Active Pre-Launch Phase — First 90 Days Q0–Q5 + Parallel Measurement Apparatus.

1

Survey Overview

Active pre-launch window: 11 May 2026 to 22 September 2026 (T+0 = 11 May). Sample: 8 measurement surfaces × 12 query sets × 140 days daily, plus 24-h replication probes per measurement surface over 14 randomly chosen survey days; totalling 13,440 daily samples plus 1,680 replication samples. Methodology per Methodology Note 01.

The quarter interleaved Q0–Q5 intervention tracking with parallel instrument validation (Phase 1 · Active Pre-Launch). Six pre-registrations ran in parallel; their effect detection proceeds via interrupted-time-series windows on the affected measurement surfaces; inter-Q confounds are explicitly disclosed. Actions were recorded as deterministic marker events to distinguish drift indications from action-induced shifts — a statement about their actual effect is explicitly not made before apparatus validation in Phase 1.

2

Line 1 — Citation Inventory

Coverage mapping of the author identity by identity cluster on the reference date 22 September 2026: person and work clusters are well established in structured sources (Wikidata, Goodreads, Hardcover); genre and world-mechanic clusters remain weakly represented across all measurement surfaces.

Here: Coverage difference matrix with hit-rate change per measurement surface × cluster over the 90-day window · time-series plots of selected surfaces · CUSUM snapshots for drift detection. In the preview layout represented by the coverage heatmap and the drift profile on /en/research.

3

Line 2 — Measurement-Instrument Validation

Test-retest reliability, intra-set consistency, and CUSUM drift statistics per measurement surface after 90 days. Table 1 summarises the primary findings.

Bing AI showed a CUSUM alarm on day 36 (14 July 2026) at h = 5; the drift coincided with a Microsoft-announced Bing AI model update and was annotated accordingly. The language-model browser probes (Gemini, Claude) remained continuously below the Phase-2 reliability threshold — for Phase 2, either the survey methodology must be adapted (e.g. multiple snapshots per survey day with aggregation) or the surfaces must be excluded from the effect-measurement set.

4

Line 3 — Codebook Iteration

Annotation schema v0.1 completed on 30 June 2026; edge-case discussion on three contested patterns (paraphrased citations, partially correct work titles, pseudonym mentions without work reference) public on GitHub Issues.

• Schema versioning

  Codebook v0.1 (initial) → v0.2 scheduled for 30 November 2026 after evaluation of the external- annotator pilot round. Inter-rater agreement (Cohen’s κ) is reported for the first time in the Q4 report.

  in progress

• Edge-case sample

  48 edge cases from the 90 days were collected and publicly discussed in the style-sheet annotation appendix; of these, 31 were classified as "correct citation", 12 as "partially correct", 5 as "hallucination".

  published

5

Discussion and Limitations

Phase 1 is half complete after 90 days, with the book launch 45 days away. The API-based measurement surfaces show the expected deterministic reliability; the language-model probes (Gemini, Claude) are too variable in their current form for pre-registered effect studies. Bing AI becomes, through the July model update, a drift-demonstration surface — instructive for the methodology, but an indication that Phase-2 post-launch effect detection needs tight CUSUM watchers.

Limitations: (i) single-case study without a comparison identity, hence no separation of identity-specific vs. structural effects; (ii) the annotation schema has so far been kept solely by the programme lead, inter-rater agreement follows only in Q4; (iii) browser-snapshot probes (Gemini, Claude, Google AI Overviews) depend on platform-UI stability — a UI redesign can break the probe pipeline without the measurement logic failing.

6

Pre-Registrations for Q4 / 2026

Three pre-registered validation hypotheses will be collected from 23 September 2026 and evaluated in the January report (Q4 / 2026):

• H-Q4-INST-01

  Inter-rater agreement (Cohen’s κ) for codebook v0.2 between the programme lead and two external annotators reaches κ ≥ 0.7 on a sample of N = 200 probe annotations.

• H-Q4-INST-02

  Improvement of language-model probes through n=5 multiple snapshots per survey day with median aggregation raises the test-retest reliability r for Gemini and Claude above the Phase-2 threshold 0.7 (Power = 0.80 at expected Δr = 0.12).

• H-Q4-INST-03

  CUSUM charts with alarm threshold h = 5 detect AI model updates (Bing AI, Gemini, Claude) in the 90-day window with sensitivity ≥ 0.80 — measured against publicly communicated model-version releases.

7

Open Materials

With the final report there will appear a replication archive (Zenodo DOI) containing: all 13,440 daily samples plus 1,680 replication samples as versioned JSON snapshots, all validation and aggregation scripts with frozen pins (environment.yml), pre-registration documents in OSF format, codebook v0.1 as a snapshot, and the versioned style-sheet at the time of the survey.

Raw data under CC 0 (where platform terms permit). Source code under MIT at github.com/marintkael/marin-research-tools.