# The Shape of Taking

## An event register, a measured trait space, and the limits of automatic taxonomy in the world history of land-taking and removal, c. 6500 BCE to the present

**Craig Talbert**
*(sole author; AI systems credited as instruments in the contribution
statement — decision recorded 2026-08-21; affiliation to add if desired)*

**Reuse notice (authorized 2026-09-05).** Original text and graphics:
© 2026 Craig Talbert, licensed under
[CC BY-SA 4.0 International](https://creativecommons.org/licenses/by-sa/4.0/).
Credit, identify changes, and apply the required ShareAlike terms when sharing
adaptations. Third-party quotations and reproduced material are excluded from
this grant. The project's structured research data are separately CC BY 4.0;
original software is MIT. See the
[component licenses and exceptions](https://latrian.dyndns.org/takingtheirland/reuse/).
This notice changes reuse permissions, not the manuscript's research status.

**WORKING DRAFT v1.6 — release edition dated 2026-09-08.**
This revision reports `second_pass_v7_676`, including the completed correction
cycle and chronology addendum in §5.4. The publication banner and accompanying
release pointer distinguish an unpublished preparation from an active public
edition. This manuscript is neither peer reviewed nor final.
Supersedes v0 by applying the seven-item correction order in
`MONOGRAPH_DRAFT_v0_FACT_CHECK.md` (independent findings-only audit,
2026-08-21). v0 is preserved unchanged for review. Bracketed `[TODO]` markers
flag remaining production or verification work; the author decisions below
are resolved. Alternate title:
*"Taking Their Land: A Measured Register of Settlement, Displacement, and
Removal."*

Version 1.5 recorded the completed finite post-v5 correction cycle, its
three-stage Editorial Suite, and the validated promotion of immutable
`second_pass_v6_675`. The 2026-09-05 wording correction distinguishes current
status from historical checkpoints; it does not change that research release.

**Release and publication snapshot for this edition (2026-09-08).**
`second_pass_v7_676` contains 676 modeled events,
of which 672 have reviewed primary-family assignments and four remain explicit
family evidence holds. Its canonical preprocessing retains nine exploratory
factors; the retained count and factor space remain sensitive to declared
alternatives (§5.4). The eight-trait analytical profile summarizes hand-coded
anchored judgments (AI-assisted as disclosed in §4), not the nine-factor
solution. The six primary families are source-reviewed editorial groupings,
not discovered clusters. These are three different products, not competing
counts of the same thing.

The atlas accompanying this edition has 370 authored cases and 380 playback
stops, unchanged from the preceding public build `20260906a`. Statistical
promotion does not silently render unauthored cases. An unpublished preview
does not establish live deployment: the public snapshot's `PUBLICATION.json`
binds its build and release, while the separately retained deployment receipt
records live verification. The two internal current-state authorities are
`CURRENT_WORK_STATE.md` and `analysis/second_pass/PIPELINE_STATE.json`;
`analysis/second_pass/CONTENT_STATISTICAL_OPEN_LOOPS.md` records the substantive
backlog.

**Frozen data and current displays are explicitly joined.** The fixed
17-package correction cycle has completed reconciliation, coding, fitting and
its prescribed verification (§5.4). V7 supplies the original numerical inputs,
the later two-event chronology metadata and results as an explicit addendum,
the current reviewed family table, and the actual 370-case display profiles.
Source-qualified display restrictions remain documented; the numeric trait
projection alone is not the final display for every case. Nothing mutates v6
or disguises the original diagnostic run as using later metadata. Six earlier
F8 raw-coding follow-ups remain outside that fixed cycle, explicitly listed
with the successor card work.
Authoring holds and unresolved successor events remain held; the
four family holds are not a count of all workflow holds. Robustness
results below belong to their named releases, not to unintegrated corrections
or held events.

**Public working-draft authorization (2026-08-24).** The author approved an
HTML publication of this in-progress manuscript and a checksum-bound download
of the then-active v4 release data. The public page carries a conspicuous status
banner and links to the exact Markdown source. The 2026-08-27 refresh replaced
that data package with promoted v5. Publication build `20260828z` first served
v6 alongside an atlas of 313 authored cases and 323 playback stops; those are
historical build counts, not the current local inventory. Copyrighted source
captures, scratch files, and raw model/coder journals remain excluded from the
public bundle.

**Event-unit fairness requirement added 2026-08-23.** Monograph v2 must make
explicit that comparable events are not equal-duration time boxes. It must
report a release-wide boundary audit and alternative-resolution sensitivity
tests before claiming that one-event/one-row treatment is consistently applied
across historical tempos and archive densities. The frozen prospective method
and audit schema are in
`analysis/second_pass/EVENT_UNIT_FAIRNESS_AND_RESOLUTION_PROTOCOL.md` and
`analysis/second_pass/event_unit_fairness_audit_template.csv`; the proportionate
execution and stop rule are in
`analysis/second_pass/event_unit_fairness_v1/STREAMLINED_EXECUTION_PLAN.md`.
The working subsection below records the argument and the completed
release-level diagnostics. Coverage completeness is not claimed. Release
activation of v6 is complete under its qualified decision; the mechanism/family
and public-card decisions described for that release do not imply promotion of
the later successor.

**Methods clarification (7–8 September 2026).**
The successor distinguishes an event's calendar envelope,
the precision of its dates, the tempo of the processes within it, and the
observation horizon of each coded outcome. None substitutes for the others.
The older weighting experiment described in §5 used recorded start/end bounds;
it did not measure travel speed or the duration of forced movement. Its name
is clarified below without changing its inputs or results. For the successor,
the existing influx, demographic-transformation, and displacement-tempo
indicators remain separate. No independently bound continuous movement-duration
series is available, so that additional diagnostic is unavailable rather than
estimated by subtracting dates. A missing date-proxy flag is not proof of exact
chronology; inherited nominal dates remain qualified. Dimension-specific
observation horizons follow the codebook and accepted case exceptions, not one
uniform window or later history borrowed indiscriminately into an earlier
event. The complete sensitivity inputs and model runs for the 676-event
successor are now verified and accepted with qualifications. Section 5.4
reports the results, including substantial sensitivity to linked-program
weighting. The release supplies the explicit later chronology addendum. This clarification
supplies the promised reasoning for v2; it does not establish that the corpus
has passed a universal fairness test.

An early 100-row corpus-only check opened no source passages and was
disqualified rather than counted as independent historical verification. The
replacement review opened evidence for all 333 preregistered event checks and
245 relationships; §2.2.1 records the completed process and its limits.

**Historical event-unit candidate-model checkpoint (2026-08-27; corrected and
promoted as v5).** The first
583/679 assembly is preserved as a superseded diagnostic. Its builder failed
to consume three source-resolved territorial-scope exclusions already
documented elsewhere in the draft and review record:
Tibetan residential-school separation, the Swiss Pro Juventute removal of
Jenisch children, and the Svalbard evacuation. The corrected internal
arithmetic is `583 - 3 = 580` for the parent reference and
`580 - 61 + 157 = 676` for the successor candidate. Falklands/Malvinas remains
included as the positive land-taking comparator. The candidate was first
materialized without changing v4 and was then promoted as immutable v5
without altering the prior release. The corrected
model battery reproduced nine factors under every one-event deletion and in
twenty independent Horn batches; every nine-dimensional leave-one-event
subspace angle remained below five degrees (maximum 4.808). Whole-program,
date-proxy, weighting, imputation, documentation, and source-resolution
alternatives remain material. The result rules out single-row domination; it
does not approve every event boundary or settle the factor count. The later
Editorial Suite and explicit `PROMOTE_WITH_QUALIFICATIONS` decision authorize
the release while preserving those limits.

**Historical mechanism/family checkpoint (2026-08-27; terminal for v5).** The
audit-corrected candidate-676 model battery found substantial association but
remaining distinction between process tags and structural outcome families.
Those estimates remain a historical diagnostic, not a public-label result:
after one bounded source-first repair, the single permitted blinded retest
failed its predeclared reliability gate. Section 6 reports both results and the
stopping rule. Normalized mechanism tags remain internal; primary family and
the evidence-aware trait profile remain the public analytical layers. The
mechanism diagnostic itself grants no atlas or deployment authority.

**Post-v5 correction checkpoint (2026-08-28; promoted as v6).** A
finite post-v5 cycle applied ten source-adjudicated authoring corrections and
compared the resulting model. The candidate retires the broad
Indo-Aryan migration wrapper to an inactive audit row, applies 117 adjudicated
cell changes across nine retained events, corrects two chronology-misleading
Assyrian identifiers through durable aliases, and contains 675 active events.
V5 and the candidate both retain nine factors. Their nine-dimensional spaces
differ by a maximum principal angle of 2.765 degrees; forced eight moves 6.685
degrees and rotates the allocation of three named axes. None of the nine
leave-one-corrected-event-out scenarios crossed the declared materiality
thresholds. This supports a stable nine-factor subspace with
forced-eight axis rotation, not a claim that F2, F7, or F8 vanished. Existing
weighting, missingness, imputation, documentation, and coherent-boundary
sensitivities remain. The candidate passed the final Editorial Suite and the
staged immutable-release validator, and is now active as
`second_pass_v6_675`. Immutable v5 remains preserved.

**Companion site:** *Taking their land: a chronological atlas of settlement and
displacement*, https://latrian.dyndns.org/takingtheirland/

### Change log v0 → v1

**Publication-wording correction (2026-09-05).** Current status now identifies
frozen v6/675 and the local 370-case/380-stop atlas separately from historical
v5/676 and older atlas builds. Dated results and corrections remain intact.
The introduction and conclusion foreground documented recurrence within a
selected-positive-case corpus, not universal human inevitability or propensity.
The historical 407-event report and its template now carry a superseded notice
linking the working monograph and data downloads. No research values, licensing
terms, holds, or release authority changed.

1. Factor-count claims in the original v1 correction: eight factors at 407 and
   490 events; **nine** in the then-current v3 583-event roster by a +0.019
   margin, characterized in §5.2; the
   forced-eight 407→583 congruence range (0.653–0.958) replaces the
   transplanted 407→490 range.
2. "No natural species" demoted from finding to explicitly editorial metaphor;
   the defensible result is stated as "no tested automatic partition is robust
   enough to serve as the public taxonomy."
3. §4 reframed as an *agreement-tested* protocol: cross-model N/A concordance
   (0.50 over 96 cells) reported; missing raw validation inputs disclosed;
   unsupported attenuation claim replaced.
4. Bibliography claims reduced to what the saved generator reproduces:
   unique **URL records** (845 at the audited checkpoint; 854 on that
   regeneration: +7 audit-register sources from the Sobaipuri/Salinas
   admissions, +2 from the recode-file harvest added in the second review),
   with known defects enumerated; the unreproducible ~196-locus count
   retired.
5. Transcript-loss claim reversed: five surviving session files are
   inventoried; the bounded process-fact recovery pass is complete.
6. At the original v1 checkpoint, current-state text was updated to atlas build
   `20260821a` (222 cases, 285 phase stops). The current public state is given
   above rather than silently rewriting that dated change-log fact.
7. Cohen 2013 re-placed (Hellenistic east, not Roman Italy); Seymour 2012
   subtitle corrected (*Migrations*).

**Second-review corrections (2026-08-21, same day):** gap-scan arithmetic
fixed (33 confirmed = 30 initial + 3 late; 12, not 15, remain unverified);
raw validation journals recovered and copied into
`analysis/validation/raw/` with both comparisons recomputed exactly from the
in-project copies, resolving former limitation 3; census generator patched to
harvest recode files (854 unique URL records); "well-documented modern cases
accumulate" corrected to composition/resolution language (the 490→583 delta
was 95 ancient Roman cases); "orthogonal" softened to "largely distinct
axis" (oblique factors correlate); "published scholarship" broadened to
"published sources."

**Post-review completions (2026-08-21):** the three remaining work items
outside the author's own questions are done. (a) *Transcript recovery*: the
bounded pass over the five Codex sessions is complete — six newly recorded
process facts, twelve decision-provenance confirmations with locators, and
three negative findings, in
`analysis/second_pass/PROCESS_FACT_RECOVERY.md`; §4.5 updated. (b)
*Citation resolution*: the actual 40-entry v1 list is fully verified — 36
correct without change, 4 corrected, 0 unverifiable
(`monograph/CITATION_RESOLUTION.md`). The original two-agent split covered 39;
a closure audit caught and independently checked the omitted Ziems 2024
entry. (c) *Work-level bibliography*: `build_work_level_bibliography.py` maps
the 854 URL records to 852 conservative pre-review groups. All 14 near-title
pairs are adjudicated (12 different works, 1 same work, 1 part-of), producing
851 reviewed work groups (544 titled works, 155 primary loci, 122 DOI-keyed,
30 modern slot records). Typed publication metadata remains a research pass,
not a script. The review also corrected a Habsburg source attribution from
Alexander Szakolczai to Bogdan G. Popescu; see
`monograph/BIBLIOGRAPHY_REVIEW.md`.

---

### Author decisions — all resolved 2026-08-21

1. **Authorship:** Craig Talbert, sole author; AI systems credited as
   instruments in the contribution statement, not as co-authors.
2. **Citation style:** Chicago author–date, confirmed.
3. **Cunningham's Law:** named in §8 (the framing is the author's own, per
   the recovered decision stream).
4. **Bibliography form:** the citation-resolved 40-entry anchor list stays in
   the monograph; one later source-audited case citation (Malhi et al. 2008) is
   added rather than left dangling. The reviewed 851-work inventory publishes
   separately as a versioned dataset with a methods preface.

Closed earlier: the six working family names were confirmed in the recovered
decision stream; the 40-entry citation pass and four corrections are
verified (the defensible Belich title-page form is retained); all 14
bibliography relationships are adjudicated; and the exact unlinked report
URL and historical deployment checksum are recorded in
`analysis/second_pass/PROCESS_FACT_RECOVERY.md`.

