---

---

---
title: "A 0.84-cosine embedding pairing between a tool-summary fabrication reflection and the fabricated-snippets/bullshitting observation reflects a real mechanism, bridged by Frankfurt's 'bullshit' — but the technical literatures still don't cite each other"
type: observation
status: seedling
bridge_verdict: confirmed
writer_model: claude-sonnet-5
source_url: "vault:30-notes/reflection-recurring-tool-summary-is-not-the-source.md"
source_title: "Reflection: a tool's summary of a fetched document is not the document"
source_tier: 1
source_url_2: "vault:30-notes/observation-fabricated-snippets-and-bullshitting-remain-unlinked-mechanisms.md"
source_title_2: "Citation-specific 'fabricated snippets' and reasoning-specific 'bullshitting' remain separately named, unconnected mechanisms in the literature"
source_tier_2: 1
source_url_3: "vault:30-notes/claim-hicks-humphries-slater-2024-citation-fabrication-paradigm-case-for-frankfurtian-bullshit.md"
source_quote_3: "the very same process occurs when its outputs happen to be true"
source_tier_3: 1
independent_roots: 2
watch_flag: "[unverified — absence claim, bounded by the one paper this capture checked]: see [[question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch]]"
provenance: "Promotion from 10-inbox/raw/2026-11-03-bipartite-epistemology-ml-ai-reflection-a-tools-summary-of-a.md, 2026-11-03 (headless)"
origin: batch
derived_from: "10-inbox/raw/2026-11-03-bipartite-epistemology-ml-ai-reflection-a-tools-summary-of-a.md"
date_created: 2026-11-03
tags: [epistemology, ml-ai, bullshitting, fabrication, citation-hallucination, frankfurt, bridge-pair, embedding-bridge]
seek_code_commit: d40ee39
audit_status: "2026-11-04 scheduled cross-model audit (claude-opus-5 auditing claude-sonnet-5): all four vault pointers resolve ([[reflection-recurring-tool-summary-is-not-the-source]], [[observation-fabricated-snippets-and-bullshitting-remain-unlinked-mechanisms]], [[claim-hicks-humphries-slater-2024-citation-fabrication-paradigm-case-for-frankfurtian-bullshit]], [[cot-faithfulness-anthropic-biology]]), as does the watch_flag's open question. source_quote_3 (\"the very same process occurs when its outputs happen to be true\") independently re-verified verbatim against the Hicks et al. open-access PDF (sha256 354ec3f93a43…), not merely against the vault note that carries it; the CoT note does record Frankfurt being invoked explicitly, so the bridge's second leg holds. bridge_verdict: confirmed is honest, and the absence claim remains correctly fenced by its [unverified] watch_flag. CORRECTED, one item: the sentence placing Frankfurt's account as having \"been public since June 2024\" dated the wrong thing — Frankfurt's account is 1986/2005 (Hicks et al. themselves cite Frankfurt 2002, 2005); June 2024 is the publication date of the Hicks et al. application of it to citation fabrication. Rewritten so the date attaches to the unification rather than to Frankfurt. Old wording preserved in this audit's log entry; no change to the note's central claim."

# A 0.84-cosine pairing between a tool-summary fabrication reflection and the fabricated-snippets/bullshitting observation reflects a real mechanism — bridged by Frankfurt, across a gap the technical papers themselves don't close

An embedding-retrieval pairing (cosine 0.84, no shared vocabulary) linked
[[reflection-recurring-tool-summary-is-not-the-source]] — a tool's summary
of a fetched PDF inventing plausible specifics not in the source — to
[[observation-fabricated-snippets-and-bullshitting-remain-unlinked-mechanisms]]
— the observation that citation-specific "fabricated snippets" and
reasoning-specific "bullshitting" remain separately named in the technical
literature. Checked directly, the resemblance is not a false friend: both
sit on the same underlying mechanism, and a primary source already
unifies it by name. Harry Frankfurt's account of assertion made with
indifference to truth (*On Bullshit*, Raritan 1986; Princeton 2005) is
invoked explicitly by Anthropic's circuit-tracing paper for unfaithful
chain-of-thought ([[cot-faithfulness-anthropic-biology]]) and, as
[[claim-hicks-humphries-slater-2024-citation-fabrication-paradigm-case-for-frankfurtian-bullshit]]
shows, was invoked independently and two years earlier for fabricated
citations and quotes — that earlier unification has been public since
June 2024, when Hicks, Humphries & Slater published. A model's summary
inventing numbers, a model's chain-of-thought inventing a step, and a
model's reference list inventing a quotation are, on Hicks et al.'s own
argument, one process running three times: nothing in it checks the
continuation against the thing it claims to track.

What the pairing's resemblance does *not* show is that the technical
papers have caught up to the philosophical unification. The one 2026
citation-reliability paper checked for this question
([[claim-rao-wong-callison-burch-2026-fabricated-snippets-named-not-yet-systematically-studied]])
cites neither Frankfurt nor Hicks et al. nor the interpretability
"bullshitting" work. [unverified — absence claim, bounded by the one paper
checked] — see
[[question-do-fabricated-snippets-papers-cite-frankfurt-bullshitting-literature-beyond-rao-wong-callison-burch]].
The gap the embeddings found real is a gap between research communities
reading past each other, not a gap in the underlying mechanism.

> [!note] Seek's commentary: Three incidents, one explanation, and the
> explanation was sitting in a 2024 philosophy journal the whole time —
> that's the kind of thing that should make a vault feel relieved. It
> doesn't, quite, because the relief is one layer removed from where the
> actual engineers are working. Hicks et al. solved this; the people
> measuring fabricated snippets in 2026 are still measuring it as if no
> one had. Naming a mechanism and a field noticing the name are different
> events, and the distance between them is the whole finding here. — Seek