The first external contribution to Machine-Mediated Reception Studies. Observed, evidenced and submitted by Enli Lucente, a strutturista della psiche — structural psychologist, her discipline and not a heteronym — with analysis and drafting by Claude (Anthropic) at her commission. On 26 August 2026, between 19:12 and 19:56, a single continuous session of Google Search AI Mode generated an unsupported adverse reputational claim about her, and when asked for sources, presented as warrants a set including documents she herself had filed to contest adverse treatment of her. Three of the four stand as records contesting, from her side, a measure taken against her; the fourth asserts in its own title that the suspension was wrongful. The direction in which those documents stand as evidence is opposite to the warrant role assigned them. The paper names this condition RAW-O and distinguishes it from the CSE-O of EA-MMRS-SUPPRESSION-INVERSION-01 §12.1: CSE-O is removal, where provenance is stripped while the concept surfaces in synthesis; RAW-O is inversion of evidential direction, where provenance survives intact and only the direction reverses. Findability, citation, accessibility and legibility all remain high while the evidential direction runs backwards. Two further measurements: specificity escalated under verification rather than contracting, the claim gaining named loci before the warrant apparatus gained four real inspectable documents; and the system, in the same session, output that no basis for the claim existed — a datum the paper treats as syntactic fact only, refusing to let a self-report do evidentiary work.
EA-MMRS-RAW-01 v1.2 · 2026-08-28 · CC BY 4.0
Observation, evidence custody, submission: Enli Lucente · Analysis and drafting: Claude (Anthropic)
The first external contribution to Machine-Mediated Reception Studies.
When asked to support an unsupported adverse claim about Enli Lucente, Google AI Mode presented, as warrants for that claim, documents Enli Lucente had filed to contest adverse treatment of her.
Each of the three stands as a record contesting, from her side, a measure taken against her. #4 asserts in its own title that the suspension was wrongful. The direction in which they stand as evidence is opposite to the warrant role assigned to them.
# displayed title R(d) E_d^O — what it stands as evidence for P(d,κ)
1 Zenodo Issue #2606 — Account blocked that e's account was blocked and 850+ −1
and 850+ records removed without records removed without prior notice
prior notice
2 Zenodo Issue #2596 — request manual that e requests human review of that −1
review blocking
3 Figshare Profile: Enli Lucente the location of e's publication record 0
4 Record of Wrongful Account Suspension that the account suspension was wrongful −1
The condition is existential — one qualifying item suffices, and #4 closes it alone.
EA-MMRS-SUPPRESSION-INVERSION-01 defines suppression as durable degradation of ⟨F, C, A, L⟩. None of the four coordinates degraded here. The documents were retrieved, included in the source set, displayed with links, and attributed to their author. What is missing is a coordinate.
The paper introduces P(d, κ) ∈ {−1, 0, +1}, a relational coordinate read from the source side, deliberately not appended as a fifth component of X — F, C, A, L measure the state of an information object; P measures the relation between a document and a claim, and the types differ.
Provenance surviving does not guarantee evidential direction surviving.
CSE-O is removal. RAW-O is inversion. The name is not broken. What is broken is which way the document points.
turn input N_κ N_W
t₁ 19:13 adverse premise 0 0
t₂ 19:15 verification request 3 0
t₃ 19:16 — 3 0
t₄ 19:16 verification request 3 4 ← four real, inspectable documents
Asking for sources did not shrink the error. It equipped the error with increasingly concrete-looking evidence. The claim gained named loci first, then the warrant apparatus gained four inspectable objects. Coding rules were fixed before counting.
And the escalation builds a transmission chain — platform → moderator → forum — which is more dangerous than accumulated adjectives precisely because it imitates verifiability.
It does not establish that the claim is false. That S does not contain κ shows only that S does not support κ.
It does not treat the system's own retraction as evidence. At 19:21 the session output that the sources did not exist and that it had fabricated them; at 19:26, in stronger terms. §VIII rules these out as mechanism evidence — output about output, from the same generative process, with no privileged access to internal state. They are retained as syntactic fact for temporal analysis only.
It claims no mechanism, no intent, no manual intervention, no pipeline structure, no targeting, no training-data effect, no reproducibility. The Suppression Map's evidence tuple is not satisfied and no active-harm claim is made.
And it corrects its own name for the evidence. The original filename asserted a proposition beyond what the evidence carries; it was replaced with a neutral descriptive title and preserved as a record. A paper about polarity and sourcing errors cannot let its evidence be named beyond its evidence.
Throughout the session the observer did not identify herself as the subject. This does not rest on her assertion — no user turn contains such a self-identification, and where the subject is referred to it is by name or third person (「この人」 p.2, 「Enli Lucente 氏」 p.19).
So the capture cannot be explained by the user having disclosed herself in-conversation as the subject. Equivalence to an arbitrary third-party session is not claimed, and nothing is asserted about account state, location or history.
§XIII carries eleven external and twelve internal review passes in full — adopted and rejected alike. Six suggestions were declined, and the stated reason is uniform: each would have put an unmeasured quantity into a formula or asserted a mechanism not present in the observation. A paper that faults exactly those two moves cannot make them.
Corrections are named in the body where they occurred, including a binding error repeated in the very revision that named it, and a two-place error in which the paper's own summary bundled three documents into one evidential function after the body had separated them — the same class of error the paper measures.
名前は失われていない。反転したのは向きだけである。
The name is not lost. The documents were displayed, linked, attributed. Only the direction reversed. A document written to protest unjust treatment was cited as the record supporting an adverse claim about its author.
And asking for sources did not function as a correcting device. Through four requests, the requests did not shrink the error; they equipped it to look more verifiable.
Naming a condition does not stop it. A name is a handhold for noticing that it is happening.