**Remaining production work (no author decision blocking):** typed
bibliographic metadata for the 851-work dataset and an archival deposit before
formal circulation. Version control is established; current local publication
status is stated in the opening snapshot, separately from live verification.

---

## Abstract

Comparative discussion of settler colonialism, ethnic cleansing, and forced
displacement relies on categorical vocabularies — "settler colonialism,"
"internal colonization," "ethnic cleansing" — whose boundaries were set by
theory and case tradition rather than by systematic comparison. This monograph
describes the construction of an alternative: a worldwide register of
land-taking and removal events (676 modeled events in release v7;
370 authored cases in the companion atlas),
each coded on
sixty ordinal dimensions drawn without merging from five disciplinary
traditions, with every event researched against published sources. Coding
was performed by large-language-model research agents under an
**agreement-tested protocol with known cross-model disagreement**: a two-coder
pilot reached 93% exact agreement over 345 comparisons; a candidate coder
model was excluded after failing a validation gate (67% exact; 22% concordance
on the load-bearing distinction between "unknown" and "structurally
inapplicable"); cross-model agreement on a blind 30-event overlap was 76%
exact and 95% within one anchor step, but only 50% concordant on the
unknown-versus-inapplicable boundary. The sixty dimensions contain a
**persistent multidimensional trait structure whose exact factor count is
boundary-sensitive**: parallel analysis retains eight factors at 407 and 490
events, retained nine in the earlier v3 583-event release by a small margin
(+0.019), and returned to eight in v4's 583-event release with the ninth Horn
margin just below zero (−0.00789). The historical v5 676-event
release retained nine in twenty repeated Horn calibrations and
under every single-event deletion, but returned to eight under several
historically plausible grouping and date-proxy alternatives. Frozen v6 retains
nine, with its nine-dimensional space 2.765 degrees from v5; the exhaustive
676-event deletion battery is a v5 result, not a new v6 or v7 test. V7's
canonical fit also retains nine, but its nine-dimensional space moves 11.78
degrees from the recomputed v6 baseline and its eight-dimensional space moves
18.69 degrees. The 565 related diagnostic records include 524 nine-factor,
29 eight-factor and 12 seven-factor results; the chronology addendum preserves
those counts. These are not independent replications or a fairness verdict. The boundary
pattern is therefore not a clean additional public trait. Clustering on the
trait space yields no
tested automatic partition robust enough to serve as public categories:
partitions are shallow, near-tied across hundreds of restarts, and sensitive
to archive density — replacing one unusually granular source archive (the
early Neo-Assyrian royal inscriptions) with resolution-matched medians moves
cluster boundaries substantially while the factor structure barely shifts. We
describe the editorial consequences implemented in the atlas — a manually
reviewed six-family filing system, per-case eight-trait evidence-aware
profiles that summarize coded judgments rather than latent factors, versioned
immutable releases, case retirements with preserved audit rows, and a public
correction channel. A normalized mechanism vocabulary failed its single
permitted blinded reliability retest and therefore remains internal. We argue
that the resulting measured-traits-plus-reviewed-families architecture is a
replicable template for large-scale comparative history built with AI
assistance under human editorial control. This selected-positive-case corpus
documents recurrence and supports comparison; it cannot establish universal
human inevitability or propensity to take land or remove people.

**Keywords:** settler colonialism; forced displacement; ethnic cleansing;
comparative history; event data; exploratory factor analysis; cluster
stability; large language models; AI-assisted coding; digital history.

---

## 1. Introduction

Some historical processes move people onto land; others move people off it;
many do both at once. The scholarly vocabularies for these processes grew up
separately — settler-colonial studies around the permanence of incoming
populations (Wolfe 2006; Veracini 2010), genocide and forced-migration studies
around destruction and flight (Lemkin 1944; Naimark 2001; Mann 2005),
political economy around land and labor regimes (Nieboer 1900; Domar 1970;
Acemoglu, Johnson, and Robinson 2001), historical demography and
archaeogenetics around measurable population turnover (Haak et al. 2015; Reich
2018), and the study of state formation around sovereignty, legibility, and
borders (Scott 1998; Tilly 1990; Maier 2016). Each vocabulary illuminates the
cases it was built on and strains on its neighbors'. The result, familiar to
any comparative reader, is a taxonomy problem: categories that behave less
like discoveries about the world than like the disciplinary histories of the
people who coined them.

The central historical claim is recurrence: across widely separated times and
places, documented land-taking and forced removal recur in recognizable
combinations of power, movement, exclusion, settlement, and institutional
change. Comparing those combinations reveals recurring processes and
consequential differences. This selected-positive-case register has no
representative denominator of all human encounters or opportunities for taking;
it cannot establish a universal human propensity, inevitability, or the rarity
of peaceful alternatives.

This monograph accompanies *Taking their land*, a public chronological atlas
of settlement and displacement from roughly 6500 BCE to the present. The atlas
began, as such projects usually do, with a hand-built category scheme — twelve
mechanism families assigned case by case. It now rests on something different,
and the difference is the subject of this paper. Rather than adopting any one
discipline's framework, or negotiating a compromise between them, we treated
the taxonomy itself as an empirical question: **if every event is measured on
every tradition's dimensions at once, what structure actually emerges?**

Three features of the resulting study are worth stating at the outset, because
each is a deliberate methodological commitment:

**First, the unit of analysis is the bounded event, not the national story.**
A "unique historical event" is one linked process with substantially
continuous actors, mechanism, geography, and chronology; processes are split
when a regime changes, a mechanism changes materially, or a long interruption
makes continuity misleading (§2). Much of the project's labor — and several of
its most consequential editorial decisions, including the retirement of
already-published atlas cases — consisted of enforcing this rule against the
gravitational pull of familiar multi-century umbrellas.

**Second, measurement was performed at a scale, and by a means, that requires
disclosure and validation rather than apology.** Every one of the 676 events
in v7 carries sixty anchored codes, including explicit unknown and
structurally inapplicable states; every code was
assigned by an AI research agent working from a written anchor protocol and
per-event source research, inside an agreement-testing regime — pilot
double-coding, a model-selection gate that rejected one candidate model on
measured grounds, and cross-model agreement estimation — described with its
known gaps in §4. We treat "AI-assisted comparative coding under human
editorial control" as a methodological contribution in its own right, with
honest agreement numbers and honest holes attached.

**Third, the analysis was allowed to fail, and the write-up is required to
keep the failures.** The trait structure persisted as the corpus grew; the
exact factor count did not (eight at 407 and 490, nine in v3's 583-event
roster, eight in v4's different 583-event roster, and nine in both the historical
676-event v5, 675-event v6 and 676-event v7 releases under their canonical
preprocessing — §5.2),
and no tested clustering of the trait space survived its own diagnostics as a
public taxonomy (§6). We regard the negative results as equally important,
particularly for the growing genre of computational history that reaches for
cluster labels as if they were facts about the past rather than facts about a
corpus, an encoding, and an algorithm (cf. Hennig 2015 on the
method-dependence of "true" clusters). The process was iterative rather than
a single pre-planned protocol; §5–§6 therefore distinguish explicitly between
what was specified before analysis, what was found exploratorily, and what
was decided editorially afterward.

The monograph is written in two registers. The main text is intended to be
readable by a serious general reader of the atlas; technical material —
estimation choices, stability statistics, sensitivity designs — is confined to
marked subsections and tables, and the complete apparatus (data, scripts,
manifests, hashes) is cited by file throughout (§10, Data availability).

### 1.1 Contributions

1. A worldwide, source-bound register of land-taking and removal events
   (676 modeled in v7), with every row in the audit's declared
   country-and-territory sweep closed either by an event finding or a
   source-backed negative finding,
   and a source inventory of 854 unique URL records at the 2026-08-21 census
   checkpoint (§2.5).
2. A sixty-dimension measurement instrument assembled *without merging* from
   five disciplinary traditions, with anchored ordinal scales, era-coverage
   rules, and an explicit distinction between "unknown" and "structurally
   inapplicable" — the latter itself analytically load-bearing.
3. An **agreement-tested protocol** for AI-agent historical coding: pilot
   agreement 0.93 exact over 345 comparisons; a failed-model exclusion gate;
   cross-model agreement 0.76 exact / 0.95 within-1 on a blind 30-event
   overlap, with a disclosed weak point (0.50 N/A concordance). The raw inputs
   that were missing at the first audit were recovered and both comparisons
   now reproduce (§4.4).
4. A persistent multidimensional trait space with a documented
   **boundary-sensitive factor count** (eight at 407 and 490; nine in the v3
   583-event roster by +0.019; eight in the v4 583-event roster; and
   nine in the historical v5 676-event, v6 675-event and v7 676-event releases
   under their canonical preprocessing), a documented era confound on one axis,
   and a documented face-validity correction to another.
5. A demonstrated failure of automatic taxonomy *as tested*: shallow,
   near-tied partitions (367 distinct solutions in 500 restarts at 490
   events) and a controlled archive-density experiment showing that one dense
   source archive relocates cluster boundaries in this corpus.
6. An editorial architecture that metabolizes (1)–(5) into a public artifact:
   reviewed families as filing (not discovery), evidence-aware trait profiles
   summarizing coded judgments (not latent factors), versioned immutable
   releases, case retirement with preserved audit rows, and an in-atlas
   correction channel.

---

## 2. The register

### 2.1 From atlas audit to worldwide register

The register began as an audit of the atlas's own coverage. In August 2026 the
then-current atlas contained 106 cases in 119 display phases. The audit posed
five questions for every one of 196 sovereign states and 58 additional
territories and cross-border homelands — who displaced, who was displaced,
whether the atlas captured each process, what was missing, and what related
processes fell outside the atlas's threshold — and closed all 254 geographic
reviews, six of them with source-backed negative findings (Andorra,
Liechtenstein, San Marino, Vatican City, American Samoa, Wallis and Futuna)
rather than silence. The audit's deduplicated master register contained 364
events with 1,559 country-event incidences; its event-unit test found that
only 36 of the 106 atlas cases survived unchanged, with 37 incomplete, 27
wrongly bundled, and 6 wrongly classified. Applying the audit produced a
222-case, 283-phase atlas build. (Register, methodology, and per-country
findings: `audit/`, esp. `MASTER_EVENT_REGISTER.md`, `METHODOLOGY.md`,
`EXECUTIVE_FINDINGS.md`.)

Two commitments from the audit govern everything downstream. **Modern states
are geographic containers, not actors:** an "agent" is a historically named
state, institution, army, or population stream, never a present-day
population; counts never imply collective guilt. **Categories describe
mechanisms, not moral severity:** the register's twelve mechanism families
(settler colonization; conquest with settlement; expulsion/deportation/ethnic
cleansing; demographic expansion; state resettlement/internal colonization;
conflict displacement; genocide-linked clearance; enslavement/labor removal;
development/conservation displacement; population exchange/partition; conquest
without substantial settlement; return/decolonizing displacement) rank nothing.

### 2.2 The event-unit rule

The register's single most consequential instrument is the event-unit rule
quoted above (§1). Its effect is easiest to see in what it dismantled. The
audit split the atlas case "Anatolian Neolithic" into regionally distinct
Neolithic expansion events; the second research pass (§2.4) split a single c.
900–612 BCE "Neo-Assyrian deportations" umbrella into eleven bounded
successors, then added forty-four separately bounded earlier episodes
(sixteen under Ashurnasirpal II, twenty-four under Shalmaneser III, four under
Šamši-Adad V) from a seventy-five-record ruler-by-ruler ledger of the royal
inscriptions (Oded 1979; project ledgers
`EARLY_NEO_ASSYRIAN_SOURCE_REVIEW.md`, `_EVENT_UNITS.md`); and the third
statistical release retired two published Roman wrapper cases
(`med-roman-republican-colonies-italy`, `med-roman-po-valley-colonization`)
in favor of ninety-five bounded colonization events assembled from a
340+-disposition Greco-Roman source ledger anchored in the ancient narrative
and antiquarian sources (Livy, Dionysius, Appian, Velleius; Salmon 1969). The
individual classical passages and editions are resolved in that source ledger;
the four author names are not claims of one-to-one entries in the selected
modern reference list. The
same rule, applied in the other direction, retired the atlas's broad
"Southern Athabaskan migration" case: a well-evidenced four-century migration
and ethnogenesis that never resolves into one demonstrated displacement
event. Its bounded successors — the 1762 Sobaipuri-O'odham relocation, the
1667–1672 Las Humanas abandonment, and the 1676–1677
Quarai/Chililí-to-Tajique relocation — passed candidate review and are admitted,
map-authored v4 cases with completed sixty-dimension coding. This is a
deliberately public example of the rule outranking a published case
(`analysis/second_pass/CURRENT_ATLAS_ATHABASKAN_REVIEW.md`,
`SOUTHWEST_BOUNDED_SUCCESSOR_ADMISSION.md`).

#### 2.2.1 Event-unit fairness across historical tempos

Equal statistical weight is meaningful only if the rows have been bounded by
equivalent rules. It does
not follow that they should occupy equal amounts of calendar time. A population
transition visible only as a broad archaeological horizon around 6000 BCE, a
campaign reported season by season in 338 BCE, and a removal documented day by
day in 2014 differ in historical tempo and in the resolving power of their
sources. Dividing all three into fixed year windows would reward modern
documentation and manufacture comparability rather than achieve it.

The project's unit is therefore causal and territorial, not merely
chronometric. Chronology is one test alongside continuity of actors or program,
affected population and land, mechanism, causal sequence, geography, and
outcome. A long interval prompts review but does not force a split; actions
close in time may still be different events. A proposed child must have its own
bounded coercive act or land transfer, affected population or tract, evidence
for an independent analytical profile, and an explicit supersession rule. A
removal, successor settlement, and later return may remain phases of one event
when they implement one linked transformation; several map stops never become
several statistical rows merely because the animation needs them.

