Evidence and reproducibility · 16 November 2026
A source you can check.
The dissertation governs the interpretation of the study. This audit separates source facts, implementation corrections and synthetic teaching choices. The dissertation, existing study PDFs and Word files, and archival code supplements remain unchanged; the build verifies their SHA-256 hashes.
Why 20–22 topics appeared
The historical result was seven topics at a maximum-topic setting of 23, with coherence 0.768 (dissertation p. 163). The shipped synthetic reference selected 20 topics at coherence 0.7896. It contains three delivery topics, five social-promotion topics, two sports topics, one restaurant topic, one gaming topic and eight background or recipe topics. Repeated templates create additional coherent lexical groups. Browser preprocessing, clustering, reduction and random-number generation also differ from Python BERTopic.
The new default walkthrough uses seven explicitly assigned teaching categories and recalculates prevalence from synthetic truth. It does not fabricate a fitted model or present seven as an independently recovered result. Exploratory discovery continues to report the topic count it actually finds.
The full Python search on the current synthetic corpus recorded 1,200 candidate ceilings across 50 draws and selected 21 observed topics at ceiling 22, with c_v = 0.7962. The measured run is preserved in the validation record.
What was corrected
- Generated benchmark labels are identified as synthetic rather than human-coded.
- API outages, refusals and ambiguous labels no longer become valid votes or “no theme” classifications.
- Browser z statistics now use the same increase-positive direction as the Python implementation.
- Checkpoints are bound to inputs, taxonomy, prompts and model settings.
- Changing a corpus or an upstream result clears dependent browser results.
- MiniLM cannot silently fall back to hashed vectors.
- Validation reports include per-theme metrics; the gate checks positive support and sensitivity as well as agreement.
- The methods distinguish the historical protocol, maintained changes and synthetic calculations.
Source discrepancies requiring author judgment
These are preserved in the original papers. The values below describe the apparent inconsistency; they do not silently revise the historical record.
| Source | Reported | Issue |
|---|---|---|
| p. 164 | Restaurant pre-to-post change −30.8% | Table 3 shows 1.9% to 1.7%, approximately −10.5% using rounded values. |
| p. 164 | Sports z = 0.083; p = 0.41 | These do not agree under a two-sided normal test. A decimal typo is possible but not established. |
| p. 164 | Gaming change −45.46% | Table 3 shows an increase from 1.1% to 1.6%. |
| p. 191 | 704 brand tweets in text; 703 in Table 4 | The displayed brand counts sum to 703. |
What the archive can establish
Study 1 occupies pages 148–179; Study 2 occupies pages 180–205. Their preprint body text was compared with the dissertation after whitespace and quotation normalization. Study 2 matched; Study 1 differed in table reading order. This supports textual fidelity but does not make reflowed files byte-identical to original chapters. The chapter viewers therefore serve the complete, unchanged dissertation PDF.
The archive does not supply the restricted corpus, original model, naming ballots, human validation labels or exact execution environment. Synthetic period denominators are calibration choices. The original S1 final-fit code also refits an unseeded model without the selected topic ceiling; the maintained code retains the winning fitted candidate instead. These limits prevent a claim of exact historical recomputation.
Read and verify
Unchanged Study 1 · Unchanged Study 2 · Revised methods · Machine-readable protocol · Source checksums
Preserved earlier methods editions, before this audit: PDF · Word. These remain archival versions and contain claims superseded by the revised methods.