A SIDERA sample from edited session logs

Published 2026-09-14. SIDERA produced this sample from session logs prepared for publication. The input was edited before processing. The sizes and hashes below identify this public edition.

The order contained 20 session logs (9 Claude Code, 11 Codex), totaling 12,225,846 input bytes, billed as 13 MB. It cost 0.26 USDC on Solana. The service reported 753 seconds of processing, about 13 minutes. Order ID: 4c00b0fd-7cad-482e-a359-a4fab7579f62. Payment transaction: 5NejD4A5aAVD36bjRngwMPZSwnVKCCEiNi4jZBsW8XDG2xU7PzP5dG4vRAHQTKjKtDSdYsaNWxPzBfzZgRbMzpCC.

What is in memory.md

20 conversation sections, with section dates from 2026-06-11 to 2026-09-02; individual conversations continue into 2026-09-03. They contain background, a timeline, conclusions, topic labels and dates, supporting conversation, and source references into the archived logs. The text is mostly Japanese with English topic labels.

The conversations cover general development workflow, document and drawing preparation, printer and email-attachment troubleshooting, chainsaw troubleshooting, fireworks image selection, and video publishing.

Try it with a question

Ask: "There is an extensionless file called noname (1) next to a PDF. What was it, and was it a duplicate?"

Search memory.md for tnef. The section dated 2026-09-02T23:51 records the identification:

noname (1) は実行ファイルではなく、Outlook由来のTNEF（winmail.dat相当）コンテナであり、中に 御見積~2.PDF が1件入っているだけだと確認できた。

The later confirmation under duplicate attachment verification records the comparison:

sample-quotation.pdf（指定PDF）と、noname (1) 内に含まれるPDFはサイズ/SHA-256/全バイトが完全一致したため、noname (1) は重複梱包データで削除してよい。

Read the corresponding conversation dated 2026-09-02T23:56:43.930Z for the supporting comparison. The email attachment recovery topic also records its first appearance on 2026-07-28. Topic dates and individual conclusion dates have different roles; compare the evidence when answering a question about the latest state.

Follow a source reference into the submitted logs

Unpack raw.tar.zst with a tar/Zstandard-capable tool. It contains USAGE.md, raw-archive.mjs, and raw/ with each JSONL file and its adjacent .sidera/ directory. Keep them together: the reader restores compressed prefixes and any remaining tail.

The duplicate-comparison conversation has this reference:

<!-- source: {"log":"ext-codex/01a06467-bbe9-7c72-9c16-79713ff6fae1.jsonl","from_byte":73567,"to_byte":89984} -->

Resolve log relative to raw/ and read the uncompressed bytes [73567, 89984). These coordinates refer to the edited input submitted for this run. See the usage guide for the reader commands.

All 20 restored logs were verified byte for byte against that input. All 1,013 source comments (531 distinct ranges) were checked against the restored logs, including UTF-8 and JSONL boundaries.

About this public edition

The session text was edited before processing. The native JSONL record sequence and conversation timestamps were retained. Placeholders mark omitted material.

The public edition contains structured memory and the corresponding source archive. Source comments link conclusions to passages in the submitted logs. The sizes and hashes above identify the downloadable files.

“Lossless” here means that the archive restores the edited submitted logs exactly.