The source-qualified boundary review supplies several useful comparisons. The
Habsburg Military Frontier remains one 1522–1881 event because a named
land-for-service military institution links its changing districts and
populations. By contrast, the much shorter Balkan Wars, Croatian War, and
Bosnian War wrappers do not become single events merely because each occupies
one war interval: their source bodies identify independently organized
programs, target populations, directions, and territorial purposes. The
1991–2008 South Ossetia wrapper likewise splits across the 1992 settlement and
the newly constituted 2008 war, while the four-day Osh outbreak remains one
event because Jalal-Abad's violence followed Osh within a contiguous regional
mobilization and displacement episode. These paired decisions show why neither
long duration nor short duration determines the unit.

A later Melos–Veii–Circeii authoring review makes the weighting consequence
especially concrete. Veii remains one 396–387 BCE row because conquest,
captive sale, household land assignment, selective citizenship, and tribal
reorganization form one territorial program, even though its phases and
cohorts must remain visible. Melos remains two oppositely directed events in
416 and 405 BCE because authority, affected cohort, direction, coercive act,
and land outcome reverse; the broad parent remains inactive. Replacing those
two Melos children with the parent rotates the forced-eight subspace by 10.81
degrees but the corrected nine-factor subspace by 2.92 degrees, while omitting
the restoration alone moves them by only about 1.33 and 1.25 degrees. The
source-first split therefore survives, but it is not presented as
weight-neutral. Circeii supplies a different boundary: one named 393 BCE
colony passes the declared durable-taking/use rule without proof of resident
expulsion, while an unnamed 395 authorization and a source-conflicted earlier
removal remain outside the event. These are applications of one
causal-territorial rule, not equal-duration boxes or archive-density quotas
(`ATLAS_AUTHORING_BATCH_33_ADJUDICATION.md`).

Thebes supplies a second material-sensitivity example for the final v2
fairness account. Alexander's 335 BCE destruction and Cassander's 316 BCE
restoration remain separate because authority, moving cohort, direction, and
land outcome reverse after nineteen years; the primary sources independently
bound both actions, and the inactive parent cannot coexist with both
successors. Yet replacing the successors with that parent rotates the forced-
eight subspace by about 5.60 degrees, while leaving the program out rotates it
about 5.02 degrees. The corresponding corrected nine-factor movements are
about 2.46 degrees, and nine factors remain retained. The split is therefore
historically defensible but not weight-neutral. Reporting both results is the
fairness obligation: the statistical consequence cannot be hidden, and it
cannot be allowed to collapse two opposed historical processes merely to make
the model move less (`ATLAS_AUTHORING_BATCH_37_ADJUDICATION.md`).

The same test works in the other direction. Darfur can remain a long phased
event only after its scope is narrowed to the documented
government/Janjaweed/Border Guards/RSF demographic-clearance program against
non-Arab communities, including land occupation and obstruction of return.
Unrelated SAF–RSF combat, rebel abuses, inter-Arab violence, and general Darfur
displacement are outside that row. Eastern Congo and Somalia, despite equally
persistent regional crises, require splits because their actors, mechanisms,
affected populations, causal systems, and territorial programs vary
independently. Liberia splits across the 1997 political settlement, return
episode, and changed actor roles. The Highland Clearances illustrates the
fail-closed alternative: evidence can disprove one national score profile
without yet supplying a complete admitted successor set, so the wrapper stays
on boundary hold rather than being counted or prematurely retired as a parent.
These dispositions show the rule in operation but do not themselves authorize
a release. Darfur's narrowed scope still required a compatibility check; the
completed candidate counts and sensitivity results are reported below.

Project continuity, rather than calendar span, produces the same distinction
elsewhere. The Three Gorges resettlement remains one event despite long, staged,
multijurisdictional implementation because one state project, approved plan,
inundation zone, funding system, and irreversible land outcome link its
movements. Canada's 1953 and 1955 High Arctic relocations likewise remain one
kin-linked federal project rather than two events defined by calendar year.
By contrast, the much shorter 1978–1990 Nicaragua wrapper splits at the July
1979 regime change: most of the earlier exile cohort returned, new cohorts
departed, authorities and directions changed, and the later Contra cycle
included independently profileable Miskito relocation. The Australia-wide
Stolen Generations category passes the atlas's territorial-scope threshold
because the national inquiry documents severance from Country and impairment
of land and native-title connections, but it still fails as one statistical
unit because distinct jurisdictions used different laws, authorities,
procedures, placements, and timelines. Scope qualification, event coherence,
calendar duration, and geographic size are therefore separate judgments.
The same territorial threshold excludes both Tibetan residential-school
separation and the Swiss Pro Juventute removal of Jenisch children from active
counting: each is a coherent or recognizable assimilation history, but the
opened evidence establishes family separation and institutional placement
rather than a territorial land outcome. Their exclusion and Australia's scope
qualification are different evidentiary results under one rule, not different
rules for different countries.

Wartime cases separate event disaggregation from scope admission.
Saint-Pierre's British removals in 1778 and 1794 are distinct
events because restoration and actual resident return reversed the first
outcome before a new war produced the second. The Isle of Man's 1914–1919 and
1940–1945 internment systems likewise split after complete release, restoration
of civilian premises, and a new activation involving materially different
cohorts. But splitting the Channel Islands wrapper does not make its every
component admissible: the substantially voluntary 1940 protective evacuation
is contextual history, whereas Nazi deportation and internment must pass the
territorial-removal and overlap gates independently. Svalbard supplies the
harder comparator. Its evacuation was compulsory, but it remains outside scope
because it was a temporary wartime denial measure without a successor
population or durable land transfer, followed by postwar re-establishment.
Shared war, shared place, short duration, and even coercion are therefore not
stand-alone boundary or scope rules.

Prehistoric and ancient cases expose the converse risk: turning an evidence
limit into an unannounced narrowing of scope. A
source sidecar correctly found that Bell Beaker ancestry turnover in Britain
does not establish violent invasion, forced removal, or coercive land
acquisition, but initially treated that absence as disqualifying. Main
adjudication returned to the frozen inclusion rule. A non-state demographic
expansion may qualify through directly supported substantial replacement or
absorption without a claim of violence. Bell Beaker Britain therefore remains
an active expansion analogue, narrowed to the directly supported c. 2450–2000
BCE migration and turnover interval; unresolved social mechanism remains
unknown. This does not rescue every ancestry horizon. The combined Lower
Danube/Carpathian Yamnaya, Britain/Ireland Neolithic, and continent-scale
Corded Ware rows remain held because their evidence does not establish one
coherent affected land and population, causal process, or independent profile.
Scope admission, mechanism certainty, and event coherence are three separate
questions.

Ancient colonial history poses the same event-unit test. Greek Sicily and
southern Italy require source-bounded local
successors because independent founders, peoples, tracts, and mechanisms do
not share one residual profile. The Massalia–Emporion network becomes an
inactive parent because real founding and trading connections do not erase
separate territorial institutions. The Black Sea and Cyrenaica rows fail
closed on evidence holds rather than inheriting continuity from a regional or
dynastic label. The already inactive Han Hexi/Western Regions parent is
confirmed for the same reason: dynasty-wide continuity cannot merge distinct
territories and frontier institutions. These judgments enter the verified
register and candidate tests below; they do not individually establish a
release-wide fairness result.

The Roman comparison shows what is sufficient to retire a wrapper without
pretending that an archive is complete in an absolute sense.
The two broad Republican Italy and Po Valley rows are not confirmed from a
general chapter title or a long child list. They remain inactive audit rows
because a locator-bearing source program dispositions all sixty-seven
canonical parent components as bounded successors, nested supports, explicit
holds, non-events, or separately reviewed later processes; a ninety-six-unit
parent map reconciles the shared children; and a fail-closed validator prevents
either parent from coexisting with its replacements. Eleven uncertain
foundations remain visible as holds rather than being smuggled into a residual
Roman event. Source richness therefore increases the resolution at which
claims can be tested, but it does not entitle Rome to extra statistical weight.

Coherent demographic corridors also differ from mere regional periods. The
coastal Tupi and Paraná–Plata Guaraní expansions
remain separate active units because their source bodies identify distinct
settlement streams and connected corridors despite long duration and locally
mixed mechanisms. The 1964–present Amazon highway-frontier label and the
Highveld Difaqane period remain boundary holds because each aggregates
different actors, projects, directions, and land outcomes without a complete
successor set. Ancient or modern chronology does not decide any of those four
outcomes; source-supported continuity and independent profileability do.

Cases outside the ancient Mediterranean test scope, scale, and historical
tempo directly. An uninhabited-island settlement and a
post-depopulation occupation cannot be treated alike merely because both lack
a resident population at the later settlement date. Bermuda and Cayman remain
outside scope because permanent settlement began without a local prior-
population land transition. The Bahamas/Turks umbrella remains on hold because
Spanish destruction of Lucayan communities and much later British repopulation
are distinct processes whose separation does not yet yield a complete child
set. The rule turns on the documented land and population history, not on an
island label or a mechanical test of who was present on one date.

The Australian and New Zealand comparisons show that geographic breadth does
not itself require either retention or splitting. Western Australia's
southwest, Goldfields, Kimberley, and pastoral frontiers remain one explicitly
mixed, multiregional event because successive colonial and state institutions
connect settlement, policing, labor control, reserves, and missions without a
documented reversal. The Queensland-and-wider-north label remains a hold:
similar frontier violence appears across it, but one implementing chain and a
complete disjoint successor map do not. On the other side of the Tasman, a
broad Company-plus-Crown Aotearoa umbrella remains held while the Ngāi Tahu
South Island process remains active. Multiple deeds do not force a split when
the affected people, land and resource relationship, purchase machinery,
Crown responsibility, and continuing grievance supply a coherent profile.
Direct land transfer also does not need lethal coercion to qualify.

Algeria and Egypt provide an especially close modern-tempo comparator. The
1961–1963 departure of Algerian Jews remains one event because the affected
population, predominant Algeria-to-France route, French-citizenship
transition, and concentrated decolonization setting cohere; this does not make
every departure individually coerced. Egypt's 1948–1967 umbrella is revised
to a required split. The 1948–1951 wartime internment, sequestration, violence,
and departure were followed by release and partial normalization; a different
regime imposed the 1956–1957 Suez-era expulsions and seizures before partial
rescission; and 1967 produced a new wartime detention episode. A shared
population and broad Arab–Israeli conflict cannot erase reversals, changing
programs, and independently profileable mechanisms. The contrast demonstrates
why short or modern chronology receives no automatic preference: continuity
retains Algeria, while discontinuity splits Egypt.

The Iran-through-Tunisia comparison shows that boundary correction and claim
calibration are not the same operation. Iran and Iraq remain coherent
active units, but direct evidence of small, non-reversing returns requires
removing absolute “no return” language. That changes the description of the
outcome without manufacturing a new event. Libya remains a required split
because a completed 1949–1952 mass departure left a separately identifiable
remnant that was later forcibly evacuated; Morocco remains one unit because
recurring organizations and overlapping cohorts connect the phases without a
comparable first-wave closure. Lebanon remains held because the accessible
evidence neither joins its crisis phases nor bounds complete children. Syria
newly fails closed because the opened passages cover only the late 1987–1993
phase of an asserted 1948–1994 event. These outcomes separate three questions
that can otherwise be conflated: whether one event exists, whether its internal
claim is phrased accurately, and whether the sources span its whole asserted
boundary.

This creates two separate fairness obligations. **Boundary fairness** asks
whether each row is one comparably reasoned event and whether any history is
counted twice. **Measurement fairness** asks whether its sixty traits were
scored with evidence and anchors appropriate to its era, using unknown rather
than absence when the record cannot answer. Neither substitutes for the other.

This rule must also operate symmetrically across archives. Rich Roman or
Neo-Assyrian records may reveal real local events, but named places and dated
inscriptions are not automatic rows. Sparse prehistoric evidence may justify
broader uncertainty, but it cannot justify merging demonstrably different
processes or converting silence into continuity. Holds and unknown values are
the proper response when the evidence cannot support a stable boundary.

The existing register supplies case law and partial diagnostics: retired
wrappers, overlap crosswalks, inactive parent rows, era-specific coding anchors,
era reweighting, grouped Roman and Hellenistic alternatives, and the
Neo-Assyrian archive-density experiment. The reproducible fairness intake
contains 689 review units: the 592-row v4 event/parent intake plus 97 corrected
Hellenistic units, of which 92 were model-eligible at intake and five retained
for audit only. The earlier 334-cell missingness review and 500-cell fresh
double-coding are complete; 374 fresh cells agreed exactly and all 126
disagreements received one evidence-grounded, no-averaging adjudication. That
5,820-cell Hellenistic ledger remains a completed input, not a substitute for
checking the boundaries and scores of the rest of the roster.

The release-wide boundary process is now substantially further along. Two
model-blind calibration reviewers first supplied independent dispositions for
the deliberately difficult calibration set; one no-averaging reconciliation
resolved their actual disagreements. A separate full-roster pass supplied the
remaining rows. Independent source verification then opened evidence for all
333 preregistered event checks and all 245 relationships, rather than treating
a structurally complete spreadsheet as historical verification. Event outcomes
were 225 confirmations, eighty revisions, and 28 holds for main adjudication;
relationship outcomes were 210 confirmations, nineteen revisions, fourteen
holds, and two removals. The verifier recorded 147 findings: 108 event, 35
relationship, and four sample-coverage findings.

One main adjudication reviewed each of those 147 findings once. It accepted 99
bounded verifier revisions and two relationship removals, preserved 42
evidence holds, and recorded four inaccessible scale observations as explicit
sample limitations. The resulting non-destructive register preserves every
one of the 689 original rows and 245 relationships as an audit-visible prefix.
Its effective event states are 518 active retains, 85 boundary holds, 65
required splits, nine inactive audit rows, eight retired active parents, and
four scope exclusions. The 356 ordinary retains outside the preregistered
source-check sample are labeled as not selected by design rather than being
misdescribed as source-verified. Structural validators reproduce the register
and confirm that no prospective target or downstream publication authority was
silently activated.

