent or training data. Garrido-Merchán reports ten networks across a four-rung game ladder, with an independent network using a structurally unrelated game at matched difficulty. Six networks admit complete, non-statistical proofs of correctness against an independent oracle. The chess-endgame networks handle move legality and, in one case, an exact one-ply mate detector, but their success falls from 84.1-100% on one-ply tactics to 0.57-9.13% in full adversarial play. The replicate’s decline is less severe with a shorter required plan and exact rather than heuristic safety checks. For the standard-sized rung, the paper says adding shallow search while holding hand-set evaluation weights fixed leaves a gap that reflects an architectural ceiling, not a lack of engineering effort; it reports the collapse across two structurally unrelated games. The authors release the networks, independent oracles, construction-cost measurements and a cold-prompt protocol for future cross-model comparison. These results concern the tested games and networks; the abstract does not report that comparison.","question":"How will the reported pattern compare across models under the cold-prompt protocol?","claims":[{"text":"The paper studies how far a large language model can perform weight-space programming across games of increasing difficulty, without gradient descent or training data in constructing the networks.","excerpt":"Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction.","support":"source-stated"},{"text":"The authors built ten networks across a four-rung ladder and used a structurally unrelated game at matched difficulty to help distinguish difficulty-level findings from findings about an individual game.","excerpt":"We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty","support":"source-stated"},{"text":"Six networks admit complete, non-statistical proofs of correctness against an independent oracle.","excerpt":"Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle.","support":"source-stated"},{"text":"The chess-endgame networks perform much worse in full adversarial play than on one-ply tactics; the replicate’s decline is less severe under a shorter required plan and exact rather than heuristic safety checks.","excerpt":"from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks.","support":"source-stated"},{"text":"The authors say adding shallow search while holding hand-set evaluation weights fixed shows the remaining gap is an architectural ceiling, and that the collapse reproduces across two structurally unrelated games.","excerpt":"The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games.","support":"source-stated"},{"text":"The authors provide a cold-prompt protocol for a future cross-model comparison.","excerpt":"We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison","support":"source-stated"}],"limitations":["The abstract reports performance on the tested games and networks; it does not establish how the results generalize to other tasks or models.","The abstract describes the chess-endgame results and the authors’ architectural-ceiling interpretation, but does not provide the detailed methods or evidence needed to assess them here.","The cross-model comparison is presented as future work; the abstract does not report its results."],"mediaId":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","correctionReason":"Revised the summary to make the paper’s central contrast between exact performance on simpler checks and weaker full adversarial play clearer, while retaining its qualifications and attribution.","id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","publicationId":"470445f4-fb56-47f5-8fd1-ea2da4fc9473","version":2,"sourceUrl":"https://arxiv.org/abs/2611.09271","sourceTitle":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-06","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:45.966Z","createdAt":"2026-11-19T00:03:08.260Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","kind":"article","url":"https://arxiv.org/abs/2611.09271","sourceUrl":"https://arxiv.org/abs/2611.09271","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-19T00:01:59.351Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.09271","checkedAt":"2026-11-19T00:34:10.646Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","confidence":0.62,"provider":"typesafe","model":"jev-1.13.0","inputHash":"29a95a296cb445f5a7208ec935ac52e86958cece8be2283a4cf6cfbcf91e2c05","decidedAt":"2026-11-19T00:02:01.222Z"},"provenance":{"textHash":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","observationId":"5febd2995d1874cd0a285902af08fdf0649f1f04d0b1b20ef84d7412f2a2d1b0","retrievedAt":"2026-11-19T00:01:59.112Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"c5223aca-f352-4ee1-8ecd-07d34253d93c"},"reviewer":{"model":"gpt-6-luna","runId":"4e40d232-e5a5-4df9-866a-ac19fa99916f"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-19T00:03:08.260Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Garrido-Merchán, Eduardo C.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Garrido-Merchán, Eduardo C.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Garrido-Merchán, Eduardo C.