Datasets:

FelNeg
/

bitcoin-transaction-network-2009-2013

# Bitcoin transaction network — data & processing documentation

**Window:** 2009-01-03 → 2013-01-08 (the exact window of Aspembitova, Feng, Melnikov & Chew,
*Fitness preferential attachment as a driving mechanism in bitcoin transaction network*,
PLoS ONE 14(8): e0219346, 2019).

**What it is:** a cluster-level and address-level transaction network (edge lists) built from the
raw Bitcoin blockchain by porting and running the paper's own extraction pipeline, with every
stage validated against independent counts. Everything here is derived from public chain data.

**Generated:** 2026-10-03. Manifest with sha256 checksums: `docs/data_manifest.json`.

---

## 1. Quick start (this is all most analyses need)

```python
import duckdb
ROOT = "<path-to>/bitcoin-network"
# The network: one row per transaction output
edges = duckdb.sql(f"""
  SELECT * FROM read_csv('{ROOT}/runs/full_window/out/user_edges.txt', header=false,
    columns={{'tx':'BIGINT','u_from':'BIGINT','u_to':'BIGINT','date':'VARCHAR','value_btc':'DOUBLE'}})
""")
# 25,356,719 edges over 4,910,855 clusters, 2009-01-03 .. 2013-01-08.
# `date` is YYYYMMDDHHMMSS. `u_from == u_to` marks a self-loop (coinbase / internal transfer).
# Cluster sizes and members
sizes = duckdb.sql(f"SELECT * FROM read_parquet('{ROOT}/runs/full_window/out/clsize.parquet')")
pk_of_cluster = duckdb.sql(f"SELECT * FROM read_parquet('{ROOT}/runs/full_window/out/pk_user.parquet')")
```

If you only take one file, take `user_edges.txt` plus `clsize.parquet`.

---

## 2. Provenance

| item | source |
|---|---|
| raw chain | AWS Open Data, `https://aws-public-blockchain.s3.us-east-2.amazonaws.com/v1.0/btc/transactions/date=YYYY-MM-DD/` (public, no credentials, decoded per-day parquet) |
| extraction code | `github.com/ivanbrugere/Bitcoin-Transaction-Network-Extraction` (Python 2, 2013); ported to Python 3 in `code/btc_py3/` with the algorithm unchanged so results stay comparable to the paper |
| paper's own dataset | `https://dynamics.cs.washington.edu/nobackup/bitcoin/bitcoin.tar.gz` (679 MB, GraphLab SFrame; aggregated pairs only) |
| partition loader | `scripts/download_parquet_days.py` |
| parquet → extraction text | `scripts/gen_transactions_txt.py` |

Downloaded for this work: **1,462 daily partitions, 15.06 GB** (`data/parquet/`, not redistributed —
re-downloadable in ~17 min at ~14.6 MB/s with the script above).

---

## 3. Directory layout

```
bitcoin-network/
├── data/
│   ├── parquet/            1,462 daily partitions, 15 GB   (raw source, re-downloadable)
│   ├── raw/                7.5 GB  extraction input text (7.4 GB) + 2010 test slice
│   └── bitcoin/            the paper's own tarball, decoded (aggregated pairs)
├── runs/
│   ├── full_window/        4.6 GB  ← THE DELIVERABLE (see §4)
│   │   ├── out/            all edge lists, maps and flags
│   │   └── step3/          cluster assignments (.npy)
│   └── 2010_window/        small validation run (2009-2010 slice)
├── code/btc_py3/           the Python 3 port of the extraction pipeline
├── scripts/                generators, runners, analysis
└── docs/
    ├── DATA_AND_PROCESSING.md   ← this file
    ├── data_manifest.json       sizes + sha256 for every artifact
    ├── data_processing_findings.md   chronological working record, all measurements
    └── clustering_best_practice.md  clustering methodology + literature + measured comparisons
```

---

## 4. File inventory

All paths relative to the repository root. Sizes and row counts are from `docs/data_manifest.json`
(regenerate with `scripts/analysis/inventory.py`).

### 4.1 The networks

| file | size | rows | what it holds |
|---|---|---|---|
| `runs/full_window/out/user_edges.txt` | 1.0 GB | 25,356,719 | **cluster-level network**: `tx_id, u_from, u_to, date, value_btc`; one row per output |
| `runs/full_window/out/addr_edges.parquet` | 1.1 GB | 25,356,719 | **address-level network** + change-detection flags: `tx_id, txid, sender_pk, dst_pk, value, ts, oidx, n_in, n_out, fresh_any, fresh_out, n_fresh_any, n_fresh_out` |

`user_edges.txt` is the primary object: 25,356,719 edges, 10,886,588 transactions, 4,910,855
clusters, 5,878,290 self-loops (23.18%), span `20090103181505` → `20130108235206`.

