# September 2026 revision record

The main article and rate reference now distinguish observed counts, model estimates, and hypothetical chronological requirements. This record explains corrections to the earlier working draft; raw sequence counts and source files were preserved.

1. **Ancient sampling times.** Branch-count ratios become annual-rate ratios only after dividing by each branch's duration. For ancestor age T and specimen age d, multiply the Neanderthal/human count ratio by T/(T−d). Extinction date and specimen date are separate inputs.
2. **Invalid branch allocation.** Removed the 30-human/170-Neanderthal allocation, its 5.7–11.4× requirements, and rejection probabilities based on it. Replaced the unconditional 1.36× ceiling with explicitly dated sensitivity scenarios.
3. **2025 egg study.** Added the methods and matched-cutoff comparison. The ≥5% coding-region estimate is 2.50 × 10⁻⁶/site/generation, versus the Icelandic 2.87 × 10⁻⁶. The egg study's lower cutoff accounts for much of the previously presented coding-rate difference; the control-region difference remains a separate issue.
4. **Eggs and transmission.** Removed the claim that sampled eggs preceded all bottleneck, drift, and selection. The study did not follow children and did not publish a 1,900-year Eve estimate.
5. **Bottleneck conversion.** Withdrew the homemade Icelandic 61,000- and 360,000-year clocks. An event count cannot be divided by a bottleneck estimate borrowed from another design without a matching model.
6. **Fixation terminology.** Separated detected heteroplasmy, transmission, sampled homoplasmy, and population fixation. Homoplasmy is not demonstrated permanence; population-wide fixation is not required for lineage divergence.
7. **Norfolk Island.** Retained the two homoplasmic events and broad interval. The Poisson comparison with Jeanson is conditional on compatible event definitions, detection, independence, and a constant rate; it is not a model-free rejection of every young history.
8. **Initial nuclear diversity.** Explained why diversity present in specified founders is distinct from diversity accumulated by mutation, while retaining the need to test founder models against genealogy, linkage, and allele frequencies.
9. **POLG.** Retained L580P as an untested candidate. Removed its dismissal by protein domain, assumed recessivity, or unadjusted mtDNA branches. Added no-lab checks, an existing variant-screening project, and purified-enzyme testing as a specialist experimental route.

Repeated argumentation and long quotations were shortened. The regional rate table moved to `../rate-reference.md`; its invalid all-study Eve-age column was removed. The wider literature catalogue remains available and is marked as containing older working assessments.

## Reproduction and limits

Run `node code/extinction-window.ts` from this research directory with Node 22.18 or later. It reads `data/neanderthal-rate-test.json` and writes `data/extinction-window.md` and `data/neanderthal-time-scenarios.json`.

The six scenarios use ancestor ages 6,000/4,500 years, specimen ages 0/2,000/3,000 years, and 25-year generations. They are assumed dates, not inferred dates. The consensus is composed of differently aged individuals and has no single real sampling date. Its conditional intervals and exact binomial tests omit full phylogenetic, alignment, recurrent-mutation, and dating uncertainty. Specimen-level comparisons are correlated and exploratory.

The original Python sequence analyses were not rerun for this revision. Their saved counts were reused; their equal-count outputs must not be presented as equal-annual-rate tests for ancient samples. No new functional experiment or independent whole-genome POLG variant validation was performed.

## Validation

The article and reference page were converted with the site's `MarkdownFile` renderer and superscripts enabled. All eight tables have consistent column spans; internal citation anchors resolve without duplicate IDs. The main article has 68 defined and used footnotes, with none missing or unused. The HTML study table uses explicit source anchors to avoid the renderer's repeated-footnote problem in Markdown-enabled cells. This checks content rendering, not the full deployed theme.

An independent PHP probability recurrence reproduced the TypeScript exact binomial results for all six scenarios. Converting the required rates back to counts reproduced both input branch means. The conditional Norfolk calculation gives P(Poisson(54.51) ≤ 2) = 3.269 × 10⁻²¹. Whitespace checks passed.

A [draft X reply](../reply-for-x.md) presents the defensible points and remaining limits without claiming that the egg study or POLG has established a young chronology.