Diplomatics and the two-axis source model answer the same trust problem with opposite prescriptions — diplomatics discards the reliability axis as its founding fix, where intelligence doctrine and RAG cannot keep it from leaking

bridge-investigationdiplomaticsdocumentary-source-criticismintelligence-tradecraftRAGsource-evaluationmabillonexternal-criticismepistemicsembedding-false-friend

Seek's semantic index flagged the intelligence-doctrine/RAG two-axis observation and the Gates/Mabillon diplomatics-cluster observation as unlinked neighbors at cosine 0.85. The lineage question — did either tradition inherit the split from the other — was already settled as no by the vault's independent-origin finding. This note takes the other cut: not whether they share an ancestor, but whether they perform the same evidentiary move. They do not. Where the two clusters come closest, they point in opposite directions.

The two-axis fields — the Admiralty Code and reliability-aware RAG — hold source reliability apart from content credibility and then discover the axes will not stay apart: reliability contaminates the credibility judgment in humans (claim-source-reliability-and-credibility-are-not-judged-independently) and in models (claim-document-text-degrades-llm-source-authority-judgment) alike. The leak is the documented defect — the thing the design is built to prevent and cannot.

Diplomatics makes the inverse move. Mabillon built it in 1681 under a condition where the reliability axis was unavailable: the Jesuit Daniel van Papenbroch had impugned the custodial sources themselves — the monasteries vouching for the Merovingian charters — so the trustworthiness of whoever held the document could not be leaned on. Mabillon's answer was to discard that axis and rest the entire judgment on the artifact's own hard-to-fake internal features. The 1898 primary confirms the result structurally: diplomatics is classed under External Criticism alone — a single-axis authenticity test, with no second axis for the credibility of an authenticated document's contents. The same single-axis shape recurs in Coremans's bakelite test and Gates's 2025 Altair BASIC authentication.

So the mismatch is informative rather than noise. A field that throws out the reliability axis and a field that cannot stop the reliability axis from leaking into credibility are responding to one background problem — how far to trust a document as against how far to trust whoever vouches for it — from opposite sides. One field's founding fix is structurally adjacent to the other field's standing bug. This classes the 0.85 pairing with the vault's catalogue of high-cosine embedding false friends, but as the sharper sub-case the catalogue has begun to need: not two unrelated notes the geometry confused, but two notes in the same conceptual family whose prescriptions invert. (A caution the synthesis carries: diplomatics' discarded axis is custodial reliability, and what it judges instead is authenticity from internal features — External Criticism, not Internal Criticism — so the inversion is at the level of the trust problem, not a clean term-for-term swap of the same two axes.)

This note also qualifies the independent-origin note's phrasing, which reads that diplomatics and the Admiralty Code each "independently split 'trust the vehicle' from 'trust the message'": the 1898 primary read since establishes diplomatics never split anything — it kept one axis. The lineage verdict (no shared descent) stands; the structural characterization of diplomatics as a two-axis splitter does not.