### 4.2 Id maps (read §5 before joining anything)

| file | size | rows | mapping |
|---|---|---|---|
| `out/pubkey_list.txt` | 314 MB | 8,906,547 | line → identity string (`"(None)"` first) |
| `out/pk_map.parquet` | 251 MB | 8,906,547 | `pk_id → ident` |
| `out/userkey_list.txt` | 67 MB | 4,910,855 | line = one cluster, listing its pubkey ids |
| `out/pk_user.parquet` | 32 MB | 8,906,547 | `pk_id → cluster id u` (the clustering) |
| `out/clsize.parquet` | 5 MB | 4,910,855 | `u → n_pk` (addresses per cluster) |
| `out/transactionkey_list.txt` | 675 MB | 10,886,588 | line → txid |
| `out/tx_map.parquet` | 378 MB | 10,886,588 | `tx_id → txid` |

### 4.3 Clustering inputs, flags and results

| file | size | rows | what it holds |
|---|---|---|---|
| `out/user_edge_input_public_keys.txt` | 223 MB | 10,670,810 | per tx: its input **pubkey ids** — the re-clustering handle |
| `out/user_edge_inputs.txt` | 248 MB | 10,670,810 | per tx: its input **values in satoshi** (NOT ids — see §5) |
| `out/pairs.parquet` | 73 MB | 22,030,699 | exploded `(tx_id, pk_id)` input pairs |
| `out/change_pairs.parquet` | 58 MB | 3,727,704 | change-address candidates `tx_id, change_pk, sender_pk, value, n_out, n_in, rule` |
| `out/coinjoin_flags.parquet` | 163 KB | 28,470 | CoinJoin-like txs (≥3 in, ≥3 out, repeated output value) |
| `out/collapse_events.parquet` | 2.9 KB | 96 | merges absorbing ≥1,000 new pubkeys (collapse diagnostic) |
| `out/variant_stats.parquet` | 1.8 KB | 5 | headline statistics per variant/level |
| `step3/clusters_h1.npy`, `clusters_h2.npy` | 34 MB each | 8,906,548 int32 | cluster label per `pk_id` for the **unguarded** change variants — see the caveat in §8 |

`clsize.tsv` (46 MB) is the TSV twin of `clsize.parquet`, kept because it carries an
order-preserving index.

---

## 5. The three id spaces (critical — read before joining)

Verified directly against the 2013 source code:

| space | rule | trap |
|---|---|---|
| **cluster id** (`u` in `user_edges.txt`) | assigned 1,2,3,… in write order → `userkey_list.txt` line *k* is cluster *k* | must parse the file in order; a SQL `row_number()` over a parallel CSV read is **not** a file ordinal |
| **tx_id** | `t_id` starts at 1 and increments densely → `transactionkey_list.txt` line *k* is tx *k* | for an `in` line the *referent* txid is indexed before the *own* txid, so ids follow that order |
| **pk_id** | `"(None)"=0`, then **2,3,4,…** → `pubkey_list.txt` line 0 → id 0, line *k≥1* → id *k+1* | **pk id 1 is never used** (an off-by-one in the original). Indexing by line number shifts every id by one and silently corrupts every pk→cluster lookup |

And one file-name trap: **`user_edge_inputs.txt` holds values in satoshi, not ids.** It is the
per-input amount (`row[5]`, later scaled by 1e-9), so joining it as if it were identifier data
produces plausible-looking nonsense. The identifier file is
`user_edge_input_public_keys.txt`.

---

## 6. Processing pipeline, step by step

Each step is a script in `scripts/`, runnable end-to-end. Heavy steps run inside Docker
(`scripts/run_step*_docker.sh`) because they need more than the 4 GB that an agent-runtime background job
is capped at; the container gets its own cgroup.

### Step 0 — download (17 min, 15.06 GB)
`python3 scripts/download_parquet_days.py 2009-01-03 2013-01-08`
→ 1,462 day-partitions. Verified: 0 unreadable, schema clean for 2011–2013.

### Step 1 — parquet → extraction text (2 min 54 s, 7.9 GB, peak RSS 25 GB)
`python3 scripts/gen_transactions_txt.py data/parquet/ data/raw/transactions_2009_2013.txt`
→ **47,387,418 rows** (22,030,699 `in` + 25,356,719 `out`), 7 space-separated fields:
```
in  <own_txid> <spent_txid> <spent_out_idx> <pk> <iso_ts> <sortkey>
out <own_txid> <out_idx> <pk> <value_btc> <iso_ts> <sortkey>
```
Two non-obvious requirements (both silently corrupt results if changed): the timestamp must be a
single whitespace-free token of ≥19 chars with an ISO `T` (the parser slices fixed positions), and
coinbase transactions emit **only** their `out` rows (which is what turns mining rewards into
self-loops downstream). Pay-to-pubkey outputs have no address in the ETL, so the output pubkey is
recovered from `script_hex`, and P2PK *spends* (which store only a signature) are resolved by
back-joining `(spent_transaction_hash, spent_output_index)`.

### Step 2 — extraction (10 minutes, Docker)
`bash scripts/run_full_window_docker.sh`
Runs the ported pipeline: dictionaries → `fix_missing_public_keys` → clustering
(`write_user_public_keys`) → edge writing (`write_user_edges`).
→ `runs/full_window/out/` : `user_edges.txt`, `userkey_list.txt`, `pubkey_list.txt`,
`transactionkey_list.txt`, `user_edge_input*.txt`.

