What genuinely connects Kelly et al. (2025)'s 'richer joint matrix' pointer and Samet's 1975 'more rating categories' critique? Real citation link, opposite prescription

intelligence-tradecraftadmiralty-codesource-evaluationepistemicsrating-scalesbridge-checkkelly-2025samet-1975icard

The task brief asks whether claim-kelly-et-al-propose-richer-joint-matrix-not-axis-collapse and claim-samet-1975-argued-for-more-rating-categories-not-fewer are a real bridge (cosine 0.87, previously unlinked) or a resemblance the retrieval index mistook for a connection.

Verdict up front, and it is not a clean yes/no. The specific bridge implied by the topic phrasing — that both sources answer "raters leak across the reliability/credibility axes" with the same fix, "more categories" — does not hold: read past the shared vocabulary, and Kelly et al./Icard's proposal and Samet's own prescription move in opposite directions on the deeper design question of how many axes to keep. But the pair is not a pure embedding false friend either, because there is a genuine, sourced citation link underneath the resemblance: Samet 1975 is one of the very papers Kelly et al. (2025) cite as evidence for the leak their discussion is responding to. The two target notes should be linked — not because they recommend the same fix, but because one is data the other's author read and chose not to follow.

Claim: Kelly et al./Icard and Samet propose opposite structural moves on axis count, dressed in the same "more granularity" language

Kelly et al.'s discussion points toward Icard's proposal as a comparison target for the Admiralty Code's two-axis scheme, without collapsing the axes: "As well, future studies could compare the reliability and perceived usefulness of the Admiralty Code to alternative methods that encode qualitative meaning at the 'cell' level. For example, Icard (2023, 2024) proposes a 3 (Honesty of Source: Honest vs. Imprecise vs. Dishonest) × 3 (Truth of Content: True vs. Indeterminate vs. False) matrix wherein the 9 categorizations (i.e., 'cells') are qualitatively well described" (verified against a fresh 2026-08-18 archive of the Cambridge page, sha256 recorded above; fragments "Honesty of Source" and "qualitatively well described" independently confirmed verbatim via quote_check against that archive). That structure crosses two named dimensions and keeps both — nine cells from two 3-point axes, not one axis at all.

Samet (1975), working the same "raters can't hold reliability and credibility apart" problem fifty years earlier, prescribed the reverse in his own implications section: "the two-dimensional evaluation should be replaced" by a single quantitative likelihood-of-truth rating, with a source's reliability index retained separately only for collection management, not for grading reports (Samet 1975, ARI Technical Paper 260, p. 20; "should be replaced" reconfirmed verbatim via quote_check against a fresh 2026-08-18 re-archive of the DTIC/archive.org OCR text, sha256 c5c1c1a799497076d66615393f7a7a446ab5d2cb49f19463c9a08f838fbe3644, identical to the sha already on file in claim-samet-1975-argued-for-more-rating-categories-not-fewer). What Samet then argued should be "more" is not axes but rungs on the single axis left standing, citing Bendig (1954): "more information can be transmitted by a rating scale that uses nine categories rather than five" (verified verbatim via quote_check against the same re-archive).

So "richer... not axis collapse" and "more rating categories, not fewer" sound like the same instinct but answer different questions. Kelly/Icard's "more" is cross-axis: keep two dimensions, subdivide each into three, get nine qualitatively distinct joint cells. Samet's "more" is within-axis: collapse two dimensions into one, then give that one axis nine rungs instead of five. On the vault's own recurring design fork — fuse the axes vs. add more structure — these two "more categories" arguments sit on opposite sides.

Claim: Samet 1975 is literally one of Kelly et al.'s own cited sources for the axes-leak evidence, but their discussion does not surface his contrary prescription

The pairing is not coincidental resemblance alone: Kelly et al.'s own §2 ("Challenges with applying the Admiralty Code") cites Samet directly, alongside Mandel et al. 2023, for the finding motivating their discussion: "These findings indicate that analysts may not be assigning source reliability and information credibility classifications independently from one another despite the instruction within the Admiralty Code to do so (Mandel et al., Reference Mandel, Irwin, Dhami and Budescu2023; Samet, Reference Samet1975)" (verified verbatim via quote_check against the 2026-08-18 archive, sha256 13396b1eac745c527068dffe83f29ec1a70acb6351bc1dc6f9387f8a67d6831b). This corroborates, from a second independent read, what the vault's own corrected audit trail on claim-kelly-et-al-propose-richer-joint-matrix-not-axis-collapse already established: Samet 1975 is named among the prior work Kelly et al. review as evidence for the leak (alongside Baker et al. 1968 and Miron et al. 1978), which their own decoding experiments then extend.

