Strengthen empirical controls and offline reproduction

.github/workflows/ci.yml

@@ -0,0 +1,64 @@
+name: CI
+
+on:
+  pull_request:
+  push:
+    branches: [master]
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ci-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  python:
+    name: Python ${{ matrix.python-version }}
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.13"]
+    steps:
+      - name: Check out source
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
+      - name: Set up Python
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - name: Install test dependencies
+        run: python -m pip install --disable-pip-version-check ".[test,data]"
+      - name: Run offline gate
+        run: python -m pytest -m "not live and not reproduction"
+
+  dss:
+    name: Python 3.13 with pinned DSS
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out source
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
+      - name: Check out DSS
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
+        with:
+          repository: bc1cindy/dense-subset-sum
+          ref: d8099bc540c9de7d0f3c9f095e8b2aad8f9d8c8d
+          path: .ci/dense-subset-sum
+      - name: Set up Python
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+        with:
+          python-version: "3.13"
+          cache: pip
+      - name: Select DSS minimum Rust toolchain
+        run: |
+          rustup toolchain install 1.88.0 --profile minimal
+          rