The Eclipse, the Clerk, and the Checker

The Assyrian king list is precise from 911 BCE and uncertain before it.

The reason is a solar eclipse. Assyrian records note one during the eponym of Bur-Sagale, and eclipses are not testimony — they are orbital mechanics. Anyone can compute which eclipses were visible from northern Mesopotamia in the ninth century BCE and find the one that fits. The chronicle stops being a claim about dates and becomes a claim that can be checked against the sky.

Everything before that point is a floating chronology: internally ordered, precisely spaced, and attached to nothing.

Three things “independent” means #

Chronology, genealogy and graph algorithms all refuse to take their own word for a result. They mean three different things by it, and the three do not substitute for each other.

Chronology means recomputable. The eclipse works because the verification route does not pass through the document. A written date tells you what a scribe wrote. An eclipse tells you what the solar system did, and the solar system did not have a position on Assyrian dynastic politics.

This is why the corpus distinguishes the eclipse anchor from the Venus tablet, which records observations that do not pin down a unique date the same way. Both are astronomical. Only one closes.

Genealogy means uncorrelated in its errors. The Genealogical Proof Standard requires reasonably exhaustive research, complete citations, analysis and correlation, resolution of conflicts, and a written conclusion. The reason all five are demanded rather than one strong record is that any record can be wrong invisibly. A clerk mistranscribes an age. An informant misremembers a birthplace. A marriage record with the wrong county on it looks precisely as authoritative as one with the right county.

Correlation works because independent records are wrong in different ways, so the errors cancel where the truth does not. Which produces the rule that matters most and gets ignored most: a derivative source — an index, an abstract, a transcription — inherits every error in its original and adds its own. It agrees with the original because it came from it. That agreement is not corroboration and looks exactly like corroboration.

Genealogists go further and classify within a source. A death certificate’s date of death is primary information from an attending physician. The birthplace field on the same certificate is often secondary, supplied by a grieving relative decades after the fact. Same page, same seal, different trust.

Graph algorithms mean checkable by a different computation. A shortest-path result is not verified by re-running the algorithm, which would reproduce whatever was wrong the first time. It is verified by reconstructing the path and summing its edges. A minimum spanning tree is verified against the cut and cycle properties, which any correct tree must satisfy no matter how it was built.

The checker is usually cheaper than the computation and shares none of its assumptions. That is the entire design requirement.

Why the three do not cover for each other #

Take a chain with a perfect anchor, perfectly resolved identities, and a structural error. The eclipse holds. Every record is independently sourced. The graph has a defect nobody tested for, and the chain is wrong.

Now take one with a checked structure and a fabricated anchor. Every property verifies. The whole thing is attached to a date somebody asserted.

Genealogy supplies the sharpest case of a set failing while its parts hold. A non-paternity event — an informal adoption never recorded as such, a concealed remarriage, a substitution at registration — produces a documentary record that is internally consistent and well sourced, because the people writing it down had no reason to write anything else. Every document passes review. The conflict appears only when genetic evidence is correlated against the whole set, and the base rate is not negligible: a 2005 review put the median around 3.7%, with later geneticists arguing for something closer to 2%.

Record-level review cannot find that. There is no defective record to find.

The audit that gets run #

Provenance work in finance — beneficial ownership, model lineage, data supply chains, collateral chains — is almost always audited on one axis.

Usually identity. Is this entity the entity we think it is, is the name matched correctly, has the record been checked. That is genealogy’s axis and it is a real one.

The anchor axis mostly goes unexamined. A chain of custody terminating in an attestation is not anchored; it is asserted. The question chronology asks is whether any link can be verified by a route that does not pass through the party attesting it, and the honest answer is often no.

The structural axis goes almost entirely unexamined, because it requires knowing what property the correct answer would have — and unlike a spanning tree, an ownership chain has no textbook certificate. Which is a reason to define one, not a reason to skip the check.

What to do with three axes #

Score them separately and look at whether the failures correlate.

If a chain that fails on identity also tends to fail on anchoring, one audit finds most of the risk and the current practice is roughly adequate. If the failures are independent — which is what three centuries of genealogical, chronological and algorithmic practice would predict, since each failure is invisible to the other two methods — then single-axis auditing is finding about a third of what is there and reporting it as a clean result.

Nobody has run that test on a real ownership register. It is a week of work and the answer is a number.

The Assyrians did not record the eclipse in order to be checked. They recorded it because it happened, and three thousand years later it is the only line in the document that cannot have been written to please anyone.