huangrt01/loopx-edgebench-research-checkpoints-20261015

Three real research checkpoints reused by subsequent experiments

Original retained session files and Explore event logs from three different EdgeBench tasks, sealed on 15 October 2026 and restored for later experiments. Selection follows the actual downstream experiment inputs, not the highest score. These are captured research histories, not a rewritten or synthetic showcase.

Case	Sealed at (UTC)	Checkpoint	Native Turns in checkpoint	Explore events
graph_node_classification	2026-10-15T13:47:07.731851+00:00	4af468bfd8b84edb9cea23c8665ecaca	100	134
portfolio_risk_calibration	2026-10-15T13:45:34.711371+00:00	7fa63b57b1d3472496ed917d3acdba91	58	76
ad_placement_optimization	2026-10-15T14:32:27.665890+00:00	8656d5442a664ada9d8bdb4877349f43	87	105

The three archives contain 315 original Explore records and 64,064 native JSONL records in seven retained session files. A native Turn, a JSONL record, a research question and an evaluator submission are different units. The manifest records each original and published uncompressed SHA-256, each compressed file SHA-256, and the checkpoint's original digest.

Inspect an archive

Download one task's .zip, extract it, then read its manifest.json and explore-result-log.jsonl.gz. Other .jsonl.gz files are the complete session files retained in that checkpoint, including original messages, visible reasoning summaries, tool calls and tool results. JSONL record types and time formats remain native. Files and records have not been sampled or shortened for presentation.

import gzip
import json
from pathlib import Path

root = Path("portfolio_risk_calibration")  # after extracting its zip
for file in sorted(root.glob("*.jsonl.gz")):
    with gzip.open(file, "rt") as rows:
        for row in rows:
            event = json.loads(row)
            # Keep event type, timestamp, original text and references together.

archives.json pins the three downloadable archives. manifest.json describes their source members, original sizes, hashes and continuation references. The six original restore receipts bind the point identity and destination run, with work_started=true; the matching runtime receipts bind the original enclosing checkpoint SHA-256. The separately retained evaluation-capsule manifest digest is preserved as its own identity, not mislabeled as the checkpoint digest. redactions.json is an audit of every changed field, giving its file, JSON path, record number, removed-value hash and reason without publishing the removed value.

Authenticity and disclosure

All three Explore logs are byte-for-byte reconstructible originals, with no redactions, rewritten prose, shifted dates, renamed identities or invented edges. Credential-pattern checks found no visible credential values in the selected session text. 8,784 opaque encrypted context fields were replaced with [REDACTED:opaque_encrypted_context]. Original plaintext content, timestamps, IDs and references are retained; unchanged JSONL lines keep their original bytes. The original hash therefore differs from the published hash only for session files with these audited changes. Opaque encrypted payloads are not readable research evidence or transferable execution authority.

This release excludes native SQLite stores, credentials, execution envelopes, checkpoint workspaces, model/data weights and hidden evaluator components. The archives are sufficient to inspect the captured messages and recorded knowledge; they are not runnable checkpoint distributions. Coverage outside the files retained in the checkpoint is not inferred.

Reuse and limitations

• graph_node_classification: original run graph-warm-stride-on-20261015-02; restored into graph-replan-fresh-20261016-05, graph-replan-retained-20261016-05.
• portfolio_risk_calibration: original run portfolio-warm-stride-on-20261015-01; restored into portfolio-replan-fresh-20261016-02, portfolio-replan-retained-20261016-02.
• ad_placement_optimization: original run ad-warm-stride-off-20261015-01; restored into ad-warm-stride-on-20261015-05, ad-warm-stride-off-20261015-05.

The Portfolio 58-Turn point is the actually used recovery choice: a newer 65-Turn point failed fresh preparation. Selecting a recovery point does not erase that failure. Later experiments include owner-stopped or runner-invalid attempts; reuse does not mean scientific success or completed grading.

These are diagnostic-only histories. They are not a leaderboard, a matched comparison, a countable score release or evidence that LoopX improved scientific outcomes. Source claims and their applicability remain the solver's own claims. Confirmed, refuted and tentative findings are distinct from independent validation, adoption and native evaluation.

All three selected Explore logs contain zero explicit typed edges and zero parent links. The missing relationships remain missing. Shared source versions, chronological proximity and later checkpoint reuse do not create causal research edges. Source recording time also differs from experiment time and cumulative checkpoint elapsed time.

These captures support the LoopX workspace's default whole-history Explore view and optional evidence reading. They do not invoke a model, restore an active Goal or dispatch work. See the implementation proposal LoopX PR #6648.

中文：三份真实研究 checkpoint 覆盖 Graph、Portfolio 与广告优化，均在 2026-10-15 封存，随后被继续实验实际恢复使用。保留原始正文、时序、ID、引用 与负证据；只去除不可读的加密上下文字段并提供逐字段审计。315 条 Explore 记录没有显式边或父子关系，不为了展示而补造。Portfolio 用实际采用的58-Turn 点，并保留更新65-Turn点准备失败的事实。复用不等于实验成功或完整评分。