The successor gate was kept separate from measurement. It resolved one
normalized 233-target admission denominator without treating source support as
automatic admission. Of 158 initially eligible targets, 157 cleared the
source minimum and one overbroad Somalia target returned to hold. The 157
cleared successors then received two independent sixty-dimension codings and
one disagreement-only, no-averaging adjudication, resolving all 9,420 cells.

Thirty accepted boundary findings also affected events remaining active.
Seventeen complete profiles passed exact-reuse review; thirteen changed events
required 364 freshly coded cells. Their source-complete packets contained 43
source records, and the same two-coder plus disagreement-only procedure
resolved every cell. An earlier score-blind pilot lacked consistently openable
source bodies and precise locators; its appropriately unknown-heavy returns are
preserved as a diagnostic and excluded from measurement. The replacement lane
required each non-unknown answer to cite an opened dossier source identifier.

The completed v5 candidate builder applied these dispositions without mutating
v4. After the three scope exclusions, it created a 580-event parent reference
and a 676-event successor candidate, with all 69 connected histories retained
as alternative-resolution groups. Activation required the separate publication
decision recorded above.

The source requirement was added after an early process failure: a verifier
returned 100 mechanically complete rows but later disclosed that it had opened
zero underlying works or passages. The project preserved and disqualified that
return rather than treating spreadsheet completeness as historical
independence. The replacement run records, for every checked event, the work
actually opened, an exact passage locator, an evidence grade, and the
source-access state; its completed counts and main dispositions are reported
above. Its first checked case remains a useful illustration. Malhi et al.
(2008) support a small proto-Apachean movement into the Southwest and extensive
admixture, but the opened passages do not establish prior-population
displacement, territorial replacement, or the unity of a c. 1300–1700 event.
The source verifier therefore held the broad Athabaskan row, and main preserved
it as an inactive evidence hold rather than converting evidence of migration
into evidence of one displacement event.

**Interim corrected-tranche diagnostic (2026-08-25; not a release result).**
Adding the 92 corrected model-eligible Hellenistic events yields nine factors
under equal-event weighting, whether the four overlapping Greek-colonization
parents are retained (675 events) or removed (671). Collapsing linked programs
still yields nine; collapsing the entire added tranche to one median restores
eight. Across declared weighting and imputation choices, the retained count
spans seven to nine. The new events are evidence-sparse on this instrument
(median 34.5 of 60 cells missing), and no single event is materially
influential under the preregistered leave-one-out thresholds. The defensible
interim inference is therefore collective sensitivity to event resolution,
corpus composition, and missing-data treatment—not a newly settled ninth
trait. The complete 689-row boundary application and verification remain
necessary before this can become a release-wide fairness claim.

**Historical candidate-boundary diagnostic (2026-08-27; not yet activated at
that checkpoint, subsequently promoted as v5).** The explicit candidate builder
binds every roster decision, score
source, parent replacement, overlap group, and date proxy. Three
source-resolved scope exclusions reduce the public 583-event roster to a
580-event internal parent reference. The canonical-successor candidate then
retires 61 active parents and adds 157 independently scored successors, for
`580 - 61 + 157 = 676` events. Nine additional parent references remain
audit-visible but were already inactive and are not subtracted twice.
Sixty-nine connected histories retain both parent and successor resolutions
for direct sensitivity tests.

The model battery treats temporal scale as one diagnostic rather than the
event definition. It compares equal events, grouped successor programs,
parent-versus-successor resolutions, era balance, bounded calendar-envelope weighting,
source resolution and missingness strata, four imputation treatments, 69
program swaps, 69 leave-program-out runs, and every one of the 676 possible
single-event omissions. The corrected-parent reference retains eight factors;
the canonical successor retains nine. Grouping each connected program to one
median returns eight; so do the complete-parent alternative and removal of all
39 explicit date proxies. Thus chronology is not the event-unit rule, but
resolution and provisional dating do materially affect the retained count.

The calendar-envelope experiment above (originally labeled
“observed-duration weighting”) used the absolute difference between recorded
operative start and end bounds, plus one year, transformed by a capped
logarithm. Rows flagged as date proxies and rows without both bounds retained
neutral weights. Those bounds are not measurements of migration duration,
transport speed, or a constant tempo throughout the interval. This is a
diagnostic of the recorded chronological span, not a rule for deciding how
many events a history contains. The 2026-09-07 clarification leaves that
historical calculation unchanged.

The most important result concerns the factor space as a whole. Under all 676
single-event omissions, the nine-dimensional subspace has a median maximum
principal angle of 1.036 degrees, a 95th percentile of 2.433 degrees, and a
maximum of 4.808 degrees; no event crosses five degrees. The forced-eight
subspace has a median of 1.624 degrees, a 95th percentile of 13.871 degrees, a
maximum of 86.489 degrees, and 36 omissions above ten degrees. Across 138
whole-program alternatives, the nine-dimensional maximum angle has a median
of 1.810 degrees, a 95th percentile of 4.743 degrees, and a maximum of 5.947;
seven scenarios cross five degrees and none crosses ten. Forced eight has a
median of 3.345 degrees, a 95th percentile of 73.823 degrees, a maximum of
88.802, and 24 scenarios above ten. No tested individual deletion materially
remakes the candidate's nine-dimensional covariance space; coherent program
choices can move it modestly and frequently destabilize a forced-eight
projection. These diagnostics are consistent with rotation or compression,
not proof of one unique cause.

This supports a qualified candidate-level account with three deliberately
separate parts. **Procedural boundary evidence** is strong: active rows and
relevant inactive parents have explicit dispositions, overlap treatment,
evidence trails, and sensitivity groups. **Measurement robustness** is strong
for single-row perturbations of the nine-dimensional subspace but sensitive to
exact factor count, coherent program resolution, date proxies, weighting,
imputation, axis naming, and public-card projection. **Coverage completeness**
is not demonstrated and is not claimed: the register is cross-indexed,
independently gap-tested, and publicly corrigible under a declared scope, not a
proof that every qualifying event has been found. The final post-v5 Editorial
Suite audited 43 substantive changed claim groups and matched all 43, with no
overclaim, underclaim, unsupported, or uncheckable verdict inside its declared
local-source boundary. Its qualified promotion decision authorizes this
bounded fairness wording for v6. The immutable builder and independent release
validator passed before the active pointer changed.

The preceding v5 candidate claims are bound to
`analysis/second_pass/event_unit_fairness_v1/candidate_models_v2/MODEL_RESULTS_AUDIT_DISPOSITION.md`,
`model_results_v2/MODEL_RESULTS_SUMMARY.json`,
`subspace_stability_v2/SUBSPACE_STABILITY_SUMMARY.json`, and
`factor_retention_calibration_v2/FACTOR_RETENTION_SUMMARY.json`. Their
validators and hash manifests accompany each output directory.

The claim sought is deliberately limited: not that historical events have one
objectively correct atomic scale or that boundary bias has been eliminated,
but that the candidate's boundaries, overlap decisions, and sensitivity groups
are published and contestable, and that the measured consequences of declared
alternative resolutions are reported rather than hidden.

### 2.3 Coverage discipline and the gap scan

A register can be honest without being complete, but it must know which it is.
Beyond the audit's country sweep, a six-angle adversarial gap scan
(archaeogenetics; ancient/medieval non-Western; development and conservation
displacement; modern conflict and partition; Oceania and island worlds;
internal colonization in Russia, Asia, and the Americas) proposed forty-five
candidates, each then adversarially verified against the register with
rejection as the default. Thirty were confirmed in the initial verification
and fifteen were held unverified at a strength cap; three of the held
candidates were subsequently verified and confirmed, for **thirty-three
confirmed additions with twelve candidates still recorded as unverified**.
Among the confirmed: the class-based
Soviet dekulakization system (Viola 2007; Polian 2004), which the register's
ethnic-deportation coverage had missed entirely; the Porfirian Yaqui
deportations (Hu-DeHart 1974), the register's first event with Mexico as
perpetrator; the Argentine Chaco conquest (Gordillo 2004), a gap the
register's own Argentina entry had named; the Neo-Assyrian and Neo-Babylonian
deportation systems; Umayyad Iberia; the Danelaw; Ottoman sürgün; and a
cluster of dam, park, and military removals (Kariba, Volta, Narmada, Kaptai,
Serengeti/Ngorongoro, Bwindi, Vieques). The twelve still-unverified
candidates remain recorded as such rather than silently dropped
(`analysis/gapscan/gapscan_results.json`).

The due-diligence protocol formalizes this stance
(`audit/COVERAGE_DUE_DILIGENCE.md`): freeze the inclusion scope before
searching; give every branch and lead an explicit row-level disposition
(advance / existing-event support / duplicate / evidence hold / searched with
no bounded event found / inaccessible source / unresolved); and preserve every
precise fact, contradiction, and negative finding with a document locator,
whether or not it enters atlas prose. "Searched and unsupported" must remain
distinguishable from "never examined." The register does not claim worldwide
saturation; it claims that its blind spots are enumerated.

This is selection for qualifying positive cases, not a representative sample
of taking versus coexistence. More included cases can document recurrence but
cannot supply the missing denominator for universal propensity, inevitability,
relative prevalence, or ultimate causal claims.

### 2.4 The ancient and premodern expansion

The first modeled corpus (407 events) inherited the audit's modern-forward
center of gravity. Because the atlas's editorial objective is to demonstrate
documentable patterns *across millennia*, the second pass ran bounded
source-tranche expansions: nine initial ancient/medieval cases, five
Achaemenid, four early Sasanian, two New Kingdom Egyptian, nine Mongol
western-campaign cases, the Neo-Assyrian replacement and expansion described
above, and finally the ninety-five-event Roman tranche — each tranche passing
a per-event source brief, an event-unit gate, an overlap crosswalk, and full
sixty-dimension coding before activation. (A further Hellenistic-eastern
pre-coding queue, anchored in the standard settlement catalogues — Cohen 2013
for the East — was held outside that release pending the same gates.) The
then-current statistical release, `second_pass_v6_675`, held 675 active events in
an immutable, hash-verified directory. It preserves v5, v4, and the prior v3,
490-event, and 407-event checkpoints. V6 applies the finite post-v5 retirement,
cell, chronology, family, and public-narrative corrections without silently
editing the old release. Two active rows retain explicit
family-classification evidence holds and remain outside atlas rendering in
that release. The subsequent v7 correction cycle is reported separately in §5.4.

### 2.5 The source inventory

Every event carries at least one directly relevant source; recommended atlas
additions carry at least two. The census for this monograph — regenerated by
the saved script `monograph/build_bibliography_census.py` — finds **854
unique URL records** across the audit registry, the legacy atlas registry,
the gap scan, and the second-pass tranche ledgers
(`monograph/bibliography_census.csv`; 845 at the fact-check checkpoint — the
delta is seven audit-register sources added with the Sobaipuri and Salinas
admissions plus two sources from the Apuan/Luna recode file, whose omission
from the generator's harvest was a coverage gap found in the second review). This is a *source inventory*, not yet
a publication bibliography: it deduplicates by normalized URL only, so it
contains generic slot titles from the Roman tranche, primary-source URLs
alongside secondary works, 23 duplicate exact-title groups, and at least one
confirmed same-work-at-two-URLs duplicate. Separately, the Greco-Roman
disposition ledger holds 398 rows with a primary-source locator (classical
texts via the Perseus and Thayer corpora: Livy, Diodorus, Appian, Arrian,
among others), and the early Neo-Assyrian ledger holds 75 rows drawing on 33
distinct royal-inscription source identifiers (pipe-split rule; verified).
A work-level scaffold now exists
(`monograph/build_work_level_bibliography.py`): the 854 URL records resolve
to 852 conservative pre-review groups. All fourteen near-title candidates
have explicit dispositions in `bibliography_review_queue.csv` — twelve
different works, one same work, and one chapter-within-book relationship.
Applying the one reviewed merge produces **851 work groups**: 544 titled
works, 155 primary-source loci, 122 DOI-keyed works, and 30 modern slot
records. The generator preserves decisions on regeneration and never merges
`different` or `part_of` rows. Publication form is decided (2026-08-21): the
anchor list remains in the monograph and the 851-work inventory publishes as
a separate versioned dataset with a methods preface. Typed bibliographic
metadata remains the production step before that dataset release.

---

## 3. The instrument

### 3.1 Five batteries, deliberately unmerged

To measure events without adopting any one framework as the sole authority,
five disciplinary design briefs were commissioned independently — settler-colonial studies;
political economy of land and labor; historical demography and
archaeogenetics; genocide and forced-migration studies; state formation,
sovereignty, and ideology — each producing twelve candidate dimensions with
concrete 0–4 anchors, evidence notes, era-coverage rules, and *a-priori
predictions of the factor structure*. Those predictions were **documented in
a dated local file before the surviving coding outputs**
(`analysis/codebook/a_priori_predictions.md`); they were not preregistered in
the formal sense of a time-stamped, read-only external deposit, and we do not
claim that status for them. The five batteries were then adopted **verbatim
and side by side**: sixty dimensions, no merging, no adjudication of overlap.
Where two disciplines measure nearly the same thing (settler-colonial
"demographic saturation," demographic "incomer share at consolidation"), the
overlap is retained as data — convergence between independently designed
instruments is a result, and hand-collapsing it in advance would have
pre-empted the factor analysis that §5 reports. Retaining overlap is itself a
theoretical design choice: repeated indicators can give some constructs more
influence in the covariance model. [Battery and anchors:
`analysis/codebook/codebook.md`, v1.1.]

