khipu-computational-toolkit

Phase 4: Geographic Patterns

Generated: 2026-03-08
Database: K-CAT SQLite database (built from KFG source data)
Script: scripts/run_phase4_geography.py
Inputs: data/processed/phase3_clusters.csv (Phase 3 feature matrix + cluster assignments)
Status: ✅ Complete

Research Question

Is structural complexity geographically patterned? Do specific summation pattern types associate preferentially with particular regions? And can structural similarity be used to attribute the 265 unprovenanced khipus (37% of corpus) to likely geographic zones?

Data & Methods

Sample: 444 provenanced khipus across 8 geographic zones; 265 unprovenanced excluded from statistical tests.

Zone construction: geo_zone consolidates ~82 provenance_display site names into 8 zones (see Phase 3). Unknown / collection-only provenance → Unprovenanced.

Analyses:

  1. Zone summary — complexity rate, mean cord count, mean pattern types per zone; 95% Wilson confidence intervals.
  2. Chi-square tests — cluster × geo_zone (overall); each of 9 binary pattern flags × geo_zone (largest zones with n ≥ 10).
  3. Pattern prevalence heatmap — mean has_* per zone for all 9 pattern types.
  4. Structural distributions — box plots of n_cords and n_pattern_types per zone.
  5. Nearest-neighbor attribution — for each unprovenanced khipu, 5 nearest neighbors in feature space among provenanced khipus (Euclidean distance over 9 binary pattern flags + scaled n_cords, n_groups, numeric_coverage); plurality-vote zone weighted by 1/distance.

Results

Complexity Rate by Zone

visualizations/phase4/complexity_by_zone.png

geo_zone n Complex Rate 95% CI
Chachapoyas 23 12 52% [33–71%]
Southern Highlands 2 1 50% — (n too small)
Nazca & Far South 33 11 33% [19–51%]
North Peru Coast 7 2 29% — (n too small)
Arica & N. Chile 11 3 27% [9–58%]
Cañete–Pisco 82 17 21% [13–31%]
Ica & Paracas 109 10 9% [5–16%]
Central Coast 177 15 8% [5–14%]
Provenanced total 444 71 16%  

Overall Geographic Signal (Chi-Square)

Test χ² dof p Significant?
Cluster × geo_zone 45.98 7 < 0.0001 ✅ Yes

Geographic zone is a statistically significant predictor of structural complexity class (p < 10⁻⁷).

Per-Pattern Geographic Signals

Chi-square tests on the 6 zones with n ≥ 10:

Pattern χ² p Significant?
has_is (indexed subsidiary sum) 37.65 < 0.001 ✅
has_sp (subsidiary → pendant) 32.81 < 0.001 ✅
has_psn (pendant sub neighbor) 25.49 0.0001 ✅ (PSN caveat applies)
has_pp (pendant → pendant) 12.47 0.029 ✅
has_adg (Ascher decreasing group) 8.69 0.122 ❌
has_ip (indexed pendant) 6.18 0.289 ❌
has_cp (color pendant) 5.93 0.313 ❌
has_gg (group → group) 7.19 0.207 ❌
has_gsb (group sum bands) 4.30 0.507 ❌

Four patterns show significant geographic variation; five do not. IS and SP — both requiring multi-level cord hierarchy (subsidiaries contributing to parents) — are the most geographically concentrated patterns. IP, CP, GG, GSB, and ADG appear at roughly consistent rates across zones.

Pattern Heatmap by Zone

visualizations/phase4/pattern_heatmap_by_zone.png

Mean Cord Count and Pattern Types by Zone

visualizations/phase4/structural_by_zone.png

geo_zone mean n_cords mean pattern types
Arica & N. Chile 283 3.4
Chachapoyas 250 4.6
North Peru Coast 177 3.9
Nazca & Far South 148 3.1
Cañete–Pisco 84 2.4
Ica & Paracas 68 2.2
Central Coast 64 2.5

Nearest-Neighbor Attribution (265 Unprovenanced Khipus)

visualizations/phase4/nn_attribution.png
Data: data/processed/phase4_nn_attribution.csv

Attributed zone Count % of unprovenanced
Central Coast 147 55.5%
Ica & Paracas 58 21.9%
Cañete–Pisco 41 15.5%
Chachapoyas 10 3.8%
Nazca & Far South 8 3.0%
Arica & N. Chile 1 0.4%

High-confidence attributions (top-zone weight ≥ 0.80): 48 khipus.

Important caveat: NN attribution is based purely on structural features. It cannot distinguish between khipus that are similar due to shared origin and those that are similar for other reasons. These attributions are hypothesis-generating, not confirmatory.

Note on museum provenance: museum_country is excluded from all analyses. It records current exhibition location, not origin.

Limitations

  1. Small zone sizes. Southern Highlands (n=2) and North Peru Coast (n=7) are too small for reliable statistics.

  2. Unprovenanced 37%. Geographic conclusions are limited to the 444 provenanced khipus. If the unprovenanced khipus are systematically different, the provenanced sample may be biased.

  3. PSN open question. PSN appears geographically significant but the KFG author considers the pattern likely coincidental. The IS and SP geographic signals are not dependent on PSN.

  4. NN attribution. The nearest-neighbor model is not calibrated — no ground-truth test set is available. The 48 high-confidence attributions are 18% of total unprovenanced.

How to Re-run

python scripts/run_phase4_geography.py

Reads data/processed/phase3_clusters.csv (Phase 3 must have run first).

Output Description
data/processed/phase4_zone_summary.csv Per-zone aggregate stats
data/processed/phase4_chi2_results.csv Chi-square test results
data/processed/phase4_nn_attribution.csv NN zone attribution for 265 unprovenanced
visualizations/phase4/ All PNG figures

See Citations and Acknowledgments in the project README for primary sources, data attribution, and toolkit provenance.

Corpus sweep run against K-CAT SQLite database. Re-run with scripts/run_phase4_geography.py to refresh.

This site is open source. Improve this page.