Sources, tiers and locators
Every public number in RetroFlops carries a source, a locator precise enough to find the number inside that source, and a confidence status. A figure that cannot be traced back to a document is not published, however plausible it looks.
Tiers
| Tier | What it is | May support publication |
|---|---|---|
| A | Manufacturer documentation, government or agency document, technical manual, official benchmark result | Yes, on its own |
| B | High-quality independent analysis: die-shot teardowns, engineering retrospectives, technical press with stated method | Yes, two agreeing sources |
| C | Wikis, encyclopedias, spec aggregators | Yes, as rumored only |
One tier A source, or two independent agreeing tier B sources, makes a value
confirmed. One tier B source with a precise locator and extract makes it
reported. A precise tier C claim is rumored. Reported and rumored values are
visible, but remain provisional and cannot feed automatic ratios, derived claims
or promoted comparisons.
Two acceptable sources that disagree are a conflict. The catalog records both values and their evidence without averaging them.
Locators
A URL is not a citation. Every source records a locator, such as a page, table, figure, section, or timecode. It must be precise enough for a reader to find the figure without reading the whole document. “Page 2-17, table 2-4” is a locator; “the datasheet” is not.
The locator field cannot be empty.
Research records: sources no adapter can read
Structured sources are read by adapters, and the value the adapter extracted can be re-derived mechanically from the cached artifact. PDFs, scans, printed manuals and photographs of a nameplate cannot be read that way, and they are exactly where the most valuable historic figures live.
For those, a figure is backed by a research record: an exact locator, a short verbatim extract containing the number, the date, and the name of whoever transcribed it. The extract is brief enough to verify the figure without republishing the source. Its SHA-256 hash is verified on every validation run, so a transcription edited or corrupted since review is caught.
The transcriber’s name is recorded so that a questionable reading has an owner. Anything ambiguous in the document, such as a smudged digit, an unclear unit, or a qualifying footnote, goes into the record’s notes instead of being resolved silently.
Validation enforces this: an approved measurement citing a government document, technical manual, book or periodical that no adapter reads, and that has no research record, is an error.
Provenance of the fetch
data:fetch is the only step in the pipeline that touches the network, and it
may only contact hosts on a reviewed allowlist. Adding a host is a deliberate
change, not a side effect of adding a source.
Redirects are followed one hop at a time and checked at every hop. An allowlist that is only consulted before the first request is decorative: an allowlisted host could hand the fetch to any origin it liked, and those bytes would be filed as evidence for a cited figure. Some legitimate redirects, including Internet Archive download URLs, point to per-item storage nodes. The allowlist declares those permitted targets on its entry, so a reviewer approves them with the host. When a redirect moves the request, the address that supplied the bytes is recorded next to the requested address.
What it retrieves is stored as evidence, never as fact: the URL, the timestamp, the HTTP status, the content type, the byte length and the SHA-256 of the response. It writes to the cache and to the provenance block of a source manifest, and nowhere else. It never writes a figure.
It also never overwrites an approved record. When an approved source’s upstream content has changed, the run reports the change and exits with a failure, leaving the manifest untouched. The change then goes through review like any other edit. This prevents an upstream document from silently changing beneath a cited figure.
Production builds never fetch anything. This is checked, not assumed:
pnpm build:offline-check runs the build with network access blocked at the
socket layer.
Archives
Sources that are not permanently hosted record an archive URL alongside the original. A separate periodic workflow re-checks links and archive availability, so link rot is found on a schedule instead of by a reader hitting a dead citation.
Licenses and usage notes
Sources carry their license where one is stated, and a usage note where quotation is constrained. SPEC data is the significant case: any figure taken from a published SPEC result carries the applicable SPEC Fair Use Rules note, which is displayed next to the figure, not buried in a colophon.
Apollo-era figures are taken primarily from NASA NTRS documents, which are public domain as United States government works, and which are cited by report number and page.