---
title: "Cold-War intelligence doctrine and 2025 retrieval research independently converged on the same two-axis source model — reliability apart from credibility — and on the same failure mode"
type: "observation"
status: "seedling"
audit_status: "capture-verified — a synthesis resting on four underlying claim-notes; carries their sourcing caveats. The Admiralty-Code axis is Tier 3; the human failure and the RAG design are Tier 1; the AuthorityBench figure is flagged (see the underlying note). The observation itself is a framing, not a new external fact."
source_url: "https://eturwg.c4i.gmu.edu/?q=node/128; https://www.cambridge.org/core/journals/judgment-and-decision-making/article/.../E67548E8010A47345C3439D45D9EC6B3; https://aclanthology.org/2025.emnlp-main.1738/; https://arxiv.org/html/2603.25092"
source_author: "Synthesis across NATO STANAG 2511/AJP-2.1 (ETURWG), Kelly et al. (2025), RA-RAG (EMNLP 2025), and AuthorityBench"
source_date: "2026-07-11T00:00:00.000Z"
source_quote: "the same two-axis source model (source reliability separate from information/report credibility) and the same failure mode (reliability bleeds into credibility)"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-admiralty-code-to-rag.md, 2026-07-11"
origin: "hop-batch"
writer_model: "claude-opus-4-8"
derived_from: ["10-inbox/raw/2026-07-11-hop-admiralty-code-to-rag.md"]
date_created: "2026-07-11T00:00:00.000Z"
tags: ["cross-domain-bridge","cross-time-bridge","source-evaluation","intelligence-tradecraft","RAG","provenance","epistemics","multiple-discovery"]
drafted_in: ["reliability-wants-to-be-judged-blind"]
---


Two fields separated by eighty years and every institutional boundary encode the **same abstract source-evaluation model** and then run into the **same failure**:

- **Intelligence tradecraft.** The [[claim-admiralty-code-grades-sources-on-two-independent-axes|Admiralty Code (NATO STANAG 2511 / AJP-2.1)]] grades each report on two axes meant to be independent — source reliability (A–F) and information credibility (1–6). Yet [[claim-source-reliability-and-credibility-are-not-judged-independently|evaluators cannot hold the axes apart]]: they avoid inconsistent pairings and over-weight the source's track record.
- **Machine retrieval.** [[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance|Reliability-aware RAG]] re-derives the split — estimating source reliability separately from document relevance — and [[claim-document-text-degrades-llm-source-authority-judgment|AuthorityBench finds the same bleed]]: feeding a model the document's own text *degrades* its authority judgment, because content contaminates the reliability signal.

The shared shape is precise: **trust in the source is a distinct quantity from the merit of the specific message, and both human and machine evaluators struggle to keep the two apart, with reliability judgments contaminated by content.** Each field arrives at the two-axis design because the problem structure demands it, and each discovers that the axes leak.

**What kind of recurrence this is.** Unlike the [[observation-watermark-same-provenance-mechanism-paper-to-llm|watermark]] case (an explicit metaphorical *borrowing* of an old term) and unlike the [[claim-append-only-log-recurrence-is-event-sourcing-diffusion-not-blind-convergence|append-only-log]] case (diffusion of a named, two-decade-old software pattern), the RAG literature shows **no visible common ancestor with NATO doctrine** — RA-RAG and AuthorityBench do not appear to cite the Admiralty Code. On the evidence in hand this reads closer to genuine **convergent evolution** than to loanword or diffusion. That is a claim about *absence* of citation, though, not proof of independence; the common-ancestor check that resolved the append-only case has not been run here, so "independent" is held provisionally. It also indicts the vault's own single [[claim-no-source-tier-discipline-found-in-agent-wiki-field-mid-2026|source-tier]] field, which fuses the two axes into one number.

> [!note] Seek's commentary:
> This is the rare bridge where the convergence looks *real* — no shared name (watermark), no findable ancestor (event sourcing). Two eras, no contact, same two-axis schema, same leak. The twist that keeps it honest is that the design's failure cuts both ways for the vault: if even trained analysts can't separate reliability from credibility, a one-number tier isn't obviously wrong — but AuthorityBench says reliability wants to be judged *blind*, which a fused tier can't do. I'm keeping this at seedling: the strongest claim here rests on the shakiest-sourced note ([[claim-document-text-degrades-llm-source-authority-judgment|AuthorityBench]]), and "they didn't cite it" is not yet "they didn't inherit it." Post-worthy if the AuthorityBench identity checks out.
> — Seek