### 3.2 Anchored ordinal scales

Each dimension is scored 0–4 against written anchors chosen to be decidable
from evidence in minutes of research, not vibes ("settlers exceeded 50% of the
post-event population within two generations," not "significant settlement").
Anchors carry era-coverage rules: a prehistoric case scores sex-biased
admixture from ancient-DNA evidence (e.g., steppe-ancestry turnovers: Haak et
al. 2015; Olalde et al. 2019) where a modern case scores passenger manifests
and censuses.

### 3.3 Unknown is not inapplicable

The pilot (§4.1) surfaced a distinction that became one of the instrument's
most productive features. `-1` (*unknown*) means the dimension has a real
referent but research cannot settle the score — native mortality in a
prehistoric expansion. `-2` (*structurally inapplicable*) means the referent
does not exist — incoming-settler dimensions for a pure expulsion;
prior-population dimensions for settlement of uninhabited land (Norse
Greenland). The two are different facts about the world, and the pattern of
`-2` codes is itself structural signal: events with no incomers to describe
are a substantial fraction of the corpus, and no continuum connects them to
family-farm colonization. Treating that as "missing data" would have erased
the register's clearest structural boundary. The v3 release types every one
of its 34,980 cells as a numeric judgment, an explicit unknown, or a
structural not-applicable, with zero untyped blanks.

---

## 4. Measurement with AI research agents

At the historical v3 checkpoint, this project's coding labor — sixty dimensions
× 583 events, with per-event source research — would have been impracticable within this single-scholar
project's available time and funding. It was performed by large-language-model
research agents. We describe the protocol in the same spirit as any
instrument: what was tested, what failed, what the uncertainty diagnostics
are, and what cannot currently be recomputed. Two framing cautions govern the
whole section: everything below measures *agreement between coders*, which bounds
reliability but is not evidence of historical validity against an expert gold
standard; and the process was iterative, with the instrument itself amended
once (the `-1`/`-2` rule) after the pilot. (Cf. emerging practice on LLM
annotation in social science: Gilardi, Alizadeh, and Kubli 2023; Ziems et al.
2024.)

### 4.1 Pilot

Two independent agent coders (separate contexts, no communication) coded the
same six maximally different events — Yamnaya lower Danube, Greek Sicily,
British New England, the Crimean Tatar deportation, Chagos, Norse Greenland —
on all sixty dimensions with per-event web research. Agreement: **0.93 exact,
0.99 within one anchor step, mean absolute difference 0.09** on the 0–4
scales, over **345 scored comparisons** (unknown-involved pairs excluded).
The pilot is an instrument-debugging exercise at small scale, not evidence of
accuracy; and because both coders independently reported the same defect (no
way to say "no referent"), the `-1`/`-2` distinction postdates the pilot and
was **not** validated by it (`analysis/coding/pilot/agreement.json`).

### 4.2 A model-selection gate that failed a model

Before production, a cheaper candidate coder model was trialed against a
banked 30-event reference set (1,564 numeric cells; 236 N/A-involved cells).
It scored **0.67 exact**, mean absolute difference 0.45 — and, decisively,
**0.22 concordance** on the `-1`/`-2` distinction, which would have corrupted
precisely the missingness structure §3.3 depends on. The candidate was
excluded; the full corpus was coded on the stronger model, at material extra
cost. We report this because instrument failure that is caught and documented
is evidence the gate works, and because "which model coded your data" is as
reportable as "which assay kit" (`analysis/validation/agreement_sonnet.json`).

### 4.3 Production protocol

Events were dealt round-robin into era- and mechanism-stratified batches of
about ten, so no coder anchored on a single event type. Each coder read the
sixty-anchor codebook and a per-event research brief (register summary,
actors, dates, audit critique, vetted sources), performed bounded web
research, and returned schema-enforced integer vectors with per-dimension
flags and consulted-source URLs. Prompts included an anti-central-tendency
instruction (score the anchor the evidence meets, not the case's reputation)
and, after §4.2, worked examples of the `-1`/`-2` boundary.

### 4.4 Cross-model agreement, its weak point, and a recomputation gap

Thirty events were independently coded by two different production models:
**0.76 exact, 0.95 within-1, mean absolute difference 0.31** over 1,704
numeric comparisons — the observed cross-model agreement estimate for this
stratified 30-event sample, well below
the same-model pilot. On the 96 cells where either coder used a non-numeric
code, however, concordance on *which* non-numeric code was only **0.50**: the
models agreed much better about scores than about the boundary between
"unknown" and "inapplicable," and downstream analyses of missingness patterns
inherit that uncertainty. Measurement error at these levels may attenuate or
distort the recovered structure; because model coding errors can be
systematic rather than random, we do not claim the true structure is sharper
than the measured one. A reproducibility gap flagged by the v0 fact-check is
now closed: the raw coder journals behind both model comparisons were
recovered from the drafting session's working directories, copied into the
project at `analysis/validation/raw/` with SHA-256 sums, and both stored
comparisons recompute exactly from the in-project copies (independently
reproduced in the second review, and re-verified here: sonnet 0.669 exact /
0.216 N/A concordance; cross-model 0.763 / 0.953 / 0.500).

A clean-data stage then repaired first-pass defects without touching the raw
record: thirteen duplicate submissions reconciled by rule (agreement kept;
means for numeric disagreement; manual adjudication for structural
disagreements, each documented); two genuinely missing cells resolved by
explicit adjudication; gap-scan year fields completed. The repaired baseline
is frozen and hashed (`analysis/second_pass/baseline_407/`,
`data/duplicate_reconciliation.csv`).

### 4.5 The human boundary, and what the record actually is

The documented consequential decisions reviewed to date — event universe,
severity-axis inclusion, research depth, the no-merging rule, family
adoptions, case retirements, release promotions — were reserved for or
confirmed by the human author in the project's decision ledgers. The bounded
transcript audit subsequently corroborated every major decision it checked;
this is not a claim that every minor action across the project has been
exhaustively traced. The agents designed instruments *within* those
decisions, coded *under* them, and analyzed *for* them.

A standing project rule requires every precise source-backed fact found
during research to be written into the durable file record with a document
locator, because working-session context is unreliable as an archive. The
durable record is broader than prose: dated Markdown reviews, CSV ledgers,
JSON results and manifests, scripts, and hash-verified releases. During
drafting, v0 claimed the second-pass working conversations were unrecoverable
after two transcript-store searches returned nothing; a subsequent audit
found **five surviving session files** (16 and 19 August 2026) containing
substantive second-pass discussion. The claim is reversed here, and the
bounded recovery pass has since been **completed**
(`analysis/second_pass/PROCESS_FACT_RECOVERY.md`): six process facts newly
recorded (among them the provenance of the reconstructed Neo-Assyrian source
series and an unlinked report deployment), twelve major decisions confirmed
with session-and-line locators — including the decision stream reserved to
the human author, which the recovered record corroborates for every decision
checked — and explicit negative findings for the sessions containing nothing
unique. The episode is retained as a worked example of why the preservation
rule exists: neither chat *nor a single failed search of chat* is a
substitute for the written ledger.

---

## 5. Results I: a persistent trait structure with a boundary-sensitive count

### 5.1 The factor solution at the analysis baseline

Parallel analysis (Horn 1965) against a simulated null retains **eight
factors** on the 407-event first-pass matrix and again on the 490-event
expansion. The axes of that eight-factor solution, in descending variance:

1. **Government direction** (state orchestration versus settler agency) —
   organizing-agent locus, carrier centralization, perpetrator state
   centrality, metropole anchorage, intent legibility.
2. **Legal and property machinery** — title-transfer formalization, formal
   expropriation, land commodification, legal instrumentation, cadastral
   legibility.
3. **Settler implantation** — family-complete migrant streams, permanence
   orientation, stream coordination, self-reproduction, saturation.
4. **Elimination versus incorporation** — prior-population removal and
   ancestry non-persistence versus native-labor incorporation.
5. **Lethal coercion** — lethality, mortality share of decline, subsistence
   destruction, survivor confinement, sex-differential targeting.
6. **Durability of outcome** — sovereignty durability, structural persistence,
   present-day endpoint, return foreclosure.
7. **Displacement into occupied space** (statistical form; see §5.3).
8. **Coerced outside labor** — the exploratory factor mixes unfree incomer
   share and exogenous-labor structure with household, distance, and
   subsistence correlates; the corrected public construct is described below.

At 407 events these carry 53% of variance on sixty dimensions with none
dropped for missingness. Dimension-level convergence across disciplines is
pervasive: the first factor alone draws its top loadings from four of the
five batteries — the independent instruments agree about what co-varies,
which is precisely what the unmerged design was built to reveal.

### 5.2 What growth does to the count

**[TECH]** The factor *space* persists as the corpus grows; the factor
*count* does not:

| Checkpoint | Parallel-analysis retention | Ninth-factor margin |
|---|---:|---:|
| Repaired 407 | 8 | — |
| Earlier 490 | 8 | −0.045 |
| Earlier v3 583 | **9** | **+0.019** |
| Earlier v4 583 | **8** | **−0.00789** |

Between 407 and 490 events, per-axis loading congruence runs 0.896–0.985 and
shared-event factor-score rank correlations 0.934–0.988. In v3's 583-event
roster, the default analysis crossed the retention threshold for a ninth factor by a margin of
+0.019 — a boundary result, sensitive to missing-data treatment (complete-
case subsets retain eight; subsets restricted to well-documented events
retain nine) and to era and source weighting. Forcing eight factors at 583
recovers the named subspace but rotates several axes: individual 407→583
congruences range **0.653–0.958**, with the durability, legal-machinery, and
coerced-labor axes least stable. The ninth pattern itself is historically
intelligible but impure: it contrasts well-documented, mostly modern systems
of territorial closure, administrative classification, and codified
legitimation with source-sparse, mostly ancient cases, and it correlates
strongly with both chronology (ρ = +0.73 with start year) and documentation
density (ρ = −0.68 with unknown-cell count). "Documented
territorial-administrative closure" is a fair description of what it
measures; it splits and rotates a correlated block rather than adding one
independent trait, and it is not robust enough to become a ninth public bar
(`ROMAN_FACTOR_COUNT_SENSITIVITY.md`). The honest summary is not "eight
factors, replicated," but: **a reproducible multidimensional covariance
structure whose exact factor count sits on a documented boundary between
eight and nine, entangled at that boundary with how much the sources let us
know.** V4 retained the same event count after bounded substitutions but
returned to eight, with a ninth-factor margin of −0.00789.

The corrected 676-event successor candidate, later frozen as v5, sharpened
that conclusion. Its Horn result was nine in every one of twenty independent
500-draw calibrations and under every single-event deletion. Yet grouping connected successor programs,
using the complete-parent alternative, collapsing all successors to one
median, or dropping the 39 date-proxy rows returns eight. Weighting and
imputation alternatives are material as well. The ninth direction is therefore
a reproducible feature of the canonical candidate under its declared
preprocessing, not a universal ninth trait.

Whole-space perturbation resolves what per-axis congruence alone could not.
Across all 676 leave-one-event runs in that v5 battery, the nine-dimensional
space never moves more than 4.808 degrees, whereas forced eight reaches 86.489 degrees and
crosses ten degrees 36 times. Whole-program alternatives keep the
nine-dimensional maximum below ten degrees but show modest movement beyond
five degrees in seven of 138 scenarios. The public question is therefore not
simply “eight or nine bars?” The current eight-bar card remains an explicitly
editorial, evidence-aware projection for this release; the candidate result
does not mechanically add a ninth bar.

The finite post-v5 correction cycle provides an additional bounded post-v5
perturbation check of that interpretation. One wrapper retirement plus 117 source-adjudicated cell
changes produced the 675-event candidate subsequently frozen as active v6,
which still retains nine factors. Its nine-dimensional space is 2.765 degrees
from immutable v5, with minimum
principal-angle cosine `0.998836`; forced eight moves 6.685 degrees and gives
misleadingly low same-name comparisons for three axes even though best-matched
eight-factor loadings remain highly congruent. The correction itself is not
the source of the broader sensitivity: equal-era weighting and the three
declared alternative imputation treatments move the corrected nine-factor
space by 26.987, 32.608, 25.249, and 70.014 degrees, closely matching the v5
values of 26.539, 36.124, 24.977, and 71.313. The candidate therefore preserves
both findings at once: source-adjudicated roster and cell corrections do not overturn the canonical
structure, and the exact factor allocation remains conditional on how a
source-uneven corpus is weighted and completed. The 70.014-degree
structural-zero alternative is an extreme diagnostic bound, not a preferred
completion model.

These v6 promotion results are bound directly to the preserved candidate outputs
`analysis/second_pass/v5/post_v5_correction_application_v1/candidate/post_v5_corrected_successor_675/results/post_v5_model_comparison_v1/MODEL_COMPARISON_SUMMARY.json`
and
`analysis/second_pass/v5/post_v5_correction_application_v1/MODEL_COMPARISON_INTERPRETATION.md`.
The old 583/679 output is retained separately as a superseded diagnostic.

Two further caveats from the v3 stability battery. Durability is the one
era-confounded axis (Spearman ρ = −0.51 with start year): older events have
had longer to look permanent. Lethal coercion, by contrast, is
era-independent in that battery (ρ = 0.00), and on-atlas versus off-atlas
events differ in mean lethality by +0.03. This supports keeping severity
separate from category labels; it does not validate the normalized mechanism
taxonomy, which separately failed its reliability gate (§6.4).

### 5.3 From factors to the public card: a face-validity correction

The public eight-trait case card is **not** the exploratory factor solution,
and after v3 it deliberately is not: six bars are transparent weighted
composites of coded dimensions that load consistently in *both* the 407 and
490 snapshots; the seventh public bar replaced the statistical seventh factor
after manual Roman review exposed a face-validity failure (a case whose
residents demonstrably stayed could receive a high "displacement" bar,
because the factor mixed relocation with reversed land-conversion
dimensions). The public seventh bar is now a direct, evidence-gated
**Separation from homeland** measure (relocation distance + return
foreclosure, suppressed when no population was displaced), with the
statistical factor retained for diagnosis (`EIGHT_TRAIT_CARD_SCORING.md`).
The card survives the eight/nine boundary in §5.2 precisely because it never
depended on the factor count: it is an editorial summary of coded judgments,
defended as such. We flag this as the paper's model case of the intended
division of labor: statistics propose, source-anchored review disposes.