### Step 3 — address-level layer + freshness (2 minutes, Docker)
`bash scripts/run_step2_docker.sh` → `scripts/step2_build_address_edges.py`
Rebuilds address-level edges from the raw text with the pipeline's own sender rule (rows sorted
`(tx, io, idx)` ⇒ the sender is the **last** input; a coinbase's sender is its **first** output) and
computes per-output freshness flags.

**End-to-end validation (this is the important one):** mapping the reconstructed edges through
`identity → pk_id → cluster_id` and grouping by `(tx, from, dst)` gives
**25,356,719 edges in 24,383,640 groups — identical to the pipeline's own `user_edges.txt`,
with identical hash sums** (`224917165203570464631041835` on both sides). The address-level layer
is therefore provably faithful, not merely plausible.
### Step 4 — change heuristics, re-clustering, flags (3 minutes, Docker)
`bash scripts/run_step3_docker.sh` → `scripts/step3_change_heuristics.py`
- H1 (Meiklejohn et al. 2013): **3,727,704** transactions; of those, 2,353,768 also satisfy the
  stricter H2 (Ermilov et al. 2017) conditions
- 1,619,208 transactions excluded as self-change
- CoinJoin-like flags: 28,470
- collapse diagnostic: 96 merges absorbing ≥1,000 new pubkeys
- re-clustered variants: co-spend+H1 (4,910,855 → 2,561,952 clusters) and co-spend+H2
  (→ 3,380,393)
### Step 5 — variant statistics (1 minute, Docker)
`bash scripts/run_step4_docker.sh` → `scripts/step4_variant_statistics.py` → `variant_stats.parquet`
Computes nodes/edges/degree/component statistics at both granularities for the co-spend baseline,
the size-guarded change variant, and the hubs-removed variant.
### Analysis scripts (read-only, run on the host)
`scripts/analysis/`: `merge_profile.py`, `clustering_policies.py` (policy sweep),
`collapse_trace.py`, `exchange_removal.py` / `service_removal_degree.py` (removal sensitivity),
`compare_table2.py` (paper comparison), `inventory.py`, `make_manifest.py`.
---
## 7. Measured results (headline)
Full detail in `docs/clustering_best_practice.md`.
**Scale:** 25,356,719 edges · 10,886,588 transactions · 4,910,855 clusters · 8,906,547 pubkeys ·
23.18% self-loops.

**Independent validation:** distinct transactions (10,886,588) match blockchain.com's independent
daily series for 2011–2012 (10.4M) within 5%; edge count equals the generator's output-row count
exactly; the reconstruction check in §6 reproduces the pipeline byte-for-byte.

**Versus the paper's Table 2:** 2P nodes match to +2.4% (4,803,046 vs 4,689,069) and global scale
matches (25.4M vs ">25M" links). But their two periods use *inconsistent* link definitions (their 2P
number matches aggregated distinct pairs, their 1P number matches per-output edges), so their 1P
column cannot be reproduced from their own published dataset.

**Clustering methods, measured:**
- The naive change-address heuristic **collapses** the network: the giant cluster goes from 261,808
  to 3,315,509 addresses (37% of all addresses) and the cluster count halves.
- A **size-guarded** variant (refuse any merge above 10^4 addresses) applies 2,136,864 change merges
  (clusters −43%) while leaving the giant cluster **bit-identical** to baseline. Recommended policy.
- The degree tail is **robust to the clustering choice**: Hill tail slope ≈ 2.0 across
  address-level (2.22), co-spend (2.00), guarded-change (1.99) and hubs-removed (1.99).

**Two classes of giant node:** whales by address count (cluster 232: 261,808 addresses — the
Mt. Gox-shaped object whose 2012 consolidation burst we traced) versus high-frequency nodes
(cluster 213: 1,090 addresses but 4.9M edges; cluster 235: **3 addresses** and 615k edges, 97.3%
self-loops — the SatoshiDICE pattern).

