# TAF Tense Index — demo guide

A working local research prototype using 2026 materials. Annual coverage and automated annotation accuracy remain unverified.

## Start

Double-click `Start.command` in the tense-index folder. Python 3 is required. Open http://127.0.0.1:8766 on this computer. Starting again opens the existing instance. Keep its Terminal window open; Ctrl+C stops a manually launched server.

The interface, annotations, economic series and events are stored locally. Showing the app does not require a model or internet connection. External source links and new collection require internet access.

## Three-minute walkthrough

1. Select “20 participants · extended corpus · automated analysis”. Explain the past, present and future shares; the fourth class contains mixed or non-temporal units. Observed participants have equal weight within each window.
2. Switch from 30 to 7 days. Hover to see participants, units, sources and words. Gaps remain gaps. The coverage strip is not model confidence.
3. Click the chart to show evidence from that window. Open a source card and point out its title, date and processed volume. Publication-date proxies are marked.
4. Select a weekly event to display its date marker. Events provide context, not a causal explanation of index movements.
5. Show S&P 500, VIX and US EPU and explain standardization. Correlations require 20 matching pairs and sufficient daily coverage, so blank cells are expected in the automated panel.
6. Switch to the short-excerpt pilot: 130 assistant-annotated passages and the full TIME100 cohort selector. The two corpora are separate and their percentages are not one continuous scale.
7. Open the speech archive and coverage report, then annotation review. Reviewers see no model labels and can export their own answers.

## Suggested introduction

“We are testing whether the temporal orientation of public speech can be measured. This prototype combines real sources, speaker extraction, semantic analysis, equal participant weights and comparisons with external series. It demonstrates the research method; coverage is uneven and independent quality evaluation is still ahead.”

## Before presenting

Use January 1–September 26, 2026 and a 30-day window. Wait for the data and clear any single-participant filter. Local charts and federal transcripts work offline; original source websites do not.

The downloadable ZIP contains the portable application, without model weights or the private research cache. See methodology-en.md, panel-report-en.md, weekly-events-en.md and project-processing-en.md for details.