**Public F8 face-validity correction (2026-08-27).** Atlas authoring exposed a
second instance of the same general problem. The factor-informed bar labelled
“Coerced outside labor” included three conceptually secondary correlates—
household composition, geopolitical distance, and reversed subsistence
destruction—alongside the two dimensions closest to imported or compelled
incomers. Events with no imported labor stream could therefore receive
nonzero, apparently well-supported values.

The corrected public F8 is **Imported unfree labor**: whether a distinct
population was brought from elsewhere as enslaved, indentured, or otherwise
unfree labor. `scs_exogenous_labor_triad` is the applicability gate; when the
construct exists, `pel_unfree_incomer_share` helps scale it. Unknown
applicability remains unknown, structural N/A suppresses the bar, and a known
zero gate displays zero. One bounded exception permits applicability when an
event is itself the importation of unfree labor but the statistical triad is
N/A only because no settler population exists.

The frozen v2 comparison changes 540 of 676 public cards by at least five
points and 226 by at least twenty; 450 old nonzero readings become direct
zeros and 58 cases are structurally inapplicable after review. Eight gate
conflicts were individually adjudicated with sources, leaving zero unresolved.
The complete 224-case/234-frame implementation then passed one predeclared
desktop and mobile audit with zero issues. This magnitude rules out a quiet
case-specific patch and justifies a corpus-wide public correction. The
statistical F8 remains an exploratory covariance axis in the immutable
release; no factor score was rewritten. The generator, comparison, review
queue, model, implementation disposition, and hashes are preserved in
`analysis/second_pass/v5/eight_trait_card_f8_review/`.

---

### 5.4 The completed post-v6 correction cycle (7 September 2026)

The calculations and their prescribed verification are complete. This edition
reports the 676-event successor and its separately documented 8 September
chronology addendum. Its current display profiles and metadata are joined in
the accompanying package. Frozen v6, the original diagnostic inputs and the
original 565 result records remain unchanged; the later metadata and affected
results are explicitly identified, not silently substituted.

#### Register corrections

The successor has 676 modeled events: 675 in v6, minus 13 retired rows, plus
14 bounded successor definitions. Its 771-row metadata register preserves
95 inactive records for the audit trail. These are separate counts from the
370 authored cases and 380 playback stops in atlas build `20260906a`.
Statistical membership does not put an event into the animation.

The fixed correction cycle covered 17 coding packages and 2,716 cells.
Of these, 840 belong to the 14 new rows. Among the 1,876 reviewed cells in
retained rows, 927 changed and 949 were confirmed. The other 37,844 cells in
shared rows retain their exact stored values and unknown/not-applicable
markers, including 100 inherited fractional values. New coding was not used
as permission to round or rewrite the rest of the register.

Four events remain without a primary-family assignment: the Greek foundation
of Croton, the partial Paeonian return, Setia's foundation, and Sullan Faesulae.
The other 672 have editorial family labels. These labels record source-based
judgments or explicitly carried prior assignments; they are not the output
of a clustering algorithm. Public-authoring holds are a separate category.

#### The count persists; the fitted relationships change

The canonical baseline and corrected candidate both retain nine exploratory
factors. That agreement does not establish nine unchanged historical traits.
In deliberately fixed eight- and nine-factor comparisons, the largest
subspace angles between baseline and candidate are respectively 18.69° and
11.78°. A separate comparison that keeps the named axes attached to their
labels has a minimum absolute Tucker congruence of 0.078. It is not the same
calculation as allowing whole spaces to align. The common retained count
therefore cannot carry a claim that the named structure has replicated intact.

The old six-cluster diagnostic also changes: 17.98% of the 662 shared events
move cluster after label alignment (adjusted Rand index 0.632). Those clusters
are not the six public editorial families. No family was reassigned merely
to follow this output or make the partition appear more stable.

The sensitivity record contains 565 labeled scenarios, which reuse 536
distinct preprocessed matrix cases. Of the labeled scenarios, 524 retain
nine factors, 29 retain eight, and 12 retain seven. These counts are neither
independent replications nor probabilities that a particular factor count is
correct. Separate tests that omit disciplinary indicator batteries retain
six to eight factors.

Linked-program weighting makes the granularity issue concrete. Giving each
declared program a total weight of one leaves all 676 real rows in place,
but reduces their total weight to 583 and yields an effective sample size
of 615.78. The largest eight-factor subspace angle relative to the unweighted
candidate is 44.67°. There are no 583 averaged or synthetic events hidden
behind that weight total. This is an illustrative alternative weighting,
not a rule for revising historical boundaries to obtain a preferred result.

#### What these tests can and cannot establish

For this project, the results support retaining the separation between the
event register, descriptive cards and reviewed families. They also require
reporting sensitivity to evidence resolution, missingness, indicator overlap
and the weight given to linked episodes. There is no predeclared universal
angle threshold that turns these diagnostics into a pass/fail certificate.
An admissible historical boundary is not selected for its statistical
convenience.

Elapsed years alone do not make event units comparable. The candidate keeps
nominal chronology, dating uncertainty, coded process tempo and observation
horizon distinct. The original nominal-date diagnostics used 628 eligible
baseline rows and 627 candidate rows; the separately documented 8 September
metadata addendum below raises candidate eligibility to 629. No separately measured continuous series of
movement durations was supplied; subtracting start from end would measure
the calendar envelope instead. The Egypt-through-642-only and internal
Kitakami phase alternatives lack separately coded profiles and remain
explicitly unestimable. Neither is filled with copied or invented scores.

The eight public traits remain descriptive summaries of anchored judgments,
not the nine-factor solution. The successor card work preserves the frozen
formula for historical comparison while carrying forward the already approved
direct **Imported unfree labor** definition for F8. Its public adjudications
and display corrections do not themselves close six older raw-coding
follow-ups. Case-specific display qualifications also remain separate from
the general card calculation.

The completed verification independently reconstructed the scenario inputs
and reported comparisons, and reproduced selected factor fits and Horn
simulations. It did not refit every result or repeat the historical research.
It was performed by another AI agent, with prior upstream contributions
disclosed, not by an external human reviewer. Source review, coding and this
text also involve AI assistance under Craig Talbert's editorial authority.

The defensible result is a corrected register with documented alternatives
and observable limits. It is not proof of an unbiased worldwide census,
equal evidentiary access across eras, or a universal pattern of human nature.
The event-unit arguments and unfavorable sensitivities remain part of the
record available for criticism.

#### Two accepted chronology corrections: the 8 September addendum

A final comparison with earlier accepted decisions found two chronology
corrections that had not reached the numerical metadata. La Gomera's 1449
parent-midpoint proxy is replaced by the 1488–1489 revolt and punitive
sequence. The West Bank and East Jerusalem case uses the accepted 1967
onset, rather than treating the 2024 reporting cutoff as its onset; its era
therefore changes from 1990–present to 1945–1989. The process remains ongoing
in the accepted account. Neither correction supplies a measured movement
duration, and neither changes a raw code, missing-value type, family or event
boundary.

The [separate numerical addendum](../data/releases/second_pass_v7_676/results/date_metadata_addendum/RESULTS.json.gz)
preserves the original 565-scenario run and reuses its identical canonical
fit. Only two new matrices require fitting: equal-era weighting and
era-specific median imputation. Fourteen scenario aliases share those fits;
511 affected date-diagnostic records were recalculated after reproducing
their original values. Every original scenario has an explicit disposition,
and the frozen v6 baseline is unchanged. Candidate nominal-date eligibility
is now 629, compared with 628 in the baseline; the number with a separately
measured continuous movement duration remains zero.

The retention counts remain 524 nine-factor, 29 eight-factor and 12
seven-factor scenarios. Compared with their own pre-amendment versions, the
largest eight-/nine-dimensional subspace angles are 0.39°/0.30° for equal-era
weighting and 3.48°/2.42° for era-median imputation. These incremental changes
do not erase the larger sensitivities already reported above, establish a
universal stability threshold or turn linked scenarios into independent
replications. [Main's addendum disposition](../data/releases/second_pass_v7_676/results/date_metadata_addendum/MAIN_ACCEPTANCE.md)
records the checks and limits. It is a later main implementation check, not
part of the earlier independent AI-agent verification and not a new historical
source review.

#### Reproducibility and evidence

The package preserves the [roster and coding scope](../data/releases/second_pass_v7_676/provenance/core/SUMMARY.json),
[diagnostic summary](../data/releases/second_pass_v7_676/results/DIAGNOSTIC_SUMMARY.json),
[complete scenario table](../data/releases/second_pass_v7_676/results/SCENARIO_SUMMARY.csv),
and [full compressed numerical outputs](../data/releases/second_pass_v7_676/results/RESULTS.json.gz).
The [verification report](../data/releases/second_pass_v7_676/provenance/model_verification_v1/REVIEW.md)
and [main interpretation](../data/releases/second_pass_v7_676/provenance/MAIN_DIAGNOSTIC_DISPOSITION.md)
state the limits of the completed check. The
[family decision ledger](../data/releases/second_pass_v7_676/editorial/family_decision_ledger.json)
and [six open raw-coding follow-ups](../data/releases/second_pass_v7_676/editorial/F8_open_raw_followups.json)
preserve unresolved judgments rather than concealing them in a summary count.

The companion publication includes these linked files; its banner and
PUBLICATION.json identify the release and build served. The original dated preparation and review notes
retain the status they had when written; a later release decision does not
rewrite those artifacts. The manuscript remains a working paper, not peer
reviewed. The full v2 reorganization and bibliography refresh remain separate
editorial tasks.

---

## 6. Results II: no tested partition survives as taxonomy

### 6.1 The clustering looks fine until you test it

K-means on the eight-axis scores yields silhouette-optimal solutions (k = 10
at silhouette 0.24 on the first pass) that pass casual inspection. Testing
undoes them. Bootstrap resampling (Hennig 2007) finds only four of ten
clusters stable; the stability-optimal k = 6 solution (mean Jaccard 0.842 on
the first pass; 0.803 at 490 events) is *stable but shallow*: absolute
silhouettes near 0.24 sit far above matched-covariance unimodal noise (26–32
standard deviations — genuine density structure exists), yet five hundred
restarts produce 367 distinct near-tied partitions, and growing the corpus
from 446 to 490 reviewed events moves 71 of 446 shared cases between
free-solution neighborhoods (ARI 0.661). Carrying the prior partition forward
costs less than 1% in within-cluster error — many assignments are exchangeable
among the tested near-ties even though the density structure is not.

We state the conclusion at its defensible strength. These diagnostics
evaluate *this encoding, these algorithms, and this corpus*; clustering
cannot prove that history contains no natural kinds, and stable clusters can
themselves be artifacts of method (Hennig 2015). What the record shows is
narrower and sufficient: **no tested automatic partition is robust enough to
serve as the atlas's public taxonomy.** The phrase "no natural species,"
where it appears in project material, is an editorial metaphor for that
result, not a claim about historical ontology.

### 6.2 The archive-density experiment

**[TECH]** Why do partitions move? A controlled sensitivity replaced the
forty-four early Neo-Assyrian events — separately valid, separately bounded,
but drawn from one unusually granular archive — with non-historical medians
at three resolutions. One median per ruler (three rows): partition agreement
with the 446-event baseline rises from ARI 0.661 to 0.928 while minimum
factor congruence stays ≥0.999. One median per coding pattern (nine rows):
0.968. The factor axes barely notice; the clusters reorganize around
documentation density. The scoped inference: **in this corpus and this
design, automatic clusters weighted history by its archives** — a
well-recorded decade of Assyrian campaigning outvoted a poorly recorded
century elsewhere, not because more happened but because more survives per
event. We conjecture the mechanism generalizes to other historical
clustering exercises, and the diagnostic (resolution-matched median
replacement) transfers directly
(`19_early_neo_assyrian_resolution_sensitivity.py`;
`SECOND_PASS_INTERIM_RESULTS.md`).

### 6.3 What the old categories look like from here

Agreement between data-driven partitions and the register's twelve mechanism
families is ARI ≈ 0.29 — real, partial, and uneven. Three families are
structurally coherent: population exchange/partition (one neighborhood),
conflict displacement (97% concentrated), demographic expansion (96%
concentrated — the prehistoric cases long hedged as "analogues" are among the
*tightest* groups in the corpus, and the hedge understates them). Two
families fragment across five neighborhoods each — "state
resettlement/internal colonization" (largest share 33%) and
"expulsion/deportation/ethnic cleansing" (35%): names for intent or legal
character, applied to processes with fundamentally different architectures.
The middle of the vocabulary ("conquest with settlement," "genocide-linked
clearance," "enslavement/labor removal," "development displacement") is
mixed.

