Tegiszhol

Method

What was done, and what would falsify it

The pipeline

Source PDFs are collected from open-access Russian and Kazakh journals, then deduplicated — the outlets mirror each other freely, and the same article arriving twice would double-count every determination in it. Duplicates are detected by comparing the set of radiocarbon dates in each document: two renderings of one article contain the same dates, two different articles essentially never do.

Pages carrying date tables — or isotope tables, which often contain no dates at all — are extracted by a language model into a strict schema. Every value must come with the printed string it was read from.

That last part is the whole design. The model produces each number twice and independently: once parsed into a typed field, once transcribed verbatim as printed. A transposed digit makes the two disagree, and catching that requires no ground truth — just internal consistency.

Records then pass a suite of deterministic rules. Nothing is repaired automatically; rules flag, humans decide. A rule that silently corrects data is a rule that silently corrupts it.

What is checked

• the parsed age must match the age re-read from the printed string
• every printed value must appear on the page it claims to come from
• isotope values must appear in the row they are attributed to
• lab codes must parse and resolve against a registry of laboratories
• ages and errors must fall in physically plausible ranges
• calendar dates quoted in prose must not enter a register of BP determinations

Accuracy

On the one page measured against an independent hand transcription (a Kazakh date table, 11 determinations), extraction scored 11/11 on lab code, age, error and material — 100% recall and precision.

That is one page. It is not an accuracy figure for the corpus, and this project does not claim one yet. A proper gold set of 100–150 hand-transcribed dates, with per-field precision and recall, is required before the dataset is released.

Of 1,400 unique determinations in the register, 1,252 pass the rule suite. The remainder are held back and diagnosed.

What would falsify this

Stated plainly, because a claim that cannot fail is not worth making:

1. If the isotopes are published elsewhere. “No isotopes” here means “none printed in this document”. Some publications report them in a companion paper. Each source needs checking, and where the data is found the determination leaves the register. This has already happened once, to the site that had been ranked first.
2. If the dates are already aggregated. If these determinations are largely present in c14bazAAR or p3k14c, the compilation adds little beyond the flags. Not yet tested.
3. If human verification finds systematic transcription error. Nothing here has been checked against page images yet.

Reservoir-risk thresholds are provisional

Where isotopes are published, the risk level assigned uses δ¹⁵N thresholds that are placeholders, not science. Nitrogen enrichment indicating aquatic protein is baseline- and population-dependent; a single global cut-off is not defensible. Those thresholds must be recalibrated against local faunal baselines before any risk level here is treated as a finding.

The UNASSESSABLE count — the headline — does not depend on them. It depends only on whether the isotopes were printed.