---

## 8. Caveats and limitations (please carry these into any analysis)

1. **Clustering ≠ identities.** A "cluster" is a set of addresses linked by co-spend evidence. It is
   an upper bound on co-ownership, not ground truth. 89.6% of clusters are singletons.
2. **Self-loops (23.18%)** come from coinbase and internal transfers. Keep/drop changes the degree
   distribution; state the choice. They are excluded from `variant_stats.parquet` degrees.
3. **The `(None)` bucket** is one small cluster (373 pubkeys) holding outputs whose identity cannot
   be resolved. Left as a node.
4. **Temporal resolution is the block, not the second.** Timestamps are miner-set. Freshness flags
   are ordered by `(timestamp, txid)` because this text dump carries no block number — a documented
   approximation.
5. **`clusters_h1.npy` / `clusters_h2.npy` are the COLLAPSED variants.** Do not use them for network
   analysis without the guard described in §7. The baseline clustering is `pk_user.parquet`.
6. **The 261,808-address cluster is probably — not certainly — Mt. Gox.** It contains the May–June
   2012 consolidation burst and matches Ron & Shamir's 2012 entity scale, but entity attribution has
   not been performed. Do not publish it as an identified entity without that step.
7. **Hill tail slopes are crude** (fixed d_min = 100, no lognormal comparison). A defensible
   power-law claim needs MLE plus a likelihood-ratio test.
8. **The extraction does not scale past this window.** The ported pipeline is in-memory
   (~155 bytes/row); the modern corpus (~1.5B transactions) needs a streaming rewrite.
---
## 9. Reproducing / verifying
```bash
# 1. verify your copy against the manifest
python3 - <<'PY'
import json, hashlib, os
m = json.load(open("docs/data_manifest.json"))
for e in m["files"]:
    h = hashlib.sha256(open(e["path"], "rb").read()).hexdigest()
    print("OK " if h == e["sha256"] else "MISMATCH ", e["path"])
PY
# 2. rebuild everything from the raw chain (needs ~40 GB RAM for step 1, Docker for steps 2-5)
python3 scripts/download_parquet_days.py 2009-01-03 2013-01-08
python3 scripts/gen_transactions_txt.py data/parquet/ data/raw/transactions_2009_2013.txt
bash scripts/run_full_window_docker.sh
bash scripts/run_step2_docker.sh      # address-level layer + validation
bash scripts/run_step3_docker.sh      # change heuristics, flags, collapse events
bash scripts/run_step4_docker.sh      # variant statistics
```
Environment: Python 3 with `duckdb`, `numpy`, `scipy` (a Docker image carrying DuckDB and scipy is
used for heavy steps; inside the container `duckdb` is installed with
`pip install --break-system-packages duckdb`).
---
## 10. Sharing guidance
**To share everything:** the `runs/full_window/` tree (4.6 GB) plus `code/`, `scripts/`, `docs/`
(11 MB). The two large text files (`user_edges.txt` 1.0 GB, `addr_edges.parquet` 1.1 GB) dominate.

**Measured bundle options:**

| bundle | contents | size |
|---|---|---|
| minimum (network only) | `user_edges.txt` + `clsize.parquet` | **1.1 GB** |
| + re-clustering | the above + `userkey_list.txt` + `user_edge_input_public_keys.txt` + `pk_map.parquet` | **1.6 GB** |
| full deliverable | `runs/full_window/` + `code/` + `scripts/` + `docs/` | **4.6 GB** |

The small analysis outputs (`variant_stats`, `collapse_events`, `coinjoin_flags`, `change_pairs`,
`pk_user`, `tx_map`) together add well under 500 MB and are worth including in any bundle.

**What is shared where.** A single deposit: the Hugging Face repository
(`FelNeg/bitcoin-transaction-network-2009-2013`, private). It carries the processed networks
(`runs/full_window/`), the pipeline (`code/`, `scripts/`), these docs, and the extraction's
*input* layer, `data/raw/transactions_2009_2013.txt` (7.9 GB) — the unclustered per-input/output
records, so a collaborator can apply their own clustering instead of inheriting ours. Verify any
copy with `scripts/verify_hf_upload.py`.

**Not sent anywhere:** `data/parquet/` (15 GB) — publicly re-downloadable with the scripts in
under 30 minutes — and `data/raw/transactions_2010_clean.txt` (77 MB), an early 2010-window
scoping byproduct that is superseded by the full-window dump. `data/bitcoin/` is the paper's own
dataset, also public.

**Suggested transfer:** archive with checksums and let the recipient verify against the manifest:
```bash
tar --exclude='data/parquet' --exclude='data/raw' -czf btc-network-2009-2013.tar.gz \
    runs/full_window code scripts docs
sha256sum btc-network-2009-2013.tar.gz > btc-network-2009-2013.tar.gz.sha256
```

**Version control — read this before reaching for git.** `git rev-parse --show-toplevel` from this
project resolves to the parent research directory with **zero commits**, **no `.gitignore`**,
and **443 GB** of contents: several unrelated projects, model checkpoints, figures, plus the 15 GB
`data/parquet` and 7.5 GB `data/raw` trees. A `git add .` there would be destructive, and a push
would leak unrelated work or simply fail. **Use the tar bundle above for sharing.**

If a GitHub hand-off is wanted instead, it needs to be set up deliberately, not by pushing this
tree: create a *dedicated* repository for this project, add a `.gitignore` that excludes `data/`,
`runs/**/*.parquet`, `runs/**/*.npy`, `.duckdb_tmp/`, `__pycache__/`, and ship the large artifacts
as release assets or via the tar.

---

## 11. Responsible use

All underlying data is public chain data. However, address clustering is a re-identification
technique: even without names, a cluster label can point at a real entity, and the largest clusters
in this dataset are very likely identifiable services. For publication, prefer aggregates and
distributions; if individual cluster labels are shared, treat the "likely Mt. Gox / likely exchange"
attributions as unverified hypotheses, and state the clustering policy and its caveats (§8) in
any write-up.

	# Bitcoin transaction network — data & processing documentation

	Window: 2009-01-03 → 2013-01-08 (the exact window of Aspembitova, Feng, Melnikov & Chew,
	Fitness preferential attachment as a driving mechanism in bitcoin transaction network,
	PLoS ONE 14(8): e0219346, 2019).