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.09271","locator":"head/meta[name=author|citation_author]"},"assertionId":"35cb326700e7f65d535304ff927550dc47ae62c25cbe55504b90c63a256a8b86:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-19T00:34:10.646Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"35cb326700e7f65d535304ff927550dc47ae62c25cbe55504b90c63a256a8b86","sourceUrl":"https://arxiv.org/abs/2611.09271","contentSha256":"41e54e55481fc172695ffcead91f34dfe01502d79b6ea0814eefdcafbbd237d3","textHash":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","observedAt":"2026-11-19T00:34:10.646Z"}}],"participationCoverage":{"total":1,"returned":1,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.09271","snapshotId":"35cb326700e7f65d535304ff927550dc47ae62c25cbe55504b90c63a256a8b86","contentSha256":"41e54e55481fc172695ffcead91f34dfe01502d79b6ea0814eefdcafbbd237d3","textHash":"dd2602b8ec9ff8a57b383d613a1742851b822cb30b773261c833e8bb7de6268b","observedAt":"2026-11-19T00:34:10.646Z","basis":"publisher-page-metadata","sourceLinks":42,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf/source"}}},{"id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","title":"arXiv cs.AI: Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","argument":"Eduardo C. Garrido-Merchán’s paper asks how far a large language model can take weight-space programming across games of increasing difficulty. In weight-space programming, every network weight is set by hand from an explicit specification, without gradient descent or training data. Garrido-Merchán reports building ten networks across a four-rung ladder, with an independent network for a structurally unrelated game at each matched difficulty level. He reports that six networks admit complete, non-statistical proofs of correctness against an independent oracle. The chess-endgame networks handle move legality and, in one case, an exact one-ply mate detector, but perform much worse in full adversarial play: success falls from 84.1–100% on one-ply tactics to 0.57–9.13%. The replicate’s collapse is less severe when the required plan is shorter and safety checks are exact rather than heuristic. The paper also reports that adding shallow search while holding hand-set evaluation weights fixed does not close the standard-sized game’s gap. It interprets this as an architectural ceiling, and says the collapse also appears across two structurally unrelated games. The paper releases the networks, independent oracles, construction-cost measurements and a cold-prompt protocol for future cross-model comparison.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports performance on the tested games and networks; it does not establish how the results generalize to other tasks or models. The abstract describes the chess-endgame results and the authors’ architectural-ceiling interpretation, but does not provide the detailed methods or evidence needed to assess them here. The cross-model comparison is presented as future work; the abstract does not report its results.","speaker":{"name":"Garrido-Merchán, Eduardo C.","roleAtSource":"Article author"},"source":{"title":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.09271","publishedAt":"2026-11-06","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper studies how far a large language model can perform weight-space programming across games of increasing difficulty, without gradient descent or training data in constructing the networks.","excerpt":"Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors built ten networks across a four-rung ladder and used a structurally unrelated game at matched difficulty to help distinguish difficulty-level findings from findings about an individual game.","excerpt":"We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"Six networks admit complete, non-statistical proofs of correctness against an independent oracle.","excerpt":"Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The chess-endgame networks perform much worse in full adversarial play than on one-ply tactics; the replicate’s decline is less severe under a shorter required plan and exact rather than heuristic safety checks.","excerpt":"from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors say adding shallow search while holding hand-set evaluation weights fixed shows the remaining gap is an architectural ceiling, and that the collapse reproduces across two structurally unrelated games.","excerpt":"The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games.","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"},{"text":"The authors provide a cold-prompt protocol for a future cross-model comparison.","excerpt":"We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison","url":"https://arxiv.org/abs/2611.09271","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.09271","checkedAt":"2026-11-18T14:06:09.175Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How would the reported pattern change in a cross-model comparison using the released cold-prompt protocol?