# Coverage due-diligence protocol Created 2026-08-16. This file defines what must be done before the project says it has made a genuinely comprehensive search for land-taking and removal events. It does not promise that a finite historical register can prove that no event has ever been missed. ## Scope to freeze before searching - Temporal range: approximately 6500 BCE to the present, matching the atlas. - Inclusion: historically identifiable direct land taking **or forced removal**, including settlement, conquest, demographic remaking, expulsion, blocked return, and land clearance for military, development, extraction, or conservation use. - Distinction: record whether newcomers conquered and stayed, residents were removed without replacement, or land was cleared for a non-settlement use. - Event unit: split when actors, affected population, mechanism, geography, or chronology changes materially. - Evidence: retain the audit's source and confidence requirements. Record uncertainty rather than filling gaps by analogy. - Source-point preservation: save every precise fact, contradiction, and negative finding with its source and page/section-level locator when available. Preserve the claim-to-source relationship in an event review, source brief, candidate ledger, or search log; do not leave it only in chat, an open browser tab, or an unindexed download. - Editorial objective: demonstrate documentable patterns across millennia. Ancient and premodern searches are therefore required coverage work, not an optional appendix; no event is added merely to satisfy an era quota. ## Release and correction policy The atlas is allowed to publish a strong provisional register before every possible expert has reviewed it. Publication should expose the event definitions, sources, confidence, known blind spots, and a practical correction channel. Public criticism and specialist feedback are expected discovery mechanisms. The correction channel must be available inside the atlas. A case-scoped report should automatically identify the case and build; a general report should allow whole-atlas criticism and missing-event proposals. Substantive reports must enter the same candidate/correction ledger used by the audit rather than disappear into an untracked inbox. This iterative approach does not lower the evidence threshold or excuse an uncrosswalked source universe. It changes outside review from a pre-publication perfection gate into a versioned post-publication process: publish a defensible version, log challenges, correct it, and rerun the affected coverage checks. ## Benchmark indexes to crosswalk There is no single index with this scope. The existing register uses individual UNHCR and IDMC sources, but as of 2026-08-16 it has not recorded a systematic record-by-record crosswalk against the following benchmark datasets: - UNHCR Refugee Data Finder, country/year refugee and related population data from 1951 onward: https://popstats.unhcr.org/refugee-statistics/methodology/data-content/ - IDMC Global Internal Displacement Database, event/trigger records from 2008 onward: https://www.internal-displacement.org/database/displacement-data/ - Uppsala Conflict Data Program event data, organized violence from 1989 onward: https://ucdp.uu.se/downloads/index.html - Government-Sponsored Mass Expulsion dataset, 139 cross-border mass-expulsion episodes from 1900–2020: https://www.belfercenter.org/publication/introducing-government-sponsored-mass-expulsion-dataset - Ethnicity of Refugees dataset, refugee-stock composition from 1975 onward: https://icr.ethz.ch/data/epr/er/beta.html - Oregon State University Dam Impacts Database, including displacement and resettlement fields for more than 500 selected dams: https://did.oregonstate.edu/ - Environmental Justice Atlas, a growing global catalog of environmental conflicts and community removals: https://ejatlas.org/ - Land Matrix, large-scale land acquisitions and their effects: https://landmatrix.org/ - RAMÛ Records of Ancient Migration, source-level imperial coerced migration from the Neo-Assyrian through Achaemenid periods: https://researchportal.helsinki.fi/en/projects/records-of-ancient-migration-locating-the-relocated-under-empire/ Add specialist regional or mechanism-specific datasets when discovered. These indexes are candidate generators, not authorities that override event-level historical research. ## Required procedure 1. **Crosswalk every benchmark record.** Give each a disposition: existing exact match, merged duplicate, new candidate, below threshold, out of temporal scope, unsupported, or unresolved. Never silently ignore a record. 2. **Build an era × region × search-mechanism review matrix.** At minimum divide the project into prehistoric, ancient, medieval, early modern, nineteenth century, 1900–1945, 1945–1989, and 1990–present cells. Record the sources and search terms used for each cell. 3. **Run specialist older-era sweeps.** Commission separate reviews of ancient imperial relocations, Greco-Roman settlement and deportation, South and East Asian imperial population policy, medieval religious and dynastic expulsions, early-modern confessional transfers, and precolonial African, American, and Oceanian cases. Modern humanitarian indexes cannot cover these. 4. **Run search-mechanism sweeps.