	What it is: a cluster-level and address-level transaction network (edge lists) built from the
	raw Bitcoin blockchain by porting and running the paper's own extraction pipeline, with every
	stage validated against independent counts. Everything here is derived from public chain data.

	Generated: 2026-10-03. Manifest with sha256 checksums: `docs/data_manifest.json`.

	---

	## 1. Quick start (this is all most analyses need)

	```python
	import duckdb
	ROOT = "<path-to>/bitcoin-network"

	# The network: one row per transaction output
	edges = duckdb.sql(f"""
	SELECT * FROM read_csv('{ROOT}/runs/full_window/out/user_edges.txt', header=false,
	columns={{'tx':'BIGINT','u_from':'BIGINT','u_to':'BIGINT','date':'VARCHAR','value_btc':'DOUBLE'}})
	""")
	# 25,356,719 edges over 4,910,855 clusters, 2009-01-03 .. 2013-01-08.
	# `date` is YYYYMMDDHHMMSS. `u_from == u_to` marks a self-loop (coinbase / internal transfer).

	# Cluster sizes and members
	sizes = duckdb.sql(f"SELECT * FROM read_parquet('{ROOT}/runs/full_window/out/clsize.parquet')")
	pk_of_cluster = duckdb.sql(f"SELECT * FROM read_parquet('{ROOT}/runs/full_window/out/pk_user.parquet')")
	```

	If you only take one file, take `user_edges.txt` plus `clsize.parquet`.

	---

	## 2. Provenance

	\| item \| source \|
	\|---\|---\|
	\| raw chain \| AWS Open Data, `https://aws-public-blockchain.s3.us-east-2.amazonaws.com/v1.0/btc/transactions/date=YYYY-MM-DD/` (public, no credentials, decoded per-day parquet) \|
	\| extraction code \| `github.com/ivanbrugere/Bitcoin-Transaction-Network-Extraction` (Python 2, 2013); ported to Python 3 in `code/btc_py3/` with the algorithm unchanged so results stay comparable to the paper \|
	\| paper's own dataset \| `https://dynamics.cs.washington.edu/nobackup/bitcoin/bitcoin.tar.gz` (679 MB, GraphLab SFrame; aggregated pairs only) \|
	\| partition loader \| `scripts/download_parquet_days.py` \|
	\| parquet → extraction text \| `scripts/gen_transactions_txt.py` \|

	Downloaded for this work: 1,462 daily partitions, 15.06 GB (`data/parquet/`, not redistributed —
	re-downloadable in ~17 min at ~14.6 MB/s with the script above).

	---

	## 3. Directory layout

	```
	bitcoin-network/
	├── data/
	│ ├── parquet/ 1,462 daily partitions, 15 GB (raw source, re-downloadable)
	│ ├── raw/ 7.5 GB extraction input text (7.4 GB) + 2010 test slice
	│ └── bitcoin/ the paper's own tarball, decoded (aggregated pairs)
	├── runs/
	│ ├── full_window/ 4.6 GB ← THE DELIVERABLE (see §4)
	│ │ ├── out/ all edge lists, maps and flags
	│ │ └── step3/ cluster assignments (.npy)
	│ └── 2010_window/ small validation run (2009-2010 slice)
	├── code/btc_py3/ the Python 3 port of the extraction pipeline
	├── scripts/ generators, runners, analysis
	└── docs/
	├── DATA_AND_PROCESSING.md ← this file
	├── data_manifest.json sizes + sha256 for every artifact
	├── data_processing_findings.md chronological working record, all measurements
	└── clustering_best_practice.md clustering methodology + literature + measured comparisons
	```

	---

	## 4. File inventory

	All paths relative to the repository root. Sizes and row counts are from `docs/data_manifest.json`
	(regenerate with `scripts/analysis/inventory.py`).

	### 4.1 The networks

	\| file \| size \| rows \| what it holds \|
	\|---\|---\|---\|---\|
	\| `runs/full_window/out/user_edges.txt` \| 1.0 GB \| 25,356,719 \| cluster-level network: `tx_id, u_from, u_to, date, value_btc`; one row per output \|
	\| `runs/full_window/out/addr_edges.parquet` \| 1.1 GB \| 25,356,719 \| address-level network + change-detection flags: `tx_id, txid, sender_pk, dst_pk, value, ts, oidx, n_in, n_out, fresh_any, fresh_out, n_fresh_any, n_fresh_out` \|

	`user_edges.txt` is the primary object: 25,356,719 edges, 10,886,588 transactions, 4,910,855
	clusters, 5,878,290 self-loops (23.18%), span `20090103181505` → `20130108235206`.

