# Cipher Corpus v0.8 — corrected website download **Status (22 October 2026):** A complete compressed JSONL download is served from the museum website. It has **no Zenodo deposit or DOI**. The published [v0.6 archive](https://doi.org/10.5281/zenodo.23297920) remains a distinct, unchanged dataset; the tagged v0.7 snapshot remains separately identifiable. Cite the v0.6 DOI only when using v0.6 data. ## What changed v0.8 is derived from v0.7 and contains **100,313 canonical records**. It corrects the `two_square` known-answer examples, whose v0.7 ciphertext was produced by an older museum-only rectangle convention. Their plaintexts and keys are retained, while their ciphertexts are regenerated with the current American Cryptogram Association horizontal rectangle rule. The museum engine's `legacy:` prefix remains available to reproduce historical corpus bytes; that compatibility mode does not make the older transformation the standard rule. The [ACA Two-square reference](https://www.cryptogram.org/downloads/aca.info/ciphers/TwoSquare.pdf) supplies the published rule and example. v0.7's separate corrections to historical-pair status, withheld records, Kryptos K1, length definitions and duplicate tags carry forward. See the [v0.7 release notes](corpus-v0.7.md) for their exact scope. The [v0.6 errata](../CORPUS-V06-ERRATA.md) still governs v0.6's immutable bytes. Five unsupported source citations are also removed from their active link fields and preserved in explicit citation-erratum metadata for auditing. One was a Gutenberg page for a different work; four were dead URLs. **This does not establish an authentic source for the associated ciphertexts.** The [link audit](../LINK_AUDIT.md) records the old links and evidence. Other unverified source claims must still be checked at record level before reuse. **Compatibility:** code relying on a Two-square ciphertext from v0.6 or v0.7 must pin that version. Do not concatenate or silently substitute v0.8 ciphertext against v0.6/v0.7 keys in a benchmark. Use the record-level change inventory and version field to compare datasets. Historical source descriptions are context, not certification that every historical pair was transcribed from the linked source. ## Download and verify - [Complete corpus (`all.jsonl.gz`)](../../public/corpus/v0.8/all.jsonl.gz) - [SHA-256 checksums](../../public/corpus/v0.8/SHA256SUMS.txt) - [Record schema](../../public/corpus/v0.8/cipher-corpus.schema.json) - [Changes by record](../../public/corpus/v0.8/change-inventory.json) - [Statistics](../../public/corpus/v0.8/stats.json) - [Corrected historical supplement (87 entries)](../../public/corpus/v0.8/historical-supplement.jsonl) - [Corrected LLM evaluation export (300 prompts)](../../public/corpus/v0.8/llm-3shot-eval.jsonl) - [Supplement manifest](../../public/corpus/v0.8/supplements-manifest.json) - [Supplement checksums](../../public/corpus/v0.8/supplements-SHA256SUMS.txt) The **historical supplement is separate from the canonical full corpus**. It retains 87 entries from the older 100-entry supplemental download after removing 13 withdrawn records. Forty-two have canonical v0.8 IDs: 41 remain labelled historical and one Enigma teaching example has been reclassified synthetic. The other 45 remain for historical context with `verified: false`, `pair_status: unverified_pair`, `benchmark_eligible: false`, and explicit `supplement_inclusion: legacy_unreviewed`; their original v0.6 verification flag is preserved only as `legacy_verified_flag` for auditing. Neither a historical title nor a contextual source link authenticates a plaintext and ciphertext pair. The Beale wrong-work citation is retired in the canonical record, and the Aeneas wrong-work citation is retired from the supplemental record, each with a `citation_erratum` retaining its old URL. The new **LLM evaluation export** copies the 300 legacy prompts and replaces the three outdated Two-square prompt records using v0.8's corrected examples and challenges. Sixty-two fixed-convention prompts that previously asked for a `value=null` key now explain that there is no external key. All 300 prompts and their 1,200 source/example pair uses were mechanically checked against canonical v0.8 and the current engines. These public prompts are not a blind held-out evaluation set, and an exact-string key target may reject a different but equivalent valid key. Their original v0.6 download remains available as a clearly labelled archive; do not mix its Two-square prompts with the v0.8 cipher engine. The 45 supplemental-only historical entries remain unverified and excluded from benchmarks. A source review checked contextual passages for 13, could not access linked full text for 27, did not independently check four, and could not substantiate one specific historical claim. Wrong-work links and overstated descriptions were corrected without treating a contextual work as proof of a ciphertext. A separate [local-engine replay audit](../../data/corpus-v0.8-supplement-replay-audit.json) found 42 pairs reproducing in both directions and three with mismatches; those calculations do not authenticate any historical pair. See the [supporting download notes](../../public/corpus/v0.8/README.md) for details. Five entries whose claimed historical deployment or source could not be supported are now explicitly modern reconstructions, with their prior classification retained for auditing. Canonical v0.8 records are unchanged. ```sh cd public/corpus/v0.8 sha256sum -c SHA256SUMS.txt gunzip -k all.jsonl.gz ``` The download is the entire dataset, rather than a patch. `all.jsonl.gz` is a compressed UTF-8 JSON Lines file: decompress it, then read one JSON object per line. The checksum manifest verifies the file you received. The site browser uses a corrected v0.8 sample. The v0.6 JSON/CSV/subset buttons still present the published archive and should not be confused with the v0.8 downloads. To regenerate v0.8 from the immutable inputs, follow [`public/corpus/v0.8/README.md`](../../public/corpus/v0.8/README.md). The build and tests replay the corrected Two-square examples using the standard rule and check that unrelated cipher pairs have not changed.