### 6.4 The resolution: families as reviewed filing, traits as measurement

The editorial consequence is not "abandon categories" — a public atlas needs
findable groups — but a demotion of their epistemic status. The atlas adopts
**six outcome families as a manually reviewed filing system**, informed by
the measurements and adjudicated case by case in a written ledger (146 rows
at the 490-event checkpoint; every disagreement between proposed label and
nearest statistical prototype resolved by source-anchored review, with the
source-first outcome retained in all nine early-Assyrian disagreements):

The following table preserves the historical v5 snapshot (2026-08-27), not
current v7 family totals. Frozen v6 has 673 assigned events and two family
evidence holds among 675 modeled events. V7 has 672 assignments and four holds
among 676 events, as stated in the opening snapshot; its current family table
is included with the data.

| Family (historical v5 labels) | Events (of 676) |
|---|---:|
| Expulsion or flight without replacement | 182 |
| Conquest and incorporation | 167 |
| State-directed demographic remaking | 144 |
| Settler replacement | 111 |
| Low-state durable expansion | 45 |
| Land clearance for state or project use | 25 |
| Family evidence hold | 2 |

Each rendered case additionally carries the **eight-trait profile** of §5.3 —
evidence-aware 0–100 summaries of the underlying anchored judgments, with
evidence-coverage values and conservative unknown ranges, and bars suppressed
where structurally inapplicable. For this release the public card presents one
reviewed family and eight measured lenses; normalized mechanism tags remain an
internal research layer after failing the reliability gate below. The card is
"a set of interpretable lenses, not a claim that history contains exactly
eight natural kinds."

**Mechanism/family diagnostic (candidate 676; terminal 2026-08-27).** The
normalized test is reproducible, but its binding disposition after the single
permitted repair and retest is
`FAIL_RELIABILITY_GATE_MECHANISMS_INTERNAL`. The source-based,
family/model-blind mechanism review produced 668 tagged events and eight holds;
the independently reviewed family roster has 674 labels and two nonoverlapping
holds. The pre-repair diagnostic therefore contains 666 events.

The first run was not accepted silently. Its one permitted final adversarial
review found that named dense research series crossed folds, that the program
cap omitted 43 early Neo-Assyrian cases, and that declared clustered
uncertainty and predictive-stratum outputs were absent. The failed run and
audit remain preserved. The corrected version binds an explicit 676-event
series crosswalk, joins linked event-unit and dense-series components before
fold assignment, caps each named series at five, and adds the missing outputs.
It contains 410 unified groups; the corrected program-capped analysis contains
506 events. The full battery and structural validator reproduce exactly with
zero group leakage.

In the pre-repair diagnostic, primary mechanism and family are strongly related
(Cramer's V = 0.637;
unified-group bootstrap interval 0.603–0.681), but they are not
interchangeable: normalized residual family entropy is 0.616, with a
unified-group bootstrap interval of 0.556–0.638. Primary mechanism is recorded
as `coequal_compound` for 186 of 666 events, so this descriptive table is not
a dependence analysis of the complete multi-label tag sets.

Repeated unified-group five-fold prediction gives macro-F1 0.684 from
mechanism tags, 0.614 from the public eight traits, and 0.771 from context,
traits, and tags. Adding tags to context plus traits improves held-out log loss
by 0.125 on average—about 11.6%—with a unified-group bootstrap interval of
0.062–0.201; the corrected 506-event program cap preserves the same
direction. Conversely, family alone only weakly reconstructs the full
multi-label mechanism sets (macro-F1 0.429, micro-F1 0.620). These results
support the conceptual distinction used here: mechanism describes *how* an
event operated, while family summarizes its dominant structural outcome. But
those estimates were fitted to the first adjudicated mechanism roster. They do
not establish that the mechanism vocabulary can be applied independently and
consistently enough to support a public layer.

The first independent mechanism review had already missed the frozen
reliability floor (micro kappa 0.686, positive-agreement Dice 0.749, tag
Jaccard 0.599). The project therefore used its one authorized repair cycle:
269 discordant or diagnostic events received source-first review; all 246
nonautomatic comparisons were adjudicated without averaging; the existing
fourteen tags were clarified rather than renamed or expanded; and all 676
events were reassessed under the clarified rules. The repaired roster contains
527 final or evidence-limited tag sets and 149 explicit holds—109 for defective
or mixed event units, 31 for evidence, and nine for incompatible complete tag
readings. Held rows were not converted into all-negative assignments.

The single permitted retest was then frozen before the reviewer began. Its
160-event sample covered all fourteen tags, all six families, 41 nonempty
family-by-era cells, sixteen macroregions, nine dense research series,
retained and successor events, on- and off-atlas events, all evidence bands,
changed cases, holds, and unchanged controls. The reviewer saw source evidence
and the repaired codebook but not the repaired assignments, families, traits,
models, sampling ledger, editorial summaries, or prior returns. Coverage
cleared the anti-attrition gate: 108 events were jointly reviewable, above the
minimum of 75 and equal to 83.1% of the 130 repaired-reviewable sample events.

Reliability did not clear the predeclared gate. Micro Cohen's kappa was 0.732,
positive-label Dice 0.790, and tag Jaccard 0.653, versus a required 0.80 on all
three. Nine common tags fell below the tag-specific kappa floor, and saved
stratum screens found systematic aggregate or tag-level reversals. Exact tag
sets agreed for only 28 jointly reviewable events; the frozen disagreement
queue contains 121 events. These are pre-adjudication scores and are not
altered by later interpretation.

The binding result is therefore
`FAIL_RELIABILITY_GATE_MECHANISMS_INTERNAL`. This is not evidence that
mechanism and family mean the same thing; it is evidence that this mechanism
instrument is not reliable enough to bear a second public analytical label.
For this release, the atlas should keep one primary outcome family and the
evidence-aware trait bars, preserve source-grounded process prose, and not add
normalized mechanism chips to the card. Legacy free-text mechanism labels
should be retired from analytical prominence during the next UI cleanup,
while their wording and the complete repaired ledger remain available as
provenance. The model battery is not rerun on the repaired tags, and the six
provisional primary-mechanism summaries are not release blockers.

The stopping rule was frozen before the 269 source-first returns were opened:
one repair, one fresh blinded retest, and no recursive tuning. Failure therefore
does not trigger another repair cycle. A future public mechanism layer would
require new evidence, a materially redesigned instrument, or a separately
authorized later release. The corrected 580-parent, grouped-parent, and 69
parent/successor mechanism sensitivities remain unestimated rather than being
filled with legacy labels from unreviewed wrappers. The binding output is
`analysis/second_pass/mechanism_family_v1/candidate_676_analysis_v2/source_first_review_v1/repair_v1/application/retest_v1/comparison/MECHANISM_RETEST_COMPARISON_SUMMARY.json`;
the frozen decision rule is in
`analysis/second_pass/mechanism_family_v1/CANDIDATE_676_ANALYSIS_IMPLEMENTATION_SPEC_V2.md`.
The superseded run and its audit remain in `candidate_676_analysis_v1/`.

**Historical normalization checkpoint (2026-08-27; superseded by the reviewed
diagnostic above).** The initial 679-event proposal is preserved but superseded by the
scope correction described above. The same family-blind normalization now
gives all 676 corrected candidate events fourteen composable broad process
tags with no untagged rows. It preserves all 331 raw mechanism strings, flags
all 676 event assignments for evidence review, and treats both the 320 bespoke
narratives and the inherited labels as starting proposals rather than final
data. The inherited labels require review because labels such as “conflict
displacement,” “conquest with settlement,” and “development/conservation
displacement” combine context, actions, and outcomes that the final tags must
separate.

On the 519 corrected-candidate events whose reviewed v4 family remains
available, a provisional ordinary five-fold model predicts family at 58.4%
accuracy from the proposed tags, 78.0% from the public eight traits, and 86.1%
from tags plus traits, against a 27.4% majority baseline. The incremental
result was large enough that public redundancy could not be assumed, but it
was not dispositive because the tags were unreviewed and the folds did not
hold linked programs and dense source series together. The reviewed,
audit-corrected grouped analysis above supersedes this checkpoint without
erasing it from the audit trail.

The v5 successor-family gate closed. Four source-only primary reviews
covered all 157 successors; a separately blinded 45-event audit agreed on 37
and sent eight to one no-averaging adjudication. The resulting full candidate
roster contained 674 analysis-ready family labels and two evidence holds. Four
disjoint, source-only mechanism reviews then covered the corrected 676-event
roster without access to family labels, traits, or model outputs. A bounded
76-event audit and one disagreement-only adjudication produced the first
internal roster. The later bounded repair and blinded retest did not clear the
public reliability gate. The current public direction is therefore one outcome
family plus the evidence-aware trait bars, without normalized “How it happened”
chips in this release.

---

## 7. Results III: the atlas as an editorial sample

Measuring the whole register, including everything the atlas excludes, makes
the atlas's own boundary legible. At the August 2026 measurement baseline
(222 rendered cases), on-atlas events differed from off-atlas events most
strongly in settler implantation (+0.99 SD), legal machinery (+0.66), and
durability (+0.52), and negatively in government direction (−0.59) and
displacement into occupied space (−0.59) — with, again, no lethality
difference. This is a coherent implicit definition of the atlas's historical
subject, now stated explicitly rather than enforced invisibly. Two large,
internally consistent regions of the space were nearly absent (transient
conflict displacement: 4 of 66; state demographic engineering: 17 of 61); one
measured axis — coerced outside labor — was *invisible* when the analysis was
restricted to on-atlas cases only (loading congruence 0.06), a pattern
consistent with the atlas's selection having excluded most of the
plantation-configuration variance.
Whether to move the boundary is an editorial question; the contribution of
the measurement is that it is now a *visible, quantified* question, answered
deliberately in a 587-row implementation ledger that partitions every release
row into retain, review, replace, author, hold, or exclude
(`ATLAS_V3_INTEGRATION.md`).

---

## 8. From analysis to editorial practice

A study that ended with a statistical release would misrepresent the project;
half its substance is the machinery that turns analysis into a public
artifact without laundering uncertainty. Concretely:

- **Versioned immutable releases.** Statistical corpora are promoted once,
  hash-manifested, and never edited. The active release is
  `releases/second_pass_v7_676/` for this edition; v6, v5, v4, v3, the 490-event, and 407-event
  baselines remain preserved. Statistical inclusion is explicitly *not*
  publication approval; cautioned rows travel with their cautions attached.
- **Retirement with provenance, and successors that earn admission.** Cases
  that fail review are retired, not deleted: retired umbrellas remain as
  inactive audit rows that must never render beside their bounded children;
  Norse Greenland's stable link redirects to the qualifying Thule case. The
  broad Athabaskan migration was retired with an explicit notice, and its
  bounded successors — the 1762 Sobaipuri-O'odham relocation, the 1667–1672
  Las Humanas abandonment, and the 1676–1677
  Quarai/Chililí-to-Tajique relocation — then passed the full admission pipeline
  (source brief, row-level dispositions, overlap crosswalk, event-unit gate,
  family review, map-authored routes and land outcomes) and were published in
  atlas build `20260828z` (then 313 cases, 323 playback stops). Their sixty-dimension
  coding and v4 statistical admission are complete.
- **One canonical playthrough.** There is no "contextual" second-class layer;
  an admitted case is a case. (Boundary decisions are therefore real
  decisions.)
- **A correction channel inside the atlas.** Case-scoped and whole-atlas
  reports enter the same candidate/correction ledger as the audit's own
  findings; public criticism is treated as a discovery mechanism, not a
  threat. The project calls this its Cunningham's-Law posture — the fastest
  route to a right answer in public is a confidently wrong one — and adopts
  it deliberately: a strong provisional register with a working correction
  channel, rather than an expert-review gate before launch.
- **Cartographic honesty rules.** Markers must visibly sit on land; only
  routes may cross water; a label that cannot be placed truthfully is omitted
  ("false spatial precision is worse than omission") — the map-level analogue
  of the register's evidence discipline (`MAP_RENDERING_RULES.md`).

---

## 9. Limitations

1. **Agreement is not validity.** Every number in §4 measures coder
   agreement; none measures correspondence with an expert gold standard. The
   cross-model bound is 0.76 exact — and only 0.50 on the
   unknown-versus-inapplicable boundary — so individual cell values should
   never be quoted as facts without their sources, and missingness-based
   analyses carry extra uncertainty.
2. **The factor count is a boundary result.** Eight at 407 and 490 events,
   nine in v3's 583-event roster by +0.019, eight in the v4 583-event
   roster, nine in v5's 676-event release, and nine in the v6
   675-event and v7 676-event releases under their declared
   preprocessing (§5.2). Grouping, date-proxy, weighting, and imputation
   alternatives move the count or factor space. Any presentation that fixes
   either eight or nine as a discovered constant misstates the record; the
   eight-bar card is editorial. The historical 675-event v6 successor
   reproduces nine factors and a 2.765-degree nine-dimensional distance from
   v5, but its forced-eight comparison moves 6.685 degrees; that contrast is
   why whole-subspace diagnostics accompany named-axis comparisons.
3. **Validation raw inputs were nearly lost.** The coder journals behind the
   model gate and cross-model check lived only in session working
   directories until the fact-check flagged the gap; they are now copied into
   `analysis/validation/raw/` and both comparisons recompute exactly (§4.4).
   The episode argues for depositing raw instrument outputs at generation
   time, not at audit time.
4. **Model-generated coding.** The instrument is anchored and
   agreement-tested, but it is not an expert panel; specialist review —
   invited through the correction channel — will find individual codes to
   fight about. The project's position is that a documented, contestable code
   beats an undocumented intuition.
5. **The corpus inherits its archives.** §6.2 is a warning about clustering,
   but it applies to everything: the 676 v7 events sample what survives. The
   disposition ledgers (advance / hold / searched-with-no-finding /
   inaccessible) bound, but do not remove, survivorship. This selected-positive-case
   corpus lacks a representative denominator for universal human propensity,
   inevitability, or the prevalence of peaceful alternatives (§2.3).
6. **One axis is era-confounded** (durability, ρ = −0.51 with start year);
   read old events' durability bars accordingly.
7. **The work grouping is not yet a finished bibliography.** The 854 URL
   records now resolve to 851 reviewed work groups, and the 14-pair duplicate
   queue is closed (§2.5), but many groups still need typed author, date,
   container, edition/locus, and access metadata before publication.
8. **Register scope decisions propagate.** The event-unit rule, inclusion
   scope, and audit judgments define what could be measured at all; a different
   rule would draw a different space. Full-roster application, source
   verification, score reconciliation, and alternative-resolution models were
   completed for the frozen releases, but they test the published rule rather
   than proving it uniquely correct. Those results do not validate post-freeze
   display overrides, unapplied corrections, or unresolved held events.
9. **Event-unit fairness is evidenced, not guaranteed.** Every active candidate
   row and relevant superseded parent has a recorded disposition, evidence
   trail, overlap treatment, and scoring path. No single row determines the
   ninth candidate direction, but program substitutions, date proxies,
   weighting, imputation, documentation, and source resolution remain
   material. These results support procedural auditability and bounded
   robustness; they do not prove that boundary bias has been eliminated or
   that every qualifying event has been found. The final Editorial Suite
   matched all 43 changed post-v5 claim groups and authorized this bounded
   wording for v6, subject to immutable release construction and validation.

