# 10 — KDMP-to-KDMP clustering: how much does the program overlap with itself? **Run**: `python reports/10-coop-clustering/run.py` · No network · **Last run**: 2026-08-13 (on 08-13 coordinates) **Source**: `kopdes_locations.csv`, `kopdes_land_assets.csv`, `kopdes_stats_village.csv`, 08's [`suspect_coordinates.csv`](../08-exact-geometry/suspect_coordinates.csv) This is B1 of `analytics-plan.md`: _how many KDMP are built close enough to each other that they compete for the same population?_ The plan proposed an 83k × 83k pairwise distance matrix — 3.5 billion pairs, which is why this report does not build one. ## The method, in one paragraph Three instruments, in order of how defensible the number is. **(1) Exact nearest-neighbour distance**: a `cKDTree` on the unit sphere gives every cooperative's true geodesic distance to its closest sibling, so the "within 500 m / 1 km / 2 km / 5 km" counts are exact. **(2) H3 co-location**: how many cooperatives share the same H3 cell at r7/r8/r9 — the plan-review's recommended replacement for the pairwise matrix, which reproduces that review's published numbers exactly. **(3) Clusters**: groups of ≥2 cooperatives in the same r8 cell (~1 km hexagon), which is the plan's "densest KDMP clusters" deliverable at the finest resolution that does not chain (see Finding 5). Performance is joined through the two-hop village link (cooperative name → land asset → village stats), which reaches **79.3%** of cooperatives. ## The load-bearing caveat, up front **798 cooperatives (0.96%) are excluded from every "clean" number here and reported separately in [`coord_artifacts.csv`](coord_artifacts.csv):** - all **798** share an _exact duplicate coordinate_ with another cooperative — 386 groups, up to 11 cooperatives at one point, concentrated in ACEH (119), PAPUA PEGUNUNGAN (79), JAWA BARAT (62). The 19 previously-impossible coordinates from 08 were corrected by the ministry on 08-13 and are no longer artifacts. Exact duplicates are the signature of a geocoding fallback — cooperatives snapped to an admin centroid, not physically co-located. **At the fine r9 scale (~350 m), 64% of apparent co-location is these artifacts.** Any sentence about cooperatives "on top of each other" must be written against the clean set. ## Finding 1 — most KDMP have a sibling nearby; few are stacked on top of each other [`nn_bands.csv`](nn_bands.csv) · [`nn_within.csv`](nn_within.csv) · exact geodesic, 82,581 clean cooperatives | Another cooperative within | Cooperatives | Share | | -------------------------- | ------------ | ---------- | | **500 m** | 3,836 | **4.65%** | | **1 km** | 18,300 | **22.16%** | | **2 km** | 48,673 | **58.94%** | | **5 km** | 75,363 | **91.26%** | Read both ends. 91% having a sibling within 5 km is the program's 1-per-desa saturation in populated Indonesia — the same geometry that produced 03's "95% of people within 1.4 km of a KDMP". But it is not nothing: **one in five cooperatives has another one within a kilometre**, and 4.6% within 500 m — which is the real footprint of the "they built them too close together" claim. ## Finding 2 — true co-location is 6.7% of cooperatives, and most of it is pairs [`co_location.csv`](co_location.csv) (clean set) · [`clusters.csv`](clusters.csv) · [`cluster_size_distribution.csv`](cluster_size_distribution.csv) | H3 res | cell edge | coops sharing a cell with ≥1 other | share | max in one cell | | ------ | --------- | ---------------------------------- | -------- | --------------- | | r7 | ~1.2 km | 35,056 | 42.4% | 23 | | **r8** | ~0.5 km | 5,572 | **6.8%** | **17** | | r9 | ~0.2 km | 450 | 0.54% | 12 | 5,572 cooperatives sit in 2,513 r8 cells with at least one sibling. The distribution is dominated by pairs — 2,100 cells hold exactly 2, and only 8 cells hold 10 or more. The maximum is **17 cooperatives in one ~1 km cell**. ## Finding 3 — the densest clusters are not where you would expect [`densest_cells.csv`](densest_cells.csv) — top 20 r8 cells by cooperative count | Size | Where | Notes | | ------ | -------------------------------------------------------------------------------------------- | ------------------------------------- | | **17** | PAPUA PEGUNUNGAN, Kab. Jayawijaya (Wamena area) | 17 _recorded_ at one spot, ~1 km cell | | 12 | PAPUA PEGUNUNGAN, Kab. Lanny Jaya | | | 8 | ACEH, Kab. Aceh Barat Daya | all "Syariah" named | | 7 | JAMBI, Kab. Kerinci | | | 6 ×4 | Manado & Sungai Penuh (kelurahan coops in city centres), Kerinci, and a Papua (Pyramid) cell | | None are in Java's cities. The densest cell is in the Papua highlands, which is the same _badly sited or badly geocoded_ ambiguity as 04: a cluster of 17 distinct-but-adjacent coordinates in the Wamena area is more plausibly many desa geocoded to a town than 17 physically co-located buildings. **Every cell in this table needs imagery before any name is cited.** The `member_cooperatives` column is the checklist. ## Finding 4 — clustering carries no measurable performance penalty (null result) [`cluster_vs_isolated.csv`](cluster_vs_isolated.csv) · [`cluster_size_vs_value.csv`](cluster_size_vs_value.csv) · village-linked cooperatives only (79.3%) | group | n linked | pct reporting a transaction | mean Rp/coop | share of linked value | | ------------------------ | -------- | --------------------------- | ------------ | --------------------- | | clustered (same r8 cell) | 3,952 | **4.50%** | 3.19M | 7.19% | | isolated | 61,548 | **3.72%** | 2.65M | 92.81% | | all | 65,500 | 3.76% | 2.68M | 100% | Spearman correlation between cluster size and per-cooperative transaction value: **−0.008** over all clustered cooperatives, **+0.097** over those reporting at all. There is no signal in either. The plan's B1 hypothesis — _clustered KDMP have lower per-unit transaction volumes_ — **is not supported**. Cooperatives that share a cell report at the same rate as isolated ones (if anything slightly more often), and being in a bigger cluster does not depress the per-cooperative figure. Read this with 02's caveat in mind: the outcome is 97% zero, so "no penalty" is the strongest statement this data permits, not a measured absence of competition. ## Finding 5 — why there is no DBSCAN-style cluster map [`chaining_check.csv`](chaining_check.csv) A density-connected partition (the plan's DBSCAN, eps≈2 km) is not a meaningful object in this data. At any linking radius, components chain across dense Java: | linking radius (r8 k-rings) | coops inside a ≥3 component | largest component | | --------------------------- | --------------------------- | ----------------- | | ~0.9 km | 25.5% | 631 | | ~1.4 km | 55.7% | 2,652 | | ~1.8 km | 82.3% | 23,129 | The 23k-member component is essentially "all of Java". That is itself a finding: **the populated half of Indonesia is a near-continuous field of KDMP, not a set of discrete clusters** — the program saturated the landscape rather than concentrating in pockets. Same-cell co-location (Finding 2) is the honest unit for "on top of each other". ## Caveats - "Co-located" here means _recorded at_ the same cell, not _built_ there. The 798 artifacts show how often the data itself is the explanation; dense Papua cells are the open question. Same imagery bar as 04. - An r8 cell is ~0.46 km edge length; two cooperatives in one cell can be up to ~0.9 km apart (opposite corners). The nearest-neighbour distances are exact geodesic; the cell-based statements are approximate by construction. - The performance comparison covers only the 79.3% of cooperatives reached by the two-hop village link. Clustered and isolated coops are both ~6% of their respective sets, so the comparison is not imbalanced by reach. - A zero transaction is "has not **reported**", never "is inactive" (01). - [`nn_distances.csv`](nn_distances.csv) is the clean set only; [`coord_artifacts.csv`](coord_artifacts.csv) lists every excluded cooperative and why. ## Output for later [`nn_distances.csv`](nn_distances.csv) is per-cooperative and map-ready: `m_to_nearest_other`, `nn_band`, `cluster_id`, `cluster_size`, plus the village-linked performance columns. It is the intended input for a screengrid view showing self-overlap alongside the population and retail layers.