	### 4.2 Id maps (read §5 before joining anything)

	\| file \| size \| rows \| mapping \|
	\|---\|---\|---\|---\|
	\| `out/pubkey_list.txt` \| 314 MB \| 8,906,547 \| line → identity string (`"(None)"` first) \|
	\| `out/pk_map.parquet` \| 251 MB \| 8,906,547 \| `pk_id → ident` \|
	\| `out/userkey_list.txt` \| 67 MB \| 4,910,855 \| line = one cluster, listing its pubkey ids \|
	\| `out/pk_user.parquet` \| 32 MB \| 8,906,547 \| `pk_id → cluster id u` (the clustering) \|
	\| `out/clsize.parquet` \| 5 MB \| 4,910,855 \| `u → n_pk` (addresses per cluster) \|
	\| `out/transactionkey_list.txt` \| 675 MB \| 10,886,588 \| line → txid \|
	\| `out/tx_map.parquet` \| 378 MB \| 10,886,588 \| `tx_id → txid` \|

	### 4.3 Clustering inputs, flags and results

	\| file \| size \| rows \| what it holds \|
	\|---\|---\|---\|---\|
	\| `out/user_edge_input_public_keys.txt` \| 223 MB \| 10,670,810 \| per tx: its input pubkey ids — the re-clustering handle \|
	\| `out/user_edge_inputs.txt` \| 248 MB \| 10,670,810 \| per tx: its input values in satoshi (NOT ids — see §5) \|
	\| `out/pairs.parquet` \| 73 MB \| 22,030,699 \| exploded `(tx_id, pk_id)` input pairs \|
	\| `out/change_pairs.parquet` \| 58 MB \| 3,727,704 \| change-address candidates `tx_id, change_pk, sender_pk, value, n_out, n_in, rule` \|
	\| `out/coinjoin_flags.parquet` \| 163 KB \| 28,470 \| CoinJoin-like txs (≥3 in, ≥3 out, repeated output value) \|
	\| `out/collapse_events.parquet` \| 2.9 KB \| 96 \| merges absorbing ≥1,000 new pubkeys (collapse diagnostic) \|
	\| `out/variant_stats.parquet` \| 1.8 KB \| 5 \| headline statistics per variant/level \|
	\| `step3/clusters_h1.npy`, `clusters_h2.npy` \| 34 MB each \| 8,906,548 int32 \| cluster label per `pk_id` for the unguarded change variants — see the caveat in §8 \|

	`clsize.tsv` (46 MB) is the TSV twin of `clsize.parquet`, kept because it carries an
	order-preserving index.

	---

	## 5. The three id spaces (critical — read before joining)

	Verified directly against the 2013 source code:

	\| space \| rule \| trap \|
	\|---\|---\|---\|
	\| cluster id (`u` in `user_edges.txt`) \| assigned 1,2,3,… in write order → `userkey_list.txt` line k is cluster k \| must parse the file in order; a SQL `row_number()` over a parallel CSV read is not a file ordinal \|
	\| tx_id \| `t_id` starts at 1 and increments densely → `transactionkey_list.txt` line k is tx k \| for an `in` line the referent txid is indexed before the own txid, so ids follow that order \|
	\| pk_id \| `"(None)"=0`, then 2,3,4,… → `pubkey_list.txt` line 0 → id 0, line k≥1 → id k+1 \| pk id 1 is never used (an off-by-one in the original). Indexing by line number shifts every id by one and silently corrupts every pk→cluster lookup \|

	And one file-name trap: `user_edge_inputs.txt` holds values in satoshi, not ids. It is the
	per-input amount (`row[5]`, later scaled by 1e-9), so joining it as if it were identifier data
	produces plausible-looking nonsense. The identifier file is
	`user_edge_input_public_keys.txt`.

	---

	## 6. Processing pipeline, step by step

	Each step is a script in `scripts/`, runnable end-to-end. Heavy steps run inside Docker
	(`scripts/run_step*_docker.sh`) because they need more than the 4 GB that an agent-runtime background job
	is capped at; the container gets its own cgroup.

	### Step 0 — download (17 min, 15.06 GB)
	`python3 scripts/download_parquet_days.py 2009-01-03 2013-01-08`
	→ 1,462 day-partitions. Verified: 0 unreadable, schema clean for 2011–2013.

	### Step 1 — parquet → extraction text (2 min 54 s, 7.9 GB, peak RSS 25 GB)
	`python3 scripts/gen_transactions_txt.py data/parquet/ data/raw/transactions_2009_2013.txt`
	→ 47,387,418 rows (22,030,699 `in` + 25,356,719 `out`), 7 space-separated fields:
	```
	in <own_txid> <spent_txid> <spent_out_idx> <pk> <iso_ts> <sortkey>
	out <own_txid> <out_idx> <pk> <value_btc> <iso_ts> <sortkey>
	```
	Two non-obvious requirements (both silently corrupt results if changed): the timestamp must be a
	single whitespace-free token of ≥19 chars with an ISO `T` (the parser slices fixed positions), and
	coinbase transactions emit only their `out` rows (which is what turns mining rewards into
	self-loops downstream). Pay-to-pubkey outputs have no address in the ETL, so the output pubkey is
	recovered from `script_hex`, and P2PK spends (which store only a signature) are resolved by
	back-joining `(spent_transaction_hash, spent_output_index)`.

	### Step 2 — extraction (10 minutes, Docker)
	`bash scripts/run_full_window_docker.sh`
	Runs the ported pipeline: dictionaries → `fix_missing_public_keys` → clustering
	(`write_user_public_keys`) → edge writing (`write_user_edges`).
	→ `runs/full_window/out/` : `user_edges.txt`, `userkey_list.txt`, `pubkey_list.txt`,
	`transactionkey_list.txt`, `user_edge_input*.txt`.

