Reproducibility checklist

Use this checklist when an NL Analytics output will support a paper, policy note, appendix, replication file, monitoring output, report, or dashboard. This page is the canonical list — other pages link here instead of repeating it.

Search definition

Record:

Exact query string.
Date range.
Selected metrics.
Search options used: section, speaker affiliation, speaker title, sector or country pre-filters, adjacent sentences. These change what the measures count.
Merge or comparison logic if searches were combined or compared.
Keyword Tool settings and the saved search protocol, if the query came out of a keyword-set review.

Review evidence

Keep notes from matched-sentence review:

Examples of true positives.
False positives and exclusions added.
Missing terms added during refinement.
Ambiguous terms and how they were handled.

Matched-sentence review supports auditability. It does not, by itself, prove construct validity.

Export context

Record:

Export date.
Corpus coverage statement used for the analysis, and the coverage check for the target sample.
Files downloaded (see the file reference).
Search metadata where available.
Product share link where used.

Search results are retained for 90 days. A search older than that must be rerun before its matched sentences can be inspected again, so do not postpone snippet review or treat a saved search as an archive. A saved search or share link is not a replication package: keep the exported files, query metadata, matched-sentence notes, and downstream code or notes together, outside the product.

Downstream analysis

Record:

Join key and identifier checks.
Sample restrictions.
Whether raw counts were normalized — and by which denominator, nr_of_sentences or nr_of_sentences_filtered (see the normalization caveat).
Any filtering of zero rows.
Version of the downstream dataset.

Zero Exposure means the selected query did not match sentences in that transcript. It does not mean the company had no real-world exposure to the topic.