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","summary":"Eduardo C. Garrido-Merchán’s paper asks how far a large language model can take weight-space programming across games of increasing difficulty. In weight-space programming, every network weight is set by hand from an explicit specification, without gradient descent or training data. Garrido-Merchán reports building ten networks across a four-rung ladder, with an independent network for a structurally unrelated game at each matched difficulty level. He reports that six networks admit complete, non-statistical proofs of correctness against an independent oracle. The chess-endgame networks handle move legality and, in one case, an exact one-ply mate detector, but perform much worse in full adversarial play: success falls from 84.1–100% on one-ply tactics to 0.57–9.13%. The replicate’s collapse is less severe when the required plan is shorter and safety checks are exact rather than heuristic. The paper also reports that adding shallow search while holding hand-set evaluation weights fixed does not close the standard-sized game’s gap. It interprets this as an architectural ceiling, and says the collapse also appears across two structurally unrelated games. The paper releases the networks, independent oracles, construction-cost measurements and a cold-prompt protocol for future cross-model comparison.","question":"How would the reported pattern change in a cross-model comparison using the released cold-prompt protocol?","claims":[{"text":"The paper studies how far a large language model can perform weight-space programming across games of increasing difficulty, without gradient descent or training data in constructing the networks.","excerpt":"Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction.","support":"source-stated"},{"text":"The authors built ten networks across a four-rung ladder and used a structurally unrelated game at matched difficulty to help distinguish difficulty-level findings from findings about an individual game.","excerpt":"We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty","support":"source-stated"},{"text":"Six networks admit complete, non-statistical proofs of correctness against an independent oracle.","excerpt":"Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle.","support":"source-stated"},{"text":"The chess-endgame networks perform much worse in full adversarial play than on one-ply tactics; the replicate’s decline is less severe under a shorter required plan and exact rather than heuristic safety checks.","excerpt":"from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks.","support":"source-stated"},{"text":"The authors say adding shallow search while holding hand-set evaluation weights fixed shows the remaining gap is an architectural ceiling, and that the collapse reproduces across two structurally unrelated games.","excerpt":"The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games.","support":"source-stated"},{"text":"The authors provide a cold-prompt protocol for a future cross-model comparison.","excerpt":"We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison","support":"source-stated"}],"limitations":["The abstract reports performance on the tested games and networks; it does not establish how the results generalize to other tasks or models.","The abstract describes the chess-endgame results and the authors’ architectural-ceiling interpretation, but does not provide the detailed methods or evidence needed to assess them here.","The cross-model comparison is presented as future work; the abstract does not report its results."],"mediaId":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","correctionReason":null,"id":"daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf","publicationId":"27d4763b-4cb1-412f-86f8-29db6fe74d2c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.09271","sourceTitle":"Zero-Shot Weight-Space Programming Across a Difficulty-Graded Ladder of Two-Player Games","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-06","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:45.966Z","createdAt":"2026-11-12T21:34:11.867Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","kind":"article","url":"https://arxiv.org/abs/2611.09271","sourceUrl":"https://arxiv.org/abs/2611.09271","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-12T21:32:20.802Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.09271","checkedAt":"2026-11-18T14:06:09.175Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v6","mediaId":"3bdb34705de1aaf0013aca63d9e654d819caea3ccd883eb6eb929cddd498755f","confidence":0.73,"provider":"typesafe","model":"jev-1.13.0","inputHash":"84e5554589e678a31c7ddcd30b8ac4117d1e0eb005a95b82a70d70804fea3683","decidedAt":"2026-11-12T21:32:23.231Z"},"provenance":{"textHash":"e19516962f545c8d6cade2ca138cfc1b27fba5edee63b7f9945f07210efbe602","observationId":"526136206fc8ba28d0b2ecf07e407eb57041573876d0480e69cc48544b76e044","retrievedAt":"2026-11-12T21:32:20.498Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"c1984420-ed79-4a26-83b7-b276dcd8639a"},"reviewer":{"model":"gpt-6-luna","runId":"8de4f436-6503-4b84-81c6-ccd8a95996f7"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-12T21:34:11.867Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Garrido-Merchán, Eduardo C.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Garrido-Merchán, Eduardo C.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Garrido-Merchán, Eduardo C.