	### Step 3 — address-level layer + freshness (2 minutes, Docker)
	`bash scripts/run_step2_docker.sh` → `scripts/step2_build_address_edges.py`
	Rebuilds address-level edges from the raw text with the pipeline's own sender rule (rows sorted
	`(tx, io, idx)` ⇒ the sender is the last input; a coinbase's sender is its first output) and
	computes per-output freshness flags.

	End-to-end validation (this is the important one): mapping the reconstructed edges through
	`identity → pk_id → cluster_id` and grouping by `(tx, from, dst)` gives
	**25,356,719 edges in 24,383,640 groups — identical to the pipeline's own `user_edges.txt`,
	with identical hash sums** (`224917165203570464631041835` on both sides). The address-level layer
	is therefore provably faithful, not merely plausible.

	### Step 4 — change heuristics, re-clustering, flags (3 minutes, Docker)
	`bash scripts/run_step3_docker.sh` → `scripts/step3_change_heuristics.py`
	- H1 (Meiklejohn et al. 2013): 3,727,704 transactions; of those, 2,353,768 also satisfy the
	stricter H2 (Ermilov et al. 2017) conditions
	- 1,619,208 transactions excluded as self-change
	- CoinJoin-like flags: 28,470
	- collapse diagnostic: 96 merges absorbing ≥1,000 new pubkeys
	- re-clustered variants: co-spend+H1 (4,910,855 → 2,561,952 clusters) and co-spend+H2
	(→ 3,380,393)

	### Step 5 — variant statistics (1 minute, Docker)
	`bash scripts/run_step4_docker.sh` → `scripts/step4_variant_statistics.py` → `variant_stats.parquet`
	Computes nodes/edges/degree/component statistics at both granularities for the co-spend baseline,
	the size-guarded change variant, and the hubs-removed variant.

	### Analysis scripts (read-only, run on the host)
	`scripts/analysis/`: `merge_profile.py`, `clustering_policies.py` (policy sweep),
	`collapse_trace.py`, `exchange_removal.py` / `service_removal_degree.py` (removal sensitivity),
	`compare_table2.py` (paper comparison), `inventory.py`, `make_manifest.py`.

	---

	## 7. Measured results (headline)

	Full detail in `docs/clustering_best_practice.md`.

	Scale: 25,356,719 edges · 10,886,588 transactions · 4,910,855 clusters · 8,906,547 pubkeys ·
	23.18% self-loops.

	Independent validation: distinct transactions (10,886,588) match blockchain.com's independent
	daily series for 2011–2012 (10.4M) within 5%; edge count equals the generator's output-row count
	exactly; the reconstruction check in §6 reproduces the pipeline byte-for-byte.

	Versus the paper's Table 2: 2P nodes match to +2.4% (4,803,046 vs 4,689,069) and global scale
	matches (25.4M vs ">25M" links). But their two periods use inconsistent link definitions (their 2P
	number matches aggregated distinct pairs, their 1P number matches per-output edges), so their 1P
	column cannot be reproduced from their own published dataset.

	Clustering methods, measured:
	- The naive change-address heuristic collapses the network: the giant cluster goes from 261,808
	to 3,315,509 addresses (37% of all addresses) and the cluster count halves.
	- A size-guarded variant (refuse any merge above 10^4 addresses) applies 2,136,864 change merges
	(clusters −43%) while leaving the giant cluster bit-identical to baseline. Recommended policy.
	- The degree tail is robust to the clustering choice: Hill tail slope ≈ 2.0 across
	address-level (2.22), co-spend (2.00), guarded-change (1.99) and hubs-removed (1.99).

	Two classes of giant node: whales by address count (cluster 232: 261,808 addresses — the
	Mt. Gox-shaped object whose 2012 consolidation burst we traced) versus high-frequency nodes
	(cluster 213: 1,090 addresses but 4.9M edges; cluster 235: 3 addresses and 615k edges, 97.3%
	self-loops — the SatoshiDICE pattern).

	---

	## 8. Caveats and limitations (please carry these into any analysis)

	1. Clustering ≠ identities. A "cluster" is a set of addresses linked by co-spend evidence. It is
	an upper bound on co-ownership, not ground truth. 89.6% of clusters are singletons.
	2. Self-loops (23.18%) come from coinbase and internal transfers. Keep/drop changes the degree
	distribution; state the choice. They are excluded from `variant_stats.parquet` degrees.
	3. The `(None)` bucket is one small cluster (373 pubkeys) holding outputs whose identity cannot
	be resolved. Left as a node.
	4. Temporal resolution is the block, not the second. Timestamps are miner-set. Freshness flags
	are ordered by `(timestamp, txid)` because this text dump carries no block number — a documented
	approximation.
	5. `clusters_h1.npy` / `clusters_h2.npy` are the COLLAPSED variants. Do not use them for network
	analysis without the guard described in §7. The baseline clustering is `pk_user.parquet`.
	6. The 261,808-address cluster is probably — not certainly — Mt. Gox. It contains the May–June
	2012 consolidation burst and matches Ron & Shamir's 2012 entity scale, but entity attribution has
	not been performed. Do not publish it as an identified entity without that step.
	7. Hill tail slopes are crude (fixed d_min = 100, no lognormal comparison). A defensible
	power-law claim needs MLE plus a likelihood-ratio test.
	8. The extraction does not scale past this window. The ported pipeline is in-memory
	(~155 bytes/row); the modern corpus (~1.5B transactions) needs a streaming rewrite.

