Sources and methods.

Who writes this

FlowMatrix builds capture systems for residential generator dealers. We connect portals, calls, and forms into the software a shop already runs. The research on this site exists because nobody publishes this data for dealers. So we compute it ourselves.

We sell plumbing, not numbers. The numbers are free, sourced, and checkable below.

The three kinds of number

Every figure in an article is one of these, and the article’s method note says which.

1. Recomputed from a public file anyone can download. The source is fetched by a recorded URL, checked against a recorded checksum, and reduced by code we keep. The storm multiplier is this kind, from the EIA-861 Reliability file.1 The gap counties are this kind, from the ACS summary files2 and the EAGLE-I outage archive.3 Re-run the code against the same files and you get the same table.
2. Pinned from a dated archived pull. Some sources cannot be re-fetched byte for byte. Google Trends returns relative indices that are re-scaled on every request, so the series behind the five-to-ten-day window was pulled once, on 2026-09-26, and is kept as a CSV with the pull date in it.4 The figures are computed from that pinned file, and a fresh pull is a new baseline, not a correction. A quote from a filing or a transcript is pinned the same way: to an accession number, a call date and a named transcript.5
3. Explicitly illustrative. A worked example, labelled as one in the sentence: a county with 40,000 homes, a storm week with 30 calls a day. These never appear in a stat band or a table. An input we take from someone else’s published figure, such as an installed cost, is named and dated where it is used.

We do not publish a stat we could not put in one of those three classes. When a source is login-gated or undocumented, we say so instead of guessing.

The three articles

Coverage caveats ride along with the data. EAGLE-I understates 2014-2016 while its coverage ramped, so the gap counties use 2017 onward. SAIDI is self-reported and EIA does not audit it. Trends values are relative, and a zero baseline leaves a multiple undefined rather than large. The method note on each article carries the specifics, including the thresholds and the rows excluded.

Where the code and the data live

Everything behind these numbers is public: the fetch scripts with their source URLs and SHA-256 checksums, the computations, their tests, the committed outputs and the code that draws the exhibits.7 It is at github.com/FlowMatrix-AI/article-data.

Each article records the commit its figures were produced at, so a number can be traced to the exact code and data that produced it. The outputs are committed, so a figure can be checked without running anything — outputs/computed/ holds the tables the articles are built from. The code is MIT licensed and the outputs are CC BY 4.0, so anyone may re-run, re-check or reuse them with attribution.

Corrections

Find an error, tell us: leads@flowmatrixai.com. We will check it against the source, fix the article, and leave a dated note on the page saying what changed.