That makes Samet a genuine ancestor in Kelly et al.'s own literature, not just a name the vault's retrieval index happens to have filed near theirs. But Kelly et al.'s discussion, in pointing forward to Icard's richer joint matrix as the next thing to try, does not engage or cite Samet's own fifty-year-old prescriptive answer to the identical evidence — fuse the two axes into one finely graduated scale. The paper that cites Samet as a diagnosis does not cite Samet as a candidate cure, and the cure it does point to runs the opposite direction from the one Samet himself proposed. This sharpens, rather than resolves, the observation already on file in claim-no-practitioner-source-found-advocating-merging-reliability-credibility-axes: Samet remains the vault's clearest practitioner-side precedent for merging the axes, and he is sitting, uncredited for that specific recommendation, inside the reference list of the paper the vault cites for the opposite direction.

Claim: the two "more categories" arguments rest on different mechanisms, not just different scopes

Samet's granularity argument is explicitly information-theoretic: it descends from Bendig's (1954) rating-scale-length work, which the vault's existing lineage note traces to Miller's 1956 channel-capacity ceiling on absolute judgment — a claim about how much information a human observer can transmit along a single dimension before saturating. Kelly et al.'s pointer to Icard's 3×3 matrix carries no such argument. Reading Icard's own 2024 paper (already in the vault, Tier 1), his motivation for structure is semantic and typological — naming nine qualitatively distinct honesty × truth combinations so that ambiguous joint cells don't get flattened — not a claim about an observer's channel capacity or a scale's transmitted-information ceiling. The two "more categories is better" arguments therefore are not the same argument applied at different scopes; they come from unrelated literatures (psychophysics/information theory vs. formal semantic taxonomy) that happen to share a design conclusion's surface phrasing.

Why this reads as a real-but-narrow link, not the bridge the topic implies

The 0.87 cosine is tracking real shared material — both notes are about the same NATO scheme, the same "raters can't separate the axes" finding, and both use "more categories, not fewer" as a headline phrase — which is more substantive overlap than the vault's prior confirmed false-friend pairs (observation-gates-jevons-cosine-pairing-is-embedding-false-friend, observation-rhw-1986-werbos-docsub-cosine-bridge-is-thematic-not-causal), where the underlying facts turned out to be about genuinely different documents or decades. Here the documents are about the same instrument and the same failure. What breaks is narrower and more specific: the two sources answer "what should replace the leaking two-axis design" in opposite ways, and the connection the topic's phrasing implies — that Kelly et al.'s discussion is in some sense continuing or updating Samet's argument — does not hold. If anything, Kelly et al.'s own reviewed evidence runs through Samet, and their prescription runs away from his.

Further leads

• A.W. Bendig's 1954 Journal of Experimental Psychology paper ("Transmitted information and the length of rating scales") has never been read at primary in the vault — every claim about it runs through Samet's citation of it. Not attempted this session.
• Whether Icard's 2023 taxonomy paper (the origin of the 3×3 grid Kelly et al. cite) engages Miller/Bendig-style channel-capacity reasoning anywhere, or is purely typological — not checked directly against Icard 2023 this session; inferred here from the 2024 dynamic-logic paper's own framing.
• Samet's subjects split 21–16 in favor of replacing the two-dimensional scale with one — whether any later intelligence-tradecraft source (post-1975) revisits or cites that specific vote has not been searched.
• Kelly et al.'s full §2 "Challenges with applying the Admiralty Code" section likely contains additional citations (Baker et al. 1968; Miron et al. 1978) not individually quote-verified this session — only the Samet/Mandel sentence was pulled and checked.

Entity candidates

• A.W. Bendig — person — the actual foundational figure behind "more categories transmit more information": his 1954 psychophysics finding is what Samet cites and what the vault's whole granularity-per-axis thread ultimately rests on, and he has no entity page in 40-entities/ despite being cited by name in two existing claim-notes. Flagged first, per the standing blind spot: the newer paper's co-authors are easy to catch; the older figure the priority/ancestry claim rests on is not.
• entity-george-miller — person — already has an entity page; the broader information-theory ancestor behind Bendig/Samet's scale-length argument. Flag-to-confirm, not flag-to-create.
• entity-michael-samet — person — already has an entity page; this capture adds the finding that his own recommendation (fuse to one axis) runs counter to where Kelly et al. point, sharpening but not creating new biographical material.
• entity-benjamin-icard — person — already has an entity page; the actual author of the 3×3 matrix Kelly et al. cite secondhand.
• entity-megan-o-kelly, entity-david-r-mandel — already have entity pages; co-authors of the 2025 paper, confirmed present but not new to this capture.