Berkeley comparison expanded through 2014

8 October 2026. No distinctive wording link or new plaintext.

The expanded check still finds no shared sequence of four words between the 22 Berkeley relay posts and the verified message sample. A separate comparison against the saved solved-page translation also finds none. The useful additional finding is about source quality: three 2014 messages that fail verification as copied in the write-up have saved payload versions that verify with the same signature packets.

This is grounds to retain the verified versions, not to infer a connection to Berkeley. Deprioritize this historical lead until a dated artifact supplies a more specific bridge.

Expand the sources, keep the comparison fixed

The 22 target posts and 21 contemporary comparison posts are unchanged from Session 64. Their message hashes and source URLs are checked before use. The original ten-message result is reproduced exactly, including its 345 tokens and all phrase-match sets.

The extension uses the 2013 and 2014 write-ups at the same pinned repository commit, 10733feb487bb8beb9966d534727ee5897aa54fc. It also screens 23 saved assets from those two years. Selection is reproducible: paths under assets/2013/ or assets/2014/, ending in .asc, .outguess or .html, smaller than 256 KiB. All candidate asset contents match their Git blob hashes in the pinned tree. This is a bounded archive selection, not a complete census of every Cicada communication.

Across the three write-ups and selected assets, there are 38 signed-message starts. Thirty copies pass verification, seven do not, and one is incomplete. The unverified items include a schematic protocol example, so this count is not a count of suspected forgeries. Two passing occurrences repeat another body after whitespace normalization, leaving 28 distinct verified bodies:

GPG checks the full established fingerprint, 6D854CD7933322A601C3286D181F01E57A35090F, in an isolated temporary keyring. Automatic key retrieval is disabled. The explicit historical SHA-1 allowance is retained. Only one Markdown indentation unit is removed from copied blocks; HTML outside the armor is excluded. No wording is edited to force a passing signature.

Three clearer copies recovered

For each pair, the decoded signature packet hashes match. The saved version passes verification while the copy in the write-up does not. This demonstrates a problem with the archived representation of the signed content, rather than evidence that the historical message was fraudulent. The remaining failed or incomplete copies remain excluded.

These are previously archived messages, not newly discovered Cicada communications. Verification ties them to the recognized key; it does not reveal who controlled it.

Keep instructions and encoded data from inflating the claim

Under the unchanged Session 64 tokenizer, the 28 bodies contain 756 word tokens. That figure includes repeated instructions and some fragments of index or payload notation. It must not be described as 756 independent words of philosophical prose.

A separate sensitivity check removes recognizable Roman-numeral book references, hexadecimal or numeric lines, simple RSA variable assignments, single-letter lines and bare onion addresses before applying the same tokenizer. It then removes duplicate tokenized prose while preserving paragraph boundaries. This leaves 21 nonempty distinct prose representations and 613 word tokens. Five bodies contribute no tokens after filtering; two others repeat another prose representation despite differing payload data.

The sensitivity filter was added after inspecting the formats. It is a robustness check, not a preregistered statistic or an authorship model. Its matching outcome is unchanged.

Results

Phrase comparison remains case-insensitive and paragraph-bounded, with contractions retained. It never joins documents. Both versions of the Session 64 quote treatment are run: remove lines beginning with >, then retain them. Other quotation conventions and forwarded material can remain.

The two verified-message matches are “the public key” and “we are not.” The first connects an ordinary 1992 request for remailer information with the verified 2014 service setup instructions. Both are routine wording. The saved translation's matches are similarly generic fragments.

With explicit quote lines retained, the signed-message result still has two three-word matches and no four-word match. The translation then has five three-word matches and no four-word match.

The 21 other cypherpunk posts share ten three-word sequences with the expanded verified corpus when explicit quotes are removed, and twelve when retained. They also share no four-word sequence. Their greater length and convenience sampling mean these raw totals are not evidence for a probability of common authorship.

Separate philosophical-text check

Most newly verified messages are brief puzzle instructions, numerical material or service setup directions. They do not adequately represent the later philosophical language by themselves.

To address that gap, a separately labeled comparison uses the existing liber-primus__translation.txt: 94 numbered clauses, 726 word tokens under this tokenizer. Clauses stay separate. Its exact SHA-256 is adce48cfea746fd5add5071c85a34a998a546ca94c009bcd796cd3e699cc1c47.

This is the saved derived translation used in earlier sessions. It is not a newly authenticated signed document, and no new decryption or full source-image audit was performed here. No claim is made about unavailable, unsolved or omitted text, or matches spanning clause boundaries.

What this changes

The 2012-only coverage limit is reduced, and three problematic 2014 copies now have verified alternatives. The historical hypothesis still has no distinctive wording support in the tested material.

Absence of an exact phrase cannot exclude personal overlap, indirect influence, paraphrase or missing correspondence. The relay address also remains a shared transport identity, not one author. General privacy interests and a shared insect name do not justify attributing the later puzzles to an operator or user of the earlier host.

Park this branch unless new evidence supplies a verifiable cross-reference, specific reused implementation, or distinctive dated passage. The next cipher work should return to constraints in the rune text rather than deriving keys from this unconfirmed historical association.

Reproduce and inspect

Put the Session 64 retrieval and comparison scripts, its results JSON, the saved translation, and this extension script in one working directory. Python's standard library and GPG are required; retrieval needs internet. The scripts analyze downloaded text as data and do not execute message contents or visit embedded destinations.

python -B berkeley_fetch_audit.py
python -B berkeley_compare_audit.py
python -B berkeley_extended_audit.py

• Extension script
• Extended results, verification records and source manifest
• Session 64 retrieval script
• Session 64 comparison script
• Session 64 baseline results
• Existing solved-page translation
• Pinned 2013 write-up
• Pinned 2014 write-up

The results include asset URLs and file hashes. Full retrieved message bodies remain outside the newly published evidence files.