The conclusion rests on checked relationships, original contexts and explicit population assumptions.

This record accompanies the readable report. Alcuin made the final assessment. Sonnet 5.5 produced the detailed research and the requested follow-up checks. “Checked by Alcuin” below means a source, script, data table or DNA read was inspected in this investigation. It does not mean that every assertion in a source was verified. The research archive includes working drafts for provenance; this record and the readable report supersede their uncorrected claims.

Alcuin checked the findings that materially changed the comparison.

The targeted read checks retain both positive and negative evidence. A DNA read is one sequenced fragment. Mapping quality estimates how reliably it was placed in the genome. Base quality estimates the confidence in a particular letter. Trimming removes letters near fragment ends, where ancient-DNA damage can be concentrated. These filters reduce common errors; they do not prove authenticity or historical identity.

The independent scripts use GRCh37/hg19 coordinates from Sahakyan’s original Supplementary Table S2, with current-tree aliases checked separately. The original alias lookup contained 232 positions, including fourteen CB-01 sites. C2 found that it missed unnamed or absent markers. The corrected audit contains 247 positions, including all 25 sites in the checked FamilyTreeDNA CB-01 panel and all three sites in its ZS2434/ZS2435 descendant panel. Sonnet checked the added coordinates against the public marker tables and assembly references. The analysis uses samtools 1.24 to read alignments, then parses CIGAR operations explicitly. It removes unmapped, secondary, supplementary, duplicate-flagged and quality-failed alignments. It compares mapping-quality thresholds 20 and 30, requires base quality at least 30, and compares removal of zero, three and seven bases at each end. It collapses repeated observations with identical start, end and strand into a single molecule for the audit. The primary modern tree file has SHA-256 38394cb274e27be2f5cca97237b5cdfb4cd03b3281bba3eb7f201a9eb0389b38. Sahakyan and colleagues, 2021

I37241’s two published double-stranded, partially UDG-treated libraries each have about 10% terminal C-to-T damage. Their mitochondrial-consensus match intervals are 0.992–1.000 and 0.989–0.999. These are authentication-related metrics, not direct estimates of Y contamination. The original table gives mean coverage 3.722 on the autosomal capture targets. That is not whole-Y coverage. Both libraries come from the same individual, so they are not two independent people. Akbari and colleagues, 2026, original tables

I4517’s retained ZS231 observations occur on distinct fragments and opposing strands; both have mapping quality 37 and base quality 36. ZS231 is a G→T transversion, a change that is not the usual ancient C→T or G→A deamination pattern. I14846’s single T at the same site is therefore useful but still thin. His original quality table reports too little data for a male-X contamination estimate. Missing sites are not coded as ancestral.

The independent audit does not turn a marker tally into an automatic terminal assignment. My initial explanation of I37241’s ten-versus-eleven count was wrong: the difference came from the marker list, not quality thresholds. The omitted ZS2401 site, hg19 Y:14,933,771, retains four derived G reads even with mapping and base quality ≥30 and seven-base trimming. I added the full defining-site panel and reran all six men. The final eleven-site result agrees with Sonnet. This correction also added a fourth ancestral CB-01 site, FGC13874, for R835.

The expanded audit retains I4517’s conflicting A at ZS2434, hg19 Y:8,511,782. Its read has mapping quality 37, base quality 32 and fourteen bases to the nearer end. The change is G→A at a CpG site, so ancient damage or recurrence is a more plausible concern than at ZS231. This single read cannot overcome the two independent ancestral G reads at ZS231 without an explanation of the inconsistency. My alias-only script had missed this position because the original spreadsheet left its marker-name field blank. Sonnet found that blind spot; the final table now displays it.

Sonnet withdrew the apparent upstream contradictions for I14846: their mapping quality was zero, so they did not reliably come from the intended Y position. The usable ZS231 read and the published manual J-Y3088 assignment agree. That supplies moderate support for CB-01, but not multiple-site confirmation. The original library table calls its treatment ss.half while AADR calls it ss.USER. Neither source publishes a usable X-contamination estimate for this boy.

