pretty_name: EdgeBench Portfolio — two redacted execution trajectories
language:
- en
- zh
tags:
- traces
- codex
- edgebench
- loopx
- exploratory
EdgeBench Portfolio: selected redacted execution traces
Two terminal Codex executions of portfolio_risk_calibration, selected for academic inspection of plateau/retention and Explore behavior. They use different feedback modes, evaluator/task versions and runner sources. They are not a matched pair and do not establish a causal Explore effect. Both scores remain integrity-unqualified (score_countable=false).
| Run | Feedback / evaluator | Recorded best | Last observed | Tool call/result pairs |
|---|---|---|---|---|
| portfolio-heartbeat-resume-blind-ledgerv2-01 | blind / ledger-v2 | 17.32 | 15.57 | 1,303 |
| portfolio-heartbeat-explore-native-qualifiedv13-01 | native / VaR-timing-v1 | 31.19 | 31.19 | 1,822 |
Both had an 18-hour ceiling, gpt-6.1-sol, xhigh, 300-second host sampling and a 120-second native submission cooldown. Per-run .settings.json files contain exact source pins, image IDs, feedback contract, worker profile, reconstructed historical defaults and explicitly unknown provenance. See redaction-manifest.json for settings/source/export hashes and filtering counts. A historical best or last submission is not an independent final-artifact replay. Solver-written score claims in trace text are not authoritative evaluator results.
Reading the sessions
Open either .jsonl file in the Hub's Codex trace viewer / Data Studio. The files retain Codex response-item structure, timestamps, ordered tool calls/results and assistant messages. [HOME] and similar markers are redactions, not runnable paths. Commands are historical data, not instructions to execute. No model inference or benchmark run is needed to inspect them.
This is a redacted execution projection, not a complete raw-session archive. Source files were about 56.1 and 96.5 MiB. The projection omits duplicated event records, internal metadata, reasoning/encrypted blocks, compaction summaries, environment snapshots and original user/developer instruction text. It withholds potentially sensitive tool inputs/results, task/verifier-source reads, repeated instruction passages and fields exceeding 32,000 characters. Personal paths, local endpoints and email/account strings are replaced. Call IDs and call/result pairing remain intact; placeholders show omissions. Benchmark-specific solver implementation and local validation work remain visible where retained.
Screening combines automated pattern/overlap filtering with sampled semantic review. It is not a line-by-line certification. The manifest reports the omissions so researchers can judge whether a question is answerable from this projection. Original task materials, credentials, private operator state and complete solutions are not distributed as a separate archive. Third-party content retains its applicable rights; this dataset card does not grant a new license to underlying benchmark materials. Contact the dataset owner through a Hub discussion for corrections or reuse questions.
Relation to the ten-arm report
This dataset is the selected-trajectory attachment to the Portfolio ten-arm retrospective report in LoopX — ten-arm report (review PR #5750). The report provides all ten designated attempts and all recorded scoring observations, with separate ledger-v2 and VaR-timing-v1 cohorts. These two sessions were chosen after results were visible and are not a representative random sample. No significance, generalization or causal uplift claim follows from them.
中文说明
这两条轨迹用于与合作者公开分享，已脱敏，并保留工具调用顺序、调用与结果对应关系、时间戳及助手消息。原始任务／系统指令、内部元数据、推理加密块、大段敏感结果等已删除或占位，故不能当作完整原始会话。每条轨迹附精确 arm 配置与摘要校验值。两组反馈方式、评分器和代码版本不同，分数尚未通过完整性核验，不能直接据此宣称 Explore 的因果收益。