1 of 5
Loading the study
Verifying the synthetic corpus…
TopicLLM / Visual demonstration
See the method in action with synthetic data. Then fork the repository or install the Python package to reproduce and adapt the analysis in your own environment.
No original research posts are used. Follow the five stages, inspect the evidence, and recalculate the published prevalence pattern.
1 of 5
Verifying the synthetic corpus…
One response you can inspect
These small examples show what the LLM contributes. A single naming response is not democratic voting, and a single correct label does not validate a classifier. For the complete procedure, fork and run the Python workflow.
Checking live-example availability…
Try local discovery and mock naming to understand the mechanics. The mock can fail validation; that is a useful demonstration of the gate. For complete research workflows, use the Python package and your own API environment.
Load a corpus, tune the search, and point Stages 2 & 4 at an LLM. Parsing and topic discovery run locally. Hosted naming and classification send selected text to OpenAI through the Cloudflare gateway.
Reference implementation: 50 iterations × 24 topic solutions = 1,200 candidate models. Browser default is 8 iterations — raise it if you have time. The UMAP/HDBSCAN sampling spaces match the paper.
The teaching mock is available without credentials.
Finds the topics. A seeded random search over the model space; every candidate is scored, every draw is logged, and the argmax wins.
Names the topics. Many independent votes per topic; the modal label wins, and the full ballot box is yours to publish.
You make the themes. Read each topic's evidence, group topics by conceptual similarity, and record the two-researcher consensus.
Labels every document — but only after a classifier earns it against your human benchmark. Every theme also needs positive examples and nonzero sensitivity. Inspect per-theme performance before drawing conclusions.
Requires the Stage 3 consensus and a human-labeled benchmark (or the demo benchmark).
Turns labels into findings: theme prevalence by period and proportional z-tests under Bonferroni correction.
One ZIP mirroring the Python pipeline's pipeline_output/ tree —
every audit artifact produced so far, ready to publish alongside your findings.
This website is a visual demonstration. The guided study shows the seven-topic construction and real calculations from synthetic labels. Live examples send only fixed, fictional inputs through this site's Cloudflare Worker to OpenAI. The browser sandbox performs approximate discovery locally and uses a teaching mock. It is not the Python BERTopic implementation.
Fork the repository or download the Python package for your own corpus. An agent can prepare inputs, run checks, assemble review evidence and execute approved analysis; researcher interpretation and independent validation remain explicit checkpoints.
Each stage streams its work to a terminal panel: $ lines echo the command being
executed with its exact parameters; › lines report steps; · lines are
configuration detail; ✓ lines confirm completed computations; ! lines
are warnings; the ▸ line at the bottom is live progress. The status on the right
reads exit 0 when a stage completes cleanly — and gate refused when
Stage 4 declines to scale an unvalidated classifier.
Study walkthrough: the published seven-topic map and a calibrated synthetic calculation. Its category assignments are known from the generator, not discovered by a fitted model. Rounded prevalence matches by design.
Exploratory workspace: a local teaching exercise using synthetic or uploaded data and a mock provider. Browser preprocessing, UMAP, clustering and topic reduction differ from the Python BERTopic implementation. Its results must not be described as a rerun of the dissertation.
Python pipeline: the maintained research implementation. Its fixes, seeds and modern dependencies differ from the original S1–S3 supplements. Record versions, model identifiers, inputs, stopwords and outputs; a seed alone is not a reproducibility guarantee.
Upload a UTF-8 CSV with a tweet or text column and optional ISO date values (YYYY-MM-DD). Remove blank text and resolve duplicates according to a documented protocol. Browser discovery is limited to 6,000 documents because clustering uses quadratic memory; larger corpora belong in Python.
Use independently labeled, held-out documents that include positive examples of every theme. Check precision, sensitivity and confusion counts, as well as overall agreement. A classifier that misses every rare positive can still have very high accuracy. Generated labels and mock predictions do not establish human agreement.
Sandbox files stay in this browser tab. The public live examples accept fixed example IDs only; they do not accept your corpus, prompts or API keys. A signed, one-day session cookie supports usage limits. Read data handling and retention. Research data belong in your approved local or institutional environment.
The guided walkthrough and local sandbox make no paid calls. Public live examples are limited to ten calls per browser session per UTC day, thirty per shared IP, and the operator's site-wide allowance. The Python workflow requires an explicit paid-call budget, records usage, and resumes completed calls from its audit ledger.
Thomas, J. E. (2026). Coherence-Optimized Topic Modeling with Democratic Large Language Model Interpretation: A Human-in-the-Loop Pipeline for Public Health Surveillance of Social Media. Preprint.