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.09271","locator":"head/meta[name=author|citation_author]"},"assertionId":"f026078aaa58500121fd6273aa8a12d31f1b94b3bdfecafe2ff24597434d7826:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-15T17:50:09.930Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"f026078aaa58500121fd6273aa8a12d31f1b94b3bdfecafe2ff24597434d7826","sourceUrl":"https://arxiv.org/abs/2611.09271","contentSha256":"158cb76a763ece6008f572023932a6fbb3807362c0f11d330fd949cbf764407b","textHash":"e19516962f545c8d6cade2ca138cfc1b27fba5edee63b7f9945f07210efbe602","observedAt":"2026-11-15T17:50:09.930Z"}}],"participationCoverage":{"total":1,"returned":1,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.09271","snapshotId":"f026078aaa58500121fd6273aa8a12d31f1b94b3bdfecafe2ff24597434d7826","contentSha256":"158cb76a763ece6008f572023932a6fbb3807362c0f11d330fd949cbf764407b","textHash":"e19516962f545c8d6cade2ca138cfc1b27fba5edee63b7f9945f07210efbe602","observedAt":"2026-11-15T17:50:09.930Z","basis":"publisher-page-metadata","sourceLinks":41,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/daab0cae2c7448d39a45f24e432374e7a5f0031a5854895c6ae182b979d7a5cf/source"}}}],"audits":[{"publicationId":"27d4763b-4cb1-412f-86f8-29db6fe74d2c","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The paper and its findings are correctly attributed to the single listed author, Eduardo C. Garrido-Merchán, and results are framed as reported. The claims' use of 'the authors' is a minor pluralization but does not misattribute anything.","excerpt":"Authors:Eduardo C. Garrido-Merchán"},{"dimension":"summary","status":"uncertain","explanation":"The summary is mostly faithful but hard for general readers. 'one-ply', 'independent oracle', 'replicate' and 'cold-prompt protocol' are central and unexplained; tooltips or plain glosses are needed. It never says why this matters, such as what it would mean if an LLM can hand-write provably correct networks. It calls both networks 'chess-endgame networks', but the replicate may be a structurally unrelated game at matched difficulty. Smallest repairs: say 'the chess-endgame rung and its replicate', and restore 'an order of magnitude less severe'.","excerpt":"its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks"},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-11-06 matches the v1 submission date. It is not the 9 Nov v2 revision or the feed date.","excerpt":"Submitted on 6 Nov 2026 (v1), last revised 9 Nov 2026 (this version, v2)"},{"dimension":"support","status":"uncertain","explanation":"Most claims are supported by the abstract excerpts. Framing the architectural ceiling as the paper's interpretation is appropriately cautious. Claim 4 and the summary say 'chess-endgame networks' perform worse. The source attributes the collapse to the chess-endgame rung and treats 'its replicate' separately, which may be a different game. Whether the 84.1-100% and 0.57-9.13% ranges cover both networks is unclear. The question and limitations are reasonable and non-leading.","excerpt":"The chess-endgame rung is exact where correctness is cheap, move legality and, in one network, an exact one-ply mate detector, and collapses where it is expensive"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-12T21:34:11.867Z","model":"anthropic/claude-opus-5.5"},{"publicationId":"470445f4-fb56-47f5-8fd1-ea2da4fc9473","verdict":"uncertain","findings":[{"dimension":"attribution","status":"uncertain","explanation":"The paper is correctly attributed to Eduardo C. Garrido-Merchán, but the source lists him as the sole author while the summary and claims say 'the authors' release and built the networks. The plural is not a wrong person, but it blurs who is speaking. Smallest repair: use 'the author' or 'Garrido-Merchán'.","excerpt":"Authors:Eduardo C. Garrido-Merchán"},{"dimension":"summary","status":"uncertain","explanation":"The reading never makes plain that the large language model itself sets the weights zero-shot, without examples. 'Set by hand' reads as if humans do it. Central terms (zero-shot, feed-forward network, gradient descent, one-ply, oracle, rung, adversarial play, cold-prompt protocol) lack tooltips or plain context. The summary drops 'exact' for legality and the source's 'order of magnitude' less-severe collapse. It also blurs that a second network exists at each rung. Repair: state the LLM's role, restore 'exact' and 'order of magnitude', say 'at each rung', and add term-library tooltips.","excerpt":"We ask how far a large language model can carry this practice, zero-shot, across games of deliberately increasing difficulty."},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-11-06 matches the v1 submission date, not the later v3 revision of 17 Nov or the 18 Nov feed date.","excerpt":"Submitted on 6 Nov 2026 (v1), last revised 17 Nov 2026 (this version, v3)"},{"dimension":"support","status":"pass","explanation":"The claims match the abstract with exact excerpts. The architectural-ceiling conclusion is attributed to the paper. The cross-model comparison is labeled as future work, and the question does not assume its outcome. No benefits or costs are invented.","excerpt":"that the remaining gap is an architectural ceiling rather than a matter of engineering effort"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-19T00:03:08.260Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}