The build
Running the collection pipeline produces a large archive of market history as a by-product. This was the attempt to turn that into something someone else could use — a small website documenting what the data contains, precisely enough that a buyer could write code against it before paying for anything. It is presented here as an engineering case study; nothing is offered for sale and there is no purchase path.
The engineering worth describing is about preventing my own mistakes. Every factual claim on the site lives in a single file, so claims can be audited in one place instead of being scattered through templates. A check runs at build time and stops the build outright if internal working material reaches the public output — added after two such leaks got through an earlier check that ran too late in the process. And the tool that produces sample data raises an error on every path I had not finished, specifically so it cannot emit a realistic-looking sample that is actually made up.
The other thing worth reporting is that I killed the product's original selling point. It claimed coverage that competitors turned out to advertise too, so the claim came off the site rather than staying up unverified.
The technical version
A static, zero-JavaScript site documenting a seven-stream order-book archive of roughly 1.05 TB as of August 2026, growing about 12.8 GiB a day. That total is self-reported from cloud storage rather than derived from a committed file, so it is dated; the curated tier beneath it (206 markets, 408,414 repricing events) is machine-verifiable and unchanged.
Provenance enforced in code
Every factual claim on the site is confined to a single data file, so claims are auditable in one place. Unknown dataset fields are flagged rather than invented. The sample-extraction tool raises on every unimplemented path, so it cannot emit a plausible-looking fabricated sample. A custom build gate fails the build if internal material reaches public output — added after two real leaks got past an earlier exclusion that worked at the rendering layer rather than the build layer.
More on this — the same discipline applied to agent work, on how I work.
Killed differentiator
The product's original headline claim, continuous coverage through resolution, was researched, found indefensible once competitors were shown to advertise the same coverage, and removed from the site rather than kept as an unverified boast. It was replaced with a structurally verifiable claim instead. A separate draft claim, that no third party redistributed the feed, was checked, found false, and corrected.
Astro with static output, self-hosted fonts, zero JavaScript by default; no backend, database, authentication or payments by design. Custom Node build gate.