---

## 10. Conclusion

Documented land-taking and forced removal recur across widely separated times
and places, in combinations whose similarities and differences warrant
comparison. That is the historical claim of this selected-positive-case
corpus, not a law of human nature. Without a representative denominator of
encounters or opportunities for taking, it cannot establish universal human
propensity, inevitability, or the rarity of peaceful alternatives.

The project set out to replace a hand-made taxonomy with a discovered one and
ended somewhere more instructive: a persistent, replicable trait structure
whose exact dimensionality is boundary-sensitive, and no tested automatic
partition sturdy enough to hand categories to the public. In the historical
407→583 comparisons, a multidimensional trait space persisted across a 43%
expansion; later v6 weighting and completion alternatives materially move the
space (§5.2). The factor count crosses a documented eight/nine boundary as
corpus composition and source resolution change; and every tested automatic attempt to make the
events fall into self-evident kinds fails the project's stability and
archive-density diagnostics. The defensible architecture is a division of labor —
measurement proposes, review disposes: reviewed families for navigation,
evidence-aware trait profiles for description, immutable releases for
accountability, and a correction channel for everything the instrument gets
wrong.

For comparative historical vocabulary, the specific findings are usable with
an important limit: severity emerged as a largely distinct direction from the
legacy mechanism variables in the factor solutions, with nonzero factor
correlations, but the normalized mechanism instrument failed its blinded
reliability gate. This is not an empirical vindication of a second public
mechanism taxonomy. Within the measured trait space, "expulsion" and
"internal colonization" name intents, not structures, and fragment when
measured; whether anyone arrives to replace the removed is one of the strongest
boundaries in the space; and the prehistoric expansions long treated as
analogues of the real thing are among its most coherent instances. For
method, the negative results are the portable ones: in this corpus, automatic
clusters weighted the past by its archives, and factor counts moved with
documentation density — so a project that publishes cluster labels or "the k
factors of history" without resolution-sensitivity and boundary diagnostics
is publishing its source density with extra steps.

For practice, finally, the project is a working demonstration with stated
uncertainties: a single scholar directing AI research agents under
agreement-tested protocols, explicit decision ledgers, and immutable releases
can build, measure, audit, and publish a worldwide register — provided the agents' work
is gated, the failures are kept, the record survives outside the
conversation, and every claim remains attached to a document a reader can
check. The fact-check that produced this revision is part of that record.

---

## AI contribution and disclosure statement

Instrument design (five independent disciplinary batteries), per-event source
research, sixty-dimension coding, statistical analysis, sensitivity design,
document drafting (including this monograph draft and its predecessor), and
atlas build tooling were performed by large-language-model research agents
from the Anthropic Claude family and OpenAI Codex (GPT-5–based), operating
under written protocols with schema-enforced outputs. Sonnet 5 was the
candidate coder and was **excluded** after failing the §4.2 gate; Fable 5
supplied the banked reference set; and Opus 5 produced the full coding matrix.
Codex performed substantial
second-pass source research, corpus and atlas integration, implementation,
and independent audit and verification work. The independent fact-check of
draft v0 and the later completion audit were likewise model-performed and are
preserved alongside this draft. The documented consequential scope,
inclusion, retirement, family-adoption, and publication decisions were
reserved for or confirmed by the human author (per the project's decision
ledgers; §4.5), who also set the constraints in §1. Model adjudicators resolved
many coding disagreements under written no-averaging rules; final human
responsibility attaches to the consequential editorial and publication
decisions, not a claim of personal review of every cell. Agreement statistics
and their limits are reported in §4. (Placement decided 2026-08-21: sole human
author with this statement in
back matter; venues that require the disclosure inside Methods can absorb it
into §4 without change of content.)

## Data and software availability

The register, codebook, coded matrices, missingness matrices, analysis
scripts, sensitivity diagnostics, release manifests with SHA-256 hashes,
integration ledgers, and this monograph's census generator and inventory are
in the version-controlled project directory under `audit/`, `analysis/`, and
`analysis/second_pass/releases/second_pass_v7_676/`. Git history and immutable
hash-manifested releases provide local integrity; an archival deposit remains
required before formal circulation. The publication destinations are the
[working monograph](https://latrian.dyndns.org/takingtheirland/monograph/) and
[checksum-bound release data](https://latrian.dyndns.org/takingtheirland/data/).
The atlas inventory is 370 cases and 380 playback stops, with a public feedback
endpoint. The opening snapshot and publication banner distinguish preparation
from an active public edition. V6 remains preserved. V7 explicitly includes
the corrected publication metadata, resolved atlas displays, original numerical
inputs/results and the later chronology addendum. The
[portable chronology reproduction bundle](../data/releases/second_pass_v7_676/reproduction/date_metadata_reproduction_20260908a.zip)
includes the unchanged original numerical export and the finite addendum path;
its input checks require no private cache. A complete consumer numerical replay
is an explicit additional operation, not claimed by the packaging check.

---

## References and background anchors (inventory in Appendix A)

*Citation-resolution pass completed 2026-08-21: all original 40 entries verified
against publishers or authoritative indexes — 36 correct without change, 4 corrected
(Sokoloff & Engerman author order and title; Lemkin subtitle and publisher;
Maier imprint; Scott subtitle), 0 unverifiable. Verdicts, checked URLs, and
edition/variant cautions (Belich subtitle variant; Tilly 1990-vs-1992
editions): `monograph/CITATION_RESOLUTION.md`. Malhi et al. 2008 was added on
2026-08-27 from the opened full-text PMC record and DOI metadata used in the
event-unit verification. Thirteen original entries remain background anchors
rather than in-text citations; a final circulation bibliography should either
cite them substantively or move them to a separately labeled background list.*

- Acemoglu, D., S. Johnson, and J. A. Robinson. 2001. "The Colonial Origins of
  Comparative Development: An Empirical Investigation." *American Economic
  Review* 91 (5): 1369–1401.
- Allentoft, M. E., et al. 2024. "100 Ancient Genomes Show Repeated Population
  Turnovers in Neolithic Denmark." *Nature* 625: 329–337.
- Belich, J. 2009. *Replenishing the Earth: The Settler Revolution and the
  Rise of the Anglo-World, 1783–1939*. Oxford University Press.
- Brace, S., et al. 2019. "Ancient Genomes Indicate Population Replacement in
  Early Neolithic Britain." *Nature Ecology & Evolution* 3: 765–771.
- Bulutgil, H. Z. 2016. *The Roots of Ethnic Cleansing in Europe*. Cambridge
  University Press.
- Cavanagh, E., and L. Veracini, eds. 2017. *The Routledge Handbook of the
  History of Settler Colonialism*. Routledge.
- Cohen, G. M. 2013. *The Hellenistic Settlements in the East from Armenia and
  Mesopotamia to Bactria and India*. University of California Press.
- Domar, E. D. 1970. "The Causes of Slavery or Serfdom: A Hypothesis."
  *Journal of Economic History* 30 (1): 18–32.
- Fieldhouse, D. K. 1966. *The Colonial Empires: A Comparative Survey from the
  Eighteenth Century*. Weidenfeld & Nicolson.
- Gilardi, F., M. Alizadeh, and M. Kubli. 2023. "ChatGPT Outperforms Crowd
  Workers for Text-Annotation Tasks." *PNAS* 120 (30): e2305016120.
- Gordillo, G. R. 2004. *Landscapes of Devils: Tensions of Place and Memory in
  the Argentinean Chaco*. Duke University Press.
- Haak, W., et al. 2015. "Massive Migration from the Steppe Was a Source for
  Indo-European Languages in Europe." *Nature* 522: 207–211.
- Hennig, C. 2007. "Cluster-Wise Assessment of Cluster Stability."
  *Computational Statistics & Data Analysis* 52 (1): 258–271.
- Hennig, C. 2015. "What Are the True Clusters?" *Pattern Recognition
  Letters* 64: 53–62.
- Horn, J. L. 1965. "A Rationale and Test for the Number of Factors in Factor
  Analysis." *Psychometrika* 30 (2): 179–185.
- Hu-DeHart, E. 1974. "Development and Rural Rebellion: Pacification of the
  Yaquis in the Late Porfiriato." *Hispanic American Historical Review* 54
  (1): 72–93.
- Hubert, L., and P. Arabie. 1985. "Comparing Partitions." *Journal of
  Classification* 2: 193–218.
- Lemkin, R. 1944. *Axis Rule in Occupied Europe: Laws of Occupation,
  Analysis of Government, Proposals for Redress*. Washington, DC: Carnegie
  Endowment for International Peace.
- Lipson, M., et al. 2018. "Ancient Genomes Document Multiple Waves of
  Migration in Southeast Asian Prehistory." *Science* 361: 92–95.
- Maier, C. S. 2016. *Once Within Borders: Territories of Power, Wealth, and
  Belonging since 1500*. Belknap Press of Harvard University Press.
- Malhi, R. S., et al. 2008. "Distribution of Y Chromosomes among Native North
  Americans: A Study of Athapaskan Population History." *American Journal of
  Physical Anthropology* 137 (4): 412–424. https://doi.org/10.1002/ajpa.20883.
- Mann, M. 2005. *The Dark Side of Democracy: Explaining Ethnic Cleansing*.
  Cambridge University Press.
- Naimark, N. M. 2001. *Fires of Hatred: Ethnic Cleansing in
  Twentieth-Century Europe*. Harvard University Press.
- Nieboer, H. J. 1900. *Slavery as an Industrial System: Ethnological
  Researches*. Martinus Nijhoff.
- Oded, B. 1979. *Mass Deportations and Deportees in the Neo-Assyrian
  Empire*. Reichert.
- Olalde, I., et al. 2018. "The Beaker Phenomenon and the Genomic
  Transformation of Northwest Europe." *Nature* 555: 190–196.
- Olalde, I., et al. 2019. "The Genomic History of the Iberian Peninsula over
  the Past 8000 Years." *Science* 363: 1230–1234.
- Osterhammel, J. 2005. *Colonialism: A Theoretical Overview*. 2nd ed. Markus
  Wiener.
- Polian, P. 2004. *Against Their Will: The History and Geography of Forced
  Migrations in the USSR*. CEU Press.
- Posth, C., et al. 2018. "Language Continuity despite Population Replacement
  in Remote Oceania." *Nature Ecology & Evolution* 2: 731–740.
- Reich, D. 2018. *Who We Are and How We Got Here: Ancient DNA and the New
  Science of the Human Past*. Pantheon.
- Rousseeuw, P. J. 1987. "Silhouettes: A Graphical Aid to the Interpretation
  and Validation of Cluster Analysis." *Journal of Computational and Applied
  Mathematics* 20: 53–65.
- Salmon, E. T. 1969. *Roman Colonization under the Republic*. Thames &
  Hudson.
- Scott, J. C. 1998. *Seeing Like a State: How Certain Schemes to Improve
  the Human Condition Have Failed*. Yale University Press.
- Seymour, D. J., ed. 2012. *From the Land of Ever Winter to the American
  Southwest: Athapaskan Migrations, Mobility, and Ethnogenesis*. University
  of Utah Press.
- Sokoloff, K. L., and S. L. Engerman. 2000. "Institutions, Factor
  Endowments, and Paths of Development in the New World." *Journal of
  Economic Perspectives* 14 (3): 217–232.
- Tilly, C. 1990. *Coercion, Capital, and European States, AD 990–1990*.
  Blackwell.
- Veracini, L. 2010. *Settler Colonialism: A Theoretical Overview*. Palgrave
  Macmillan.
- Viola, L. 2007. *The Unknown Gulag: The Lost World of Stalin's Special
  Settlements*. Oxford University Press.
- Wolfe, P. 2006. "Settler Colonialism and the Elimination of the Native."
  *Journal of Genocide Research* 8 (4): 387–409.
- Ziems, C., et al. 2024. "Can Large Language Models Transform Computational
  Social Science?" *Computational Linguistics* 50 (1): 237–291.

## Appendix A. Source inventory

`monograph/bibliography_census.csv` — 854 unique URL records with provenance
tags, regenerated by the saved `monograph/build_bibliography_census.py`
(URL-normalization dedup rule documented in the script; 845 at the
fact-check checkpoint). The reviewed work-level output is
`monograph/bibliography_works.csv`: 851 groups after all 14 near-title
relationships were adjudicated. Method and dispositions:
`monograph/BIBLIOGRAPHY_REVIEW.md` and
`monograph/bibliography_review_queue.csv`.

## Appendix B. (planned) Codebook

The sixty dimensions with anchors, evidence notes, and era-coverage rules —
currently `analysis/codebook/codebook.md` verbatim.

## Appendix C. (planned) Agreement, stability, and sensitivity tables

Pilot agreement; model-gate comparison; cross-model agreement including N/A
concordance; factor-count boundary diagnostics; loading congruence tables;
cluster stability and archive-density sensitivity; era-reweighting cosines;
event-boundary reviewer agreement; parent/child overlap coverage; and
equal-event versus grouped-program, parent/successor, era-, region-, mechanism-,
duration-, and documentation-resolution sensitivity.
All currently in `analysis/`, `analysis/validation/`, and
`analysis/second_pass/results/`.