** Independently review settler replacement, state-directed demographic remaking, expulsions and blocked return, conflict flight, population exchange, coerced labor movement, dams and infrastructure, military clearance, conservation removal, and extraction or plantation land seizure. These ten terms are discovery prompts only. They are not a public taxonomy or coded event field, and they do not revive the normalized mechanism vocabulary that failed its reliability retest. 5. **Keep a candidate ledger.** Every proposed event receives its sources, closest existing event, inclusion reasoning, rejection reasoning when applicable, and reviewer/date. Preserve negative findings. For older-era work, keep the source/search disposition log separately from the candidate list so a search that finds only an existing case or an access blocker is not silently lost. When a source supplies a precise date, count, place, route, artifact, legal act, or evidentiary limitation, save that point and its locator even if it is too detailed for the eventual atlas prose. 6. **Red-team the register independently.** Search specifically for omissions, bundled events, Eurocentric and modern-era bias, and events hidden by modern country containers. A verifier should not simply repeat the first searcher's source list. 7. **Publish an auditable provisional register and solicit correction.** Make sources, event boundaries, confidence, exclusions, and known blind spots visible. Invite public and specialist challenges, prioritize weak cells, and disposition every substantive correction in the candidate ledger. 8. **Repeat the implicated cause class after corrections.** Cause-code each omission and rescan the discovery method or scope that missed it, while retaining periodic cross-class checks. Do not restart every historical search indefinitely. ## Relationship to the statistical second pass Ancient and premodern coverage work is a required corpus-preparation stage for the queued second classification pass. The sequence is: 1. Run a declared older-era crosswalk and specialist search, beginning with RAMÛ and the regional sweeps above. 2. Apply the normal inclusion, event-splitting, sourcing, and confidence rules. 3. Code every accepted addition with the same eight-trait codebook and preserve unknowns where the evidence cannot support a value. 4. Freeze and identify that expanded corpus version. 5. Rerun the classification tests and compare them directly with the original 407-event result, including era sensitivity and documentation-density effects. This does not require delaying analysis until historical completeness can be claimed. It requires a substantial, auditable older-era search before the next corpus is frozen; later corrections can produce later versioned reruns. No era quota and no weaker ancient evidence threshold should be used to manufacture balance. ### Current progress, 2026-08-20 The first iteration of that sequence is complete, but the due-diligence protocol is not: - the repaired 407-event corpus is frozen as `analysis/second_pass/baseline_407/`; - 84 bounded ancient/premodern cases across nine source tranches passed source briefs, all 60 coding judgments, an expanded 490-event rerun, and manual family review; - the c. 900–612 BCE Neo-Assyrian umbrella has been source-reviewed and replaced without parent/child coexistence. Eleven later public-RINAP children retire the parent, while a completed ruler-by-ruler official-inscription review adds 44 separately bounded c. 900–745 BCE episodes. The active arithmetic is `407 + 84 gross additions - 1 retired umbrella = 490`; - the early Neo-Assyrian source ledger preserves 75 explicit source records. Forty-two advance dispositions become 44 non-overlapping event units after documented splits: 16 under Ashurnasirpal II, 24 under Shalmaneser III, and four under Šamši-Adad V. Holds, army-only taking, cumulative duplicates, ambiguous terminology, and source-poor years remain visible rather than being silently promoted; - the eight-trait structure remains stable, while thirteen of the 84 additions surface as current nearest-prototype hybrids; - the free automatic k=6 neighborhoods remain near-degenerate. The 446-to-490 test moves 71 of 446 shared cases in independent fits and 20 under continuity initialization; carrying the prior partition costs only 0.9827% more within-cluster error. These clusters are diagnostic and cannot define the public taxonomy; - a new source-resolution sensitivity test shows that much of that apparent movement is produced by archive density. Replacing the 44 early Neo-Assyrian records with three ruler medians, nine evidence-pattern medians, or thirteen ruler-by-pattern medians reduces shared-case movement to 12, six, or seven. Future tranches dominated by one archive must receive the same diagnostic; - the era-balanced test shows that state orchestration, legal machinery, and coerced labor redistribute some loadings within a stable combined block, reinforcing the need for broader older-era coverage before label freeze; - the Greco-Roman balancing ledger now preserves **407 explicit source dispositions**, including 97 direct Greek/Macedonian/Hellenistic prospective event units. Its separate 67-row Roman Republican colony catalogue has completed northern, central/southern, and remaining named-site priority reviews. Five northern event units advance behind broad-parent supersession, while Luna is a probable same-event Apuan land-use phase. Nine further catalogue rows yield ten central/southern event units because Fregellae's Roman foundation, intervening Samnite capture, and later