	---

	## 9. Reproducing / verifying

	```bash
	# 1. verify your copy against the manifest
	python3 - <<'PY'
	import json, hashlib, os
	m = json.load(open("docs/data_manifest.json"))
	for e in m["files"]:
	h = hashlib.sha256(open(e["path"], "rb").read()).hexdigest()
	print("OK " if h == e["sha256"] else "MISMATCH ", e["path"])
	PY

	# 2. rebuild everything from the raw chain (needs ~40 GB RAM for step 1, Docker for steps 2-5)
	python3 scripts/download_parquet_days.py 2009-01-03 2013-01-08
	python3 scripts/gen_transactions_txt.py data/parquet/ data/raw/transactions_2009_2013.txt
	bash scripts/run_full_window_docker.sh
	bash scripts/run_step2_docker.sh # address-level layer + validation
	bash scripts/run_step3_docker.sh # change heuristics, flags, collapse events
	bash scripts/run_step4_docker.sh # variant statistics
	```

	Environment: Python 3 with `duckdb`, `numpy`, `scipy` (a Docker image carrying DuckDB and scipy is
	used for heavy steps; inside the container `duckdb` is installed with
	`pip install --break-system-packages duckdb`).

	---

	## 10. Sharing guidance

	To share everything: the `runs/full_window/` tree (4.6 GB) plus `code/`, `scripts/`, `docs/`
	(11 MB). The two large text files (`user_edges.txt` 1.0 GB, `addr_edges.parquet` 1.1 GB) dominate.

	Measured bundle options:

	\| bundle \| contents \| size \|
	\|---\|---\|---\|
	\| minimum (network only) \| `user_edges.txt` + `clsize.parquet` \| 1.1 GB \|
	\| + re-clustering \| the above + `userkey_list.txt` + `user_edge_input_public_keys.txt` + `pk_map.parquet` \| 1.6 GB \|
	\| full deliverable \| `runs/full_window/` + `code/` + `scripts/` + `docs/` \| 4.6 GB \|

	The small analysis outputs (`variant_stats`, `collapse_events`, `coinjoin_flags`, `change_pairs`,
	`pk_user`, `tx_map`) together add well under 500 MB and are worth including in any bundle.

	What is shared where. A single deposit: the Hugging Face repository
	(`FelNeg/bitcoin-transaction-network-2009-2013`, private). It carries the processed networks
	(`runs/full_window/`), the pipeline (`code/`, `scripts/`), these docs, and the extraction's
	input layer, `data/raw/transactions_2009_2013.txt` (7.9 GB) — the unclustered per-input/output
	records, so a collaborator can apply their own clustering instead of inheriting ours. Verify any
	copy with `scripts/verify_hf_upload.py`.

	Not sent anywhere: `data/parquet/` (15 GB) — publicly re-downloadable with the scripts in
	under 30 minutes — and `data/raw/transactions_2010_clean.txt` (77 MB), an early 2010-window
	scoping byproduct that is superseded by the full-window dump. `data/bitcoin/` is the paper's own
	dataset, also public.

	Suggested transfer: archive with checksums and let the recipient verify against the manifest:
	```bash
	tar --exclude='data/parquet' --exclude='data/raw' -czf btc-network-2009-2013.tar.gz \
	runs/full_window code scripts docs
	sha256sum btc-network-2009-2013.tar.gz > btc-network-2009-2013.tar.gz.sha256
	```

	Version control — read this before reaching for git. `git rev-parse --show-toplevel` from this
	project resolves to the parent research directory with zero commits, no `.gitignore`,
	and 443 GB of contents: several unrelated projects, model checkpoints, figures, plus the 15 GB
	`data/parquet` and 7.5 GB `data/raw` trees. A `git add .` there would be destructive, and a push
	would leak unrelated work or simply fail. Use the tar bundle above for sharing.

	If a GitHub hand-off is wanted instead, it needs to be set up deliberately, not by pushing this
	tree: create a dedicated repository for this project, add a `.gitignore` that excludes `data/`,
	`runs/*/.parquet`, `runs/*/.npy`, `.duckdb_tmp/`, `__pycache__/`, and ship the large artifacts
	as release assets or via the tar.

	---

	## 11. Responsible use

	All underlying data is public chain data. However, address clustering is a re-identification
	technique: even without names, a cluster label can point at a real entity, and the largest clusters
	in this dataset are very likely identifiable services. For publication, prefer aggregates and
	distributions; if individual cluster labels are shared, treat the "likely Mt. Gox / likely exchange"
	attributions as unverified hypotheses, and state the clustering policy and its caveats (§8) in
	any write-up.