Provenance

How a number here traces back to a sentence on a page, and what happens when that page can no longer be checked.

Home · The measurement · The engine · Provenance · Tests

The graph is provenance, not prediction.

It exists to answer one question about any number on this site: which sentence, on which page, in which book, does this come from — and can that still be checked?

The chain

Every measurement traces back the same way, and the chain is refusable at every link:

a claim	names the rule that produced it
a rule	names the book and page it was implemented from
a page	carries the sha256 of the text that was read
a quote	is verbatim, paraphrase or inferred, and must say which

A rule that cannot verify its page refuses to claim it paraphrases one. It demotes itself from paraphrase to inferred and emits a null hash beside the claim, rather than asserting a provenance it can no longer support.

The layers, which do not pool

Two transcriptions of the same shelf exist, and they are different evidence:

publisher text layer	the quoting layer — what a verbatim quote may be checked against
OCR re-derivation	the findings layer — 144 books, 21,471 files. A hit is a candidate. An absence is never a finding.

That rule earned its keep, on the one page where it mattered most.

The OCR layer read a damaged 0 as a 9 in Gann’s own corn table — the table underneath the only claim in 144 books that beat its null. We published the mismatch as an error in his book. It was an error in our transcription of his book, and his own prose had said so all along.

What the graph cannot do now, and why that is said out loud

The publisher text layer was destroyed in a container reset on 2026-09-07 and is not recoverable from this side. Its consequences are visible rather than papered over:

every rule citing a page has demoted its own provenance, exactly as designed
1,377 tests that assert exact strings from that layer are skipped with the reason attached — not silently passed, and not reported as code failures
one test fails on purpose, so a missing corpus produces one loud failure that names the cause instead of a thousand quiet skips

Separately: the bar snapshots that priced the forward claims were destroyed in the same reset. 0 of 5,435 claims and 0 of 1,327 grades reference a snapshot that still exists. They were not re-stamped with a surviving snapshot id — that would have made the ledger look reproducible while making every row lie about its own provenance.

A claim whose bars are lost is unverifiable, which is a different fact from wrong. The graph's job is to keep those two apart.