Roman destruction require separate boundaries. Cales and Ariminum add two more event units; Luca is a negative control and Dertona remains an evidence hold. The corrected 28-row Sullan benchmark adds eight direct prospective units, including one medium-confidence Alban Hills regional unit resolved through a source-critical land-surveyor review, while preserving colony-only sites as holds and the national program as a parent control. The separate 23-row Sullan punishment sibling review adds seven prospective units, preserves five holds and four support or overlap records, and rejects six booty or spoil leads without a resident-land outcome. The first eighteen-row standard-tier foundation batch adds eleven prospective 329-241 BCE units, four same-event or origin-land supports, and three evidence holds. It groups linked Auruncian and Aequian sequences and prevents Firmum from duplicating the existing Picentine transfer. The second eighteen-row checkpoint covers all seventeen remaining standard rows plus the Castrum control: thirteen prospective 194-122 BCE units, three support nodes, one standard evidence hold, and one related textual-identity hold. It groups the four Campanian coastal nodes and keeps different affected communities and source lands separate. A subsequent twelve-row pre-Gracchan viritane-and- controls benchmark advances eight prospective units, preserves one location and overlap hold, rejects two false or non-qualifying event leads, and kept Labici as a qualifying scope control that opened a separate pre-338 BCE foundation census. It also identifies Capua's 211-210 BCE removal, confiscation, and civic dismantling without inventing an incoming settler population, and treats the 232 BCE *lex Flaminia* as a later settlement phase rather than a duplicate Senone expulsion. The subsequent nine-row Gracchan review advances five bounded regional land-division implementations, preserves Polla and Calabria/Bruttium as explicit source-resolution holds, and keeps both the national program and the 111 BCE regularization out of the event count. The subsequent twelve-row late-Republic residual review advances four prospective units, preserves four source or implementation holds, and records four rejected, annulled, or unlocated-program controls. It separates the 146 BCE destruction of Carthage from Junonia's later allotments, records Narbo as land taking with Indigenous continuity, and groups the disputed Marian African evidence into one medium-confidence program rather than one event per late title. The following two-row Salassian review advances one 25–24 BCE mass-enslavement, best-land-transfer, and Augusta Praetoria event and preserves the early Salassian *incolae* dedication as same-event survivor- continuity evidence. The completed thirty-three-row pre-338 BCE foundation census then advances eighteen bounded event units and preserves fifteen uncertain foundations, same-site supports, failed or unidentifiable proposals, and extension controls. Labici is updated in place, and the review freezes proposal/implementation, parent/child, sequence, and paired- versus-local sensitivities. All 22 priority catalogue rows remain dispositioned. A final row-level residual audit gives all 67 canonical parent components and all 96 Roman coding units explicit dispositions. It freezes complete retirement of both broad Roman parents when children activate, without using eleven evidence holds as a fictional residual event. The queue is 95 prospective events plus one count-neutral Apuan/Luna recode; **all ninety-six Roman coding units** now have a validated combined source-brief row and a frozen codebook-1.1 Roman coding protocol, but remain inactive or unimplemented behind the sixty-dimension coding and sensitivity gates. The 546 parent-only and 583 full-candidate counts are projections only and do not alter the active 490-event matrix; - a separate 35-row Hellenistic royal-transfer inventory now fully dispositions the first western/Macedonian synoikism and relocation benchmark. All records are source-frozen: twenty event-unit advances, eleven controls, and four evidence holds closed under the declared search. No Cohen catalogue lead or named first-pass gate remains open. The review requires implemented removal, clearance, land transfer, or bounded successor settlement rather than treating royal foundation or political union as proof of displacement. - a separate 19-row eastern Hellenistic priority inventory now preserves nine prospective event units, five negative controls, and five evidence holds. It tests destruction and removal outside foundation catalogues, counts Ptolemy's 312 BCE four-city razing as one linked operation, separates the reversed Jerusalem Akra removals, and rejects generic Ptolemaic cleruchies and a mass-Babylonian-transfer claim as event units; - a follow-on 25-row Cohen-2013 eastern catalogue-trigger inventory preserves five prospective event units, twelve evidence or scope holds, and eight negative or overlap controls. It completes candidate disposition across the catalogue's geographic divisions, corrects Apameia on the Silhu to December 140 BCE, and keeps Ai Khanoum inside one broad Bactrian sequence rather than double-counting it. This closes the declared catalogue trigger pass but does not close the non-catalogue search; - the first full Alexander Indian city-treatment sibling pass adds a separate 27-row source-frozen inventory: eleven prospective event units, one