rldebruijn/potion-preprint-rows

Potion preprint: per-row score release

Companion data for the preprint Potion: Constant Context over Growing Histories, and What a Memory Leaderboard Measures (Robin L. de Bruijn, independent researcher, v1.1, 28 September 2026).

Record of first disclosure and the citation of record: https://doi.org/10.5281/zenodo.22966163 (all versions). This repository is a byte-identical mirror of the files inside the v1.1 Zenodo deposit potion-preprint-rows-v1.1.zip; MANIFEST-2026-09-28.json carries each file's SHA-256 prefix. If the two ever differ, Zenodo is the record.

Contact: robin@potionmemory.com. ORCID 0009-0003-3598-0133.

What is here

One CSV per number reported in the paper, named by benchmark tier, whose answers, configuration, hint condition, ruler and the number itself. Examples: beam10m_potion_certified_hinted_63.1.csv (AMB's judge is the default ruler and is not named in the file name), beam10m_potion_certified_hinted_beamscorer_65.4.csv, beam500k_potion_preceding_nohint_preview_74.6.csv. Each score file has five columns:

column	meaning
row_key	the BEAM benchmark's row identifier: conversation index within the tier file, ability, question hash
conversation	the conversation the question belongs to, as the source run labelled it (10M runs count c1..c10; the 100K and 500K runs count from c0)
ability	one of BEAM's ten abilities
score_<ruler>_draw1	the row's score under the ruler named in the header, judge draw 1
score_<ruler>_draw2	the same, judge draw 2; for a one-draw ruler the two columns are equal

Rulers in the headers: amb_judge (AMB's judge configuration: gemini-2.5-flash-lite on AMB's April scoring code, two draws); beam (BEAM's own scorer, gpt-4.1-mini at temperature 0, one draw); preview_judge (Gemini 3 Flash Preview on AMB's current prompts, the paper's "before" rater in section 3; two draws at 10M, one at 100K and 500K).

Whose answers: files beginning beam…_potion_… score Potion's answers (its own 27B reader, or AMB's answerer reading Potion's contexts where the name says ambprotocol; _reader_gemini31pro_ and _reader_qwen38_27b_ are the section 7 pair on one conversation). The two files beginning hindsight_published_answers_rescored_ score the verified leaderboard entry's published 10M answers, re-scored under each ruler, so that the paper's two pairings (section 5a) recompute from files; those answers are the entry's own public data.

MANIFEST-2026-09-28.json lists, per file, the row count, the mean that reproduces the paper's number, the file's SHA-256 prefix, and a note with every rule a reader needs to recompute it (rows carried from an earlier read, event-ordering rows that no arm read and enter as 0.0, single-draw rows). A score is the mean over the ten ability means (equal counts per ability), as in the paper.

Two files release section 5c's pre-June number on conversation 1, one per answer draw: beam10m_c1_potion_ambprotocol_prejune_A1_68.7.csv and beam10m_c1_potion_ambprotocol_prejune_A2_69.6.csv; the paper's 69.2 is the mean of the two.

beam500k_potion_preceding_blindcontrol_67.0.csv releases section 5b's 500K blind control: the 67.1 rows with the 140 abstention and knowledge-update rows as read without the hint lines; its abstention and knowledge-update means are the paper's 50.4 and 60.7.

context_tokens_per_question.csv gives the size of the context handed to the reader for every question the paper reports at 100K, 500K and 10M (400 + 700 + 200 rows), counted as in the paper's section 2: row_key, tier, conversation, ability, tokens_qwen3_8_27b (Qwen3.8-27B's tokenizer, used for every token count in the paper except where cl100k_base is named) and tokens_cl100k_base (the tokenizer of AMB's own context_tokens field). Three 10M event-ordering questions received an empty context and count 0.

Dataset Viewer configs

The viewer groups the files by ruler, because each ruler has its own score columns: amb_judge (18 files, columns score_amb_judge_draw1/2), beam_scorer (7 files, score_beam_draw1/2), preview_judge (4 files, score_preview_judge_draw1/2) and context_tokens (1 file). A viewer config concatenates its files; a paper number is always one file's mean, never a config's.

How to recompute a paper number

Load a score file, average the two draw columns per row, take the mean per ability, then the mean of the ten ability means. The manifest gives the expected result and the rules for the few files with carried or single-draw rows.

What is not here, and why

No retrieved contexts, no answers, no judge verdict text, no rubric text, no reference answers. The memory system that built the contexts is closed; the paper's Appendix A states what is withheld and why. The questions and rubrics are the benchmark's own (Tavakoli et al., 2026) and are available from its authors under their terms. Everything needed to check every number in the paper against its per-row file is in this release.

Citation

de Bruijn, R. L. (2026). Potion: Constant Context over Growing Histories, and What a Memory Leaderboard Measures (preprint v1.1).
Zenodo. https://doi.org/10.5281/zenodo.22966163

Versions

v1.1 (this release, 28 September 2026): the paper's corrections of text, figures and references; adds context_tokens_per_question.csv, the two section 5c files and the 500K blind-control file; the other 26 files are byte-identical to v1's. v1 (25 September 2026): https://doi.org/10.5281/zenodo.22966164.