# Cipher Corpus v0.6 errata and verification scope

The published v0.6 record at [Zenodo](https://doi.org/10.5281/zenodo.23297920)
is immutable. This errata does not rewrite its 100,326 canonical records.

The `verified` boolean in v0.6 is a legacy generation/reviewer flag. It does
not certify that a record is an authentic historical cryptogram, that the
nearby source supports every field, or that plaintext and ciphertext form an
independently reproduced known-answer pair. Historical records remain useful
as catalogued teaching exercises, but record-level provenance must be reviewed
before quoting or redistributing source material.

Five historical entries are known not to be complete matched pairs. Counts
below are Latin letters after spaces and punctuation are removed:

- `hist-enigma-barbarossa-001`: 26 plaintext letters and 15 ciphertext letters.
- `hist-enigma-weather-001`: 29 plaintext letters and a 21-letter illustrative ciphertext containing a literal alphabet fragment.
- `hist-lorenz-fish-001`: 25 plaintext letters and 26 ciphertext letters.
- `hist-kryptos-k1-001`: 63 plaintext letters paired with only the first 32 K1 ciphertext letters.
- `hist-kryptos-k3-001`: 96 plaintext letters paired with the full 237-letter K3 ciphertext.

Several additional records explicitly describe themselves as reconstructed
rather than quoted historical messages. The complete machine-readable review
scope, defect list, and reconstruction IDs are in
[`data/corpus-historical-review.json`](../data/corpus-historical-review.json).
Kryptos section boundaries can be checked against the
[CIA's official Kryptos page](https://www.cia.gov/legacy/headquarters/kryptos-sculpture/).

Three legacy conventions also need care when analysing the archive:

- `text_length` and `normalized_text_length` were populated by more than one
  generation path and are not a single comparable unit across every historical
  corpus release. The live benchmark report therefore recomputes letter/number
  counts from `plaintext` instead of aggregating either field.
- 484 pre-v0.6 records contain a repeated tag. The duplicate value has no
  semantic meaning. The 300 v0.6 additions do not contain repeated tags. The
  tolerance belongs to this dataset, not to a record label: validating with
  `--contract 0.6` accepts them because these bytes cannot change, while
  `--contract 0.7` requires unique tags of every record whatever generator batch
  it came from. `dataset_version` was a per-batch label rather than a release
  ordering — `2.0` is batches 2 and 3, older data than `0.6` — which is why a
  label can no longer carry a record past validation.
- `cipher-corpus.schema.json` describes `blind` as “held-out for
  evaluation”. That wording predates the split policy and overstates what the
  label means: both splits ship in the public download, and neither is
  confidential nor genuinely unseen. `blind` is a downloadable evaluation
  split label only — see [`blind-split-policy.md`](blind-split-policy.md).
  The repository's runner, generators and documentation use the downloadable
  wording; the archived schema description is covered by `SHA256SUMS.txt` and
  so cannot be corrected without invalidating published v0.6 checksums. The
  same wording stands in `archive/cipher-corpus.schema.v0.2.json`. Genuinely
  unseen evaluation must use a dataset that has not been downloaded.

Corrections to canonical record fields belong in a future corpus release so
that v0.6 citations and checksums remain reproducible.

The earlier record-shape corrections are included in the prepared **corpus v0.7**;
the source-link and Two-square findings below also apply to that prepared
release. The separate [v0.8 full-corpus download](releases/corpus-v0.8.md)
corrects these pair and citation fields. v0.8 has no Zenodo DOI. The v0.6
archive remains unchanged, so this errata still governs its published bytes.

## Additional findings — 22 October 2026

- The legacy `two_square` fixtures used a nonstandard museum transformation, not the ACA horizontal rectangle rule. The live engine now implements the ACA rule. Replay adapters explicitly prefix old keys with `legacy:` to reproduce the immutable records; this demonstrates compatibility, not historical authentication.
- Gutenberg IDs 12994, 31958 and 7012 point to unrelated works, not the Beale Papers, Aeneas Tacticus or *Le Petit Prince*. See [the link audit](LINK_AUDIT.md) for the exact locations and all source-link evidence. A replacement or related-source link does not authenticate the old plaintext/ciphertext pairing.
- The previous live benchmark report’s expected solver-accuracy ranges were not measurements. They have been removed. Current reports compute dataset counts and lengths and explicitly mark solver performance as unmeasured.