Peucelaotis identity/resident-outcome hold, and fifteen negative, nested, or same-event controls. Connected Assacenian and Mallian narratives are grouped rather than weighted one case per city; peaceful submissions remain visible; and Rhambakia advances only inside the wider Oreitai conquest-and-colony sequence. These units are not coded or active; - the first bounded named Ptolemaic Egypt/Cyrenaica pass adds a separate 25-row source-frozen inventory: four prospective event units, four evidence or overlap holds, and seventeen negative, overlap, nested, or parent controls. Alexandria's 145–124 BCE evidence is one repeated sequence; the Barke/Ptolemais transfer is resolved as a non-event control; and ordinary foundations, reclaimed-land allotments, and the Great Revolt wrapper do not become cases without bounded resident–land outcomes. These new units are not coded or active; - the bounded Hasmonean regional sibling pass adds a separate 32-row source-frozen inventory: thirteen prospective event units—twelve newly identified plus the already counted Akra reversal—five evidence holds, and fourteen negative, nested, military-only, or parent controls. Three identifiers already existed, so 29 rows enter the shared ledger. Joppa's 164/3 BCE resident killing, 147 BCE negotiated capture, and 143 BCE expulsion remain distinct; Simon's Joppa, Gezer, and Akra operations remain separate; and source-dense Gilead and Jannaeus sequences are grouped to avoid archive- density inflation. Nothing is coded or active, and the 490-event matrix is unchanged; - the completed cross-inventory Hellenistic overlap audit gives all 97 direct Greek/Macedonian/Hellenistic advances a row-level shared-place, chronology, active-parent, and event-unit disposition. It finds no duplicate unit, no active-event supersession, and no parent/child coexistence. Seven rows can coexist with four active archaic Greek colonization parents only after explicit scope narrowing. Forty-two rows fall into twelve compact ruler, campaign, or shared-source groups that must receive both one-group-median and group-by-outcome-pattern sensitivity tests after coding. Four `advance_pending_*` Greek rows remain outside the freeze. These are nine coded and modeled source tranches plus western and eastern pre-coding Hellenistic source checkpoints, not saturation. RAMÛ remains blocked at record level, the other declared benchmark indexes have not all been crosswalked, and later Sasanian and other regional source series still require event splitting. The Neo-Assyrian official-inscription series is now bounded, coded, modeled, and manually reviewed for both the later replacement children and the admitted early tranche, but it is not a benchmark-complete crosswalk: specialist and non-royal evidence can still refine the record, and neither the public children nor the official corpus substitutes for RAMÛ's 4,327-record disposition. The Mongol parent is partially dispositioned after nine western-campaign additions; European captives and later workshop transfers remain explicit community-unit and overlap work. The New Kingdom source series is only partially dispositioned: two Levantine cases are coded, Konosso remains held against the broad Nubia wrapper, and the reported 74-case record table is unavailable. Three non-admitted Achaemenid children and two early Sasanian reports now have documented evidence-limited holds rather than open reviews; they reopen only if RAMÛ or a new source supplies the missing event or coercion evidence. The first manual scope tranche moved all 26 formerly open removal-family rows to presumptive inclusion under the forced-removal rule. The second manual tranche advanced 16 more and recorded nine explicit scope/evidence/event-unit holds. No omission remains in an undefined open bucket, and no pair of successive independent zero-new-event scans has occurred. The project therefore may describe a validated first expansion, but not completed comprehensive due diligence. ## Declared-scope coverage-report evidence The project may publish a finite coverage report for a declared scope when all of the following are true: - 100% of records from the declared benchmark datasets have a documented disposition. - Every applicable era × controlled-region × search-mechanism cell has at least two methodologically distinct search channels, or an explicit non-applicable or evidence/access-limited disposition. Any unsearched cell is listed. - Every included event has stable provenance, source minimums, and an event-unit review; every exclusion has a reason. - Cross-dataset duplicates and umbrella events are resolved explicitly. - Every located candidate has a current disposition, including explicit holds; search leads without real source locators remain separately identified. - Channel yields and overlaps, valid benchmark-disposition fractions, and known unsearched cells are published. No global coverage percentage is reported without a specified and validated estimator. - The remaining uncertainties and known blind spots are published alongside the coverage claim. The defensible wording is then: **this coverage report states the declared scope, where and how the atlas searched, what it found, every current candidate disposition, and the remaining limitations.** It closes that declared search program, not history. Later evidence may reopen a cell. Do not claim saturation, historical completeness, or a mathematically complete census.