Datasets

Every dataset in this project is a text file in the repo (data/), validated in CI against a JSON Schema (schema/), and served here verbatim.

Datasets

data/packs_index.yml is generated from data/sites/ for the index pages. It is derived, not authoritative, and CI fails if it drifts.

The directory is called sites and the entity registry is called operators because both names belong to a shared data contract this project did not invent and does not get to rename. In this dataset a “site” is a pack and an “operator” is a maker.

The provenance chain

A published number can be walked back to a page someone read:

a record in data/sites/  →  the signals it cites  →  the source each
signal names  →  the exact URL that signal was read on

CI enforces every link in that chain. A record with no signals fails. A signal naming a source that is not in the registry fails. It is the one property of this dataset that is not allowed to degrade.

How the data was collected

By an AI agent, one page at a time, on request. Not by a crawler, not on a schedule, and not by a person.

Say the exact thing, because the vague thing flatters us. The founding corpus was assembled on 2026-07-09 by an agent session (Claude, driving the Bussetech studio’s tooling). For each maker, the agent searched, then fetched each cited page once, read what came back, and recorded the facts with the URL. Nothing was crawled recursively. Nothing was polled on a timer. No page was fetched twice. And no human being sat and read these pages — an earlier version of this page said one had, and that was not true.

There is an unresolved tension here, and it is ours, not yours. Evergoods and GoRuck forbid using their sites “to spider, crawl, or scrape”. A single agent-initiated fetch of one product page is not spidering, and a browser fetches pages too — but we are not going to pretend the question is obviously settled in our favour, on a page whose whole argument is that a permissive robots.txt does not license conduct an operator forbids in prose. We applied that standard to the scheduled fetch layer and did not, at first, apply it to ourselves. The question is filed as an open studio decision rather than answered by whoever happened to be convenient. If either maker would rather we did not, we will remove their specifications; every one of them is a fact they published themselves, and none of them is worth this.

What follows is a statement about the fetch layer — the scheduled machinery — and it remains exactly true. Every source in data/sources.yml is status: manual; nothing is on a cron; each carries its reason in its own robots: block:

• Evergoods and GoRuck publish terms of service that prohibit using their sites “to spam, phish, pharm, pretext, spider, crawl, or scrape”. A permissive robots.txt does not license conduct the operator forbids in prose. Where the two disagree, the stricter one governs.
• Alpha One Niner asks that its materials not be copied or transmitted without written consent. We record facts and cite pages; we reproduce no materials.
• Mystery Ranch publishes no feed. There is nothing to poll politely.
• Filip Raboch and Sample permit reading, and we still do not poll them. They are one-person shops. A scheduled fetcher aimed at a solo maker’s store is rude in a way robots.txt has no vocabulary to express.
• Carryology, The Perfect Pack and Suburban publish clean, permissive feeds, and we do not touch those either. They are the people who have documented this subject for years. We read them, we credit them by name, and we link to them. We do not aggregate them on a cron.

Not one of the ten robots.txt files reviewed on 2026-07-09 named an AI or LLM crawler — no GPTBot, CCBot, ClaudeBot, anthropic-ai, or Google-Extended directive anywhere. We record that as a finding. It is an absence of instruction, not a grant of permission.

That finding lands closer to home than we first wrote it. The agent that collected this corpus is an AI agent, and it fetched these pages while those files said nothing about agents like it. We did not read that silence as permission to build a crawler, and none exists. Whether it was permission to fetch ten pages once is the open question above.

Transformations applied: unit conversion only (pounds and ounces to grams), always with the maker’s original figure preserved in a note. No value was normalized, averaged, or reconciled across sources.

Where sources disagree

They are left disagreeing. The dataset records both claims and says so on the record. Sample’s own page for the Article 404C reads 20L in its title and 18 L in its spec block; both are stored as signals from that one URL, and the record shows the conflict rather than choosing.

What is missing

Gaps are data. Each record and each maker carries a gaps list naming what we looked for and did not find — an unpublished weight, an unconfirmed founding year, an acquisition whose effects no source describes. An empty field would read as an answer. A named gap reads as the truth: we don’t know.

License

The datasets published by this project are licensed under CC BY 4.0 (studio default, ADR-0002). The project’s code is MIT-licensed (see the repo’s LICENSE).

The specifications are facts, and facts are not owned. The prose describing them on the makers’ own pages is theirs, and none of it is reproduced here.

Commerce

There are no affiliate links on this site, and no monetization plumbing in this repo. Whether there ever should be is an open question the studio has recorded and deliberately not answered.