yyoraa/evomem-swe-terminal-results-20260927

EvoMem: SWE results and Terminal paired scores

A sanitized public snapshot of existing experiment results, published on 2026-09-27. This repository contains evaluation records and sanitized task-output archives, not a full benchmark or model distribution. The test split label is a storage convention, not a new benchmark split assignment.

中文概览

• SWE 评分尝试已完成：290 条逐题记录，两组各 145 条，其中正式目标各 133 条、热身各 12 条。3 条评分异常保留 null，不改记失败。
• 严格同题对比：131 对，两组各答对 45 题（34.35%）；有记忆 11 胜、11 负、109 平。不能据此推断记忆普遍无效。
• Terminal 同题配对成绩：100 个有效同题配对中，无记忆 baseline 通过 79/100（79%），EvoMem 通过 78/100（78%）；两组都通过 74 题、都未通过 17 题，EvoMem 单独通过 4 题、baseline 单独通过 5 题。异常和缺少有效配对的任务不纳入；计划范围为每组 127 道题。
• Terminal 历史进度：2026-09-26 22:41 UTC 的快照记录 219/254 次执行已有结果（含 7 次任务异常），另有 35 次未启动。这不是通过数，也不是现在的实时运行状态。
• 新增逐题输出包：SWE 290 题、Terminal 219 个已有 trial 的可公开输出，含脱敏轨迹、补丁、memory 和测试输出。SWE 40 份、Terminal 24 份主轨迹因保护规则未公开；另有录屏、格式等删减，见清单。原始文件未改动，未重跑或重评分。

Download per-task outputs / 下载逐题输出

• SWE output archive — 77.3 MB
• Terminal output archive — 39.3 MB
• Coverage and redaction policy
• SWE file manifest / Terminal file manifest
• Browse the swe_output_index and terminal_output_index configurations to find each task's archive directory.

These are sanitized available outputs, not an unredacted complete copy. Core trajectory coverage is 250/290 SWE and 195/219 existing Terminal trials; withheld files are explicit. Final shared memory is not automatically the historical memory visible to each earlier task. See the policy before attempting replay.

SWE results

Arm	Official targets	Pass	Fail	Grader error (null)	Warmup pass / total
nomem	133	45	86	2	3 / 12
evomem-repo_sequence	133	46	86	1	4 / 12

The strict paired comparison requires identical instance ID and official chain/index, and valid native Boolean scores on both sides. It includes 131 target pairs: 34 both pass, 75 both fail, 11 EvoMem wins, and 11 EvoMem losses. Both arms score 45/131 = 34.35% on this same-instance comparison. The extra unpaired EvoMem pass corresponds to a missing valid no-memory grade. Scored-only arm accuracies use different denominators (131 and 132); their difference must not be presented as a paired improvement. Warmups are excluded from the main comparison. All 270 new grading attempts completed; the original 20 grades were retained. Completion does not mean every attempt yielded a valid grade. The source audit is dated 2026-09-25; publication did not run another evaluation or modify scores.

Files and schemas

• data/swe_records.jsonl — 290 rows. arm, repo, instance_id, position, is_target, official_chain, official_index, provenance, status, resolved, error_kind. resolved is native true/false, or null for a grading exception. status=graded includes both pass and fail; grader_error is not a scored assertion failure. provenance=original means one of the 20 retained grades; derived means one of the 270 new attempts. Position/index fields retain the source convention (warmups have index 0).
• data/swe_pairs.jsonl — 131 target pairs: ID, both scores, and outcome.
• data/swe_excluded_pairs.jsonl — 2 excluded target IDs and both score/status fields. Exclusion from the paired statistic does not delete records from the 290-row table.
• summaries/swe_summary.json — counts independently recomputed from the exported rows.
• reports/swe_comparison_zh.md — existing detailed Chinese report. Evidence filenames in that report identify retained source artifacts, not additional files distributed here.

Terminal: same-task paired scores

Overview updated on 2026-09-28 from existing public outputs; no rerun or regrading.

Arm	Valid same-task pairs	Pass	Fail	Paired accuracy
baseline (no memory)	100	79	21	79%
evomem	100	78	22	78%

The paired outcomes are 74 both pass, 17 both fail, 4 EvoMem-only pass, and 5 baseline-only pass. EvoMem minus baseline is -1 percentage point on this same-task paired set, not a demonstrated general or causal memory effect.

Eligibility requires the same task_id in both arms, empty exception_info, native numeric 0/1 rewards agreeing with each trial's verifier/reward.txt, and nonempty matching task checksums. All 7 exception trials are excluded, including those with a recorded reward; tasks missing a valid counterpart are also excluded. Exclusion changes only the comparison subset, not the stored raw results. This covers 100 of the 127 planned tasks per arm. Matching task checksums do not prove that every execution condition matched; an EvoMem arm label does not establish actual memory injection or use. The permanent sampler/EvoMem prior-history gap remains a qualification.

Sources: the already-public Terminal output index, manifest, and result/reward files in the Terminal archive. These outputs were collected on 2026-09-27, 13:19–13:20 UTC (sequentially, not atomically). The scores were recomputed from those existing outputs, not from a new runtime observation. The disagreement-case index includes all 9 Terminal pairs with different outcomes; actual delivered memory examples are available separately. These examples do not establish a causal memory effect.

Historical progress snapshot (separate from scores)

Observation timestamp: 2026-09-26T22:41:03.799574+00:00. This timestamp is not the upload time and does not assert live machine state on the publication date.

• Planned scope: 26 families, 127 tasks per arm, 52 family/arm sequences, 254 attempts.
• 219 trial directories started and 219 finalized records, including 7 task-error records.
• 7 sequences / 35 tasks were unstarted in the snapshot.
• A permanent sampler/EvoMem prior-history gap remains part of the run qualification.
• Preparation changed only dependency installation for 44 tasks; 83 direct-pytest tasks retained their scripts. Assertions, scoring, models and budgets were not changed for this publication.

data/terminal_sequence_progress.jsonl contains 52 rows, including unstarted sequences: family, arm, expected_tasks, started_trial_directories, finalized_records, task_error_records, and observed_at. Arms retain their source labels: baseline and evomem. finalized_records counts result artifacts, not successful solutions. A task error is not necessarily an assertion failure or a nonzero sequence-runner exit. Progress diagnostics do not establish controller admission or completion of the full experiment. summaries/terminal_snapshot.json repeats these boundaries in machine-readable form.

Provenance, privacy, and limitations

The initial metric tables use an explicit scalar-field allowlist. Their source digests remain in provenance.json; this original provenance file describes the initial metrics-only snapshot. The later task-output archives have their own versioned manifests, raw-source and public-output digests, coverage counts, redaction status, and timestamps under outputs/20260927/. checksums.sha256 lists checksums for the root README, initial metrics, and the 20260927 output-release files named in that manifest; it is not an inventory of every later addition. Its README entry is updated with this display-only correction. Later comparison and memory additions retain their own versioned checksum files. The original provenance.json is unchanged and does not attest to this later overview update.

Credentials, private configuration, private evidence and unreviewed binary formats are excluded. Generated patches, recorded tool output, and memory are now available only in sanitized form, with explicit withheld files and missing outputs. They may contain upstream task or code fragments; upstream licenses continue to apply and no new blanket license is asserted. This is not a full reproducibility bundle: complete workspace/container state and model configuration are not provided. Selected results do not establish a general or causal memory effect, and neither experiment proves the proposed wrong-acceptance-memory mechanism.