The broader analysis excluded inconsistent sites such as Z2293 and Z27715 when they conflicted across otherwise well-supported branches. R835’s source label B877 is especially misleading if read as a modern terminal label: the 2021 marker list spans the present ZS241, FGC13873, FGC13863, ZS227, ZS222 and Z18271 levels. Its label alone cannot establish membership in all of them. Sahakyan and colleagues, 2021

R835’s earlier-related placement also has positive evidence: a non-damage-type T→C change at Z18264/FGC13870 and one at F23384 in the BY143137 family. The lone derived ZS241 G→A read is weaker evidence because damage can cause that substitution. AADR places him at BY143137 and flags hapConX at 1.8–2.9%. FamilyTreeDNA places him more deeply at FTH61646, but that one-man branch was built from this same ancient file; it is not independent confirmation. The ENA sample description establishes Civitanova Marche and an archaeological period range of 27 BC–AD 300. The Antonio burial-detail supplement could not be opened. None of the original sources inspected assigns him a Jewish identity.

The date estimates remain tied to their own methods. FamilyTreeDNA Discover

The YFull file used by the earlier selection was version 14.04.0: 68,881 nodes and 43,250 terminal haplogroups. Terminal haplogroups are not counts of tested men. Sonnet checked a later public 14.06 tree, which changes the examined candidate ages modestly. The main report uses explicit calendar estimates where available. YFull’s “before present” footnotes for ancient specimens must not automatically be assumed to define every modern-clade clock field. The archive preserves the printed values and the follow-up on that convention.

Lipson uses a stated 0.76 × 10⁻⁹ changes per site per year in its dating pipeline; Sahakyan uses 6.95 × 10⁻¹⁰ with its own uncertainty and Bayesian dating method. Closely matching estimates built from overlapping men and similar clocks are not independent confirmations. No rate was recalibrated here. No estimate for repeated exceptionally late fatherhood was invented. The preprint’s cited supplementary mutation Tables 1–9 were not found among its three deposited supplementary files; public marker tables were used for the targeted tests instead. Lipson and colleagues, 2025 preprint · Sahakyan and colleagues, 2021

The population calculations explain how a paternal frequency can arise before calculating detection.

Let x be the share of resident men on the proposed father line. Let ρ be their surviving-son production relative to other men. Between arrivals, the deterministic calculation uses x_next = ρx / (1 − x + ρx). A value of ρ = 1 preserves the share even when both groups grow. A value below 1 reduces it. A value above 1 increases it. Each incoming cohort adds the assumed number of paternal descendants to the region. The shared denominator follows the source population growth.

The detailed sensitivity uses ρ = 0.95, 1.00, 1.05 and 1.15; it uses 30-year generations and seven modelled arrival cohorts. The earliest main cohorts arrive near 2369 BC; smaller cohorts arrive later. It varies the assumed paternal membership of arriving males through 10%, 50% and 100%. These are alternative assumptions, not fitted estimates. The source population is about 1,002,557 people near 2362 BC, halved to an assumed 501,278 males. Population growth is about 1.0113 per generation and is extrapolated after the last available source-model date.

D3 traced the retained example pedigree instead of inferring Y from overall ancestry. In that one run, 67 of 67 male members of the two departing Levant households have a father chain to Noah. The incoming outside partners in those households are women. After local marriages, only 133 of 183 retained Levant-family males at 2200 BC have that father chain. The latter denominator contains followed family members, not all regional men. The observation describes the programmed example, not real Bronze Age households. Pruning and absent servants or retinues limit it.

The deterministic table uses the expected arrival cohorts, so it is a mean-field reference path. Because reproduction becomes nonlinear when the share is large, it is not exactly the average over random histories. D2 separately simulated 40,000 histories per setting, including random surviving sons and possible extinction. Its household-presence assumptions are independent even though the original runs could contain correlations. D3’s additional survival approximation uses Poisson offspring and density feedback from the deterministic path; it was checked against the full-founder simulation but not separately validated for the smaller founder fractions.

For an already recovered collection, let w be the relative probability that a carrier enters the collection, compared with another man. The carrier share among recovered men is xw / (1 − x + xw). Let r be the conditional chance that a recovered carrier’s informative mutations can be read and recognised. The expected recognised fraction is then r × xw / (1 − x + xw), before allowing false assignments. Under independent sampling, the probability of at least one recognised carrier among n men is 1 − (1 − O)ⁿ. Actual collections require site, date, burial, preservation, kinship and readability information to estimate those terms. A general preservation probability should not be multiplied in again when the denominator already consists of recovered men.

Recognition of a genetic branch is not recognition of Noah. If an older branch was already present among the receiving population, its ordinary local descendants contribute the same broad marker. The observed branch frequency then combines the proposed introduced family with that background. A source-specific private mutation or a securely contextualised narrower relationship is needed to distinguish the histories. This limitation applies even to perfect DNA.

The later forward calculation has a different denominator. It begins with Jacob’s household and computes A_next = (1 − m)A + mA_outside. Its paternal share is a product of (1 − p), where p = 2mφ is the assumed contribution of outside fathers and φ is the male fraction assigned to incorporation. A three-step counting delay is an input, not an established historical fact. The pure-Python independent check reproduces the modern Ashkenazi Y values 0.9137302243, 0.4992375692 and 0.1259263745 for the stated low, middle and high input sets. The zero-delay middle result is 0.4942294721.

The early map, the later pedigree, the community forward calculation and the branch-detection simulations are separate models. The map imports the earlier population model. The pedigree supplies selected ancestry-threshold counts. The forward model constructs an approximate outsider mean from those counts and carries it forward. It does not generate the full early-to-modern population history. The fitted 5–7% or 20% Jacob-share variants target imposed numbers; their resulting fit cannot count as another observation.

The review corrected material errors instead of treating worker drafts as established results.

The source record distinguishes inspected data from located material.

AADR was checked against Dataverse version 14.0, released 8 June 2026, carrying v66.p1. The local annotation’s MD5 matched a2db1ac16f0f3558ed66fb251e1d5c7d. Its 23,089 rows represent 21,433 distinct individual identifiers; repeat genetic versions are not additional people. Some important 2026 individuals are absent from that release, so Sonnet also inspected the original later datasets. The inventory covers nineteen dataset/publication entries and 183 individual records, with evidence tiers. Those records are a searched collection, not proof of an exhaustive worldwide census. AADR, v66.p1, Dataverse release 14.0

Primary archaeological and genetic material inspected includes Agranat-Tamir’s sample and kinship workbook; Feldman’s Philistine contexts; Haber’s Sidon and Beirut tables; Waldman’s Erfurt sample and quality workbook; Brace’s Norwich report and supplements; the Tàrrega paper; Akbari’s original individual and library tables; Sahakyan’s marker spreadsheet; Lipson’s preprint and available supplements; Shen and Oefner’s Samaritan studies; Waas’s original figure; Hammer’s original text; Rootsi’s Table 2; and Levy-Toledano’s original PDF and figures. The final supplementary context checks are retained with the relevant workers.

The most consequential access limits are the 2026 study’s anonymised archaeological metadata, deposited under an embargo to 2041; the missing Lipson mutation-table files; Waas’s unrecovered sample supplement; uncertain public documentation of the finer Tsedaka and Danfi family assignments; raw-only Mödling data that were not aligned here; the unverified I14918 burial association; and the Antonio supplement describing burial details for R835. No researchers were contacted. No new laboratory work was commissioned. The investigation did not use an autosomal profile to infer a withheld individual’s religion or precise location.

The execution record confirms Sonnet delegation and discloses auxiliary usage.

The main research model is recorded as claude-sonnet-5-5 in completed runs. Original A, B, C and A2 records also contain Haiku usage. Original D additionally records Opus usage. The saved usage totals do not identify the purpose of each auxiliary call, and no conversation transcript was retained. Claude documents separate advisor and auxiliary processing features, so an advisor is a possible explanation for the Opus entry, not a verified account of this run. No intentional replacement of the requested main model occurred. Later launches explicitly disabled the advisor and set auxiliary defaults to Sonnet. The final A3, B2, C2 and D3 usage records list only claude-sonnet-5-5. The official advisor documentation describes the optional mechanism; it does not establish why the earlier run recorded Opus.

The launcher used Claude CLI 2.1.296 with an explicit main-model name, safe mode, no session persistence, read/write tools limited by the assignments, and no fallback-model option. Changes to advisor and auxiliary defaults applied only to these process environments. The original projects were treated as read-only. The supervisor retained the final research design and interpretation. Raw worker drafts contain the substantive results but also the errors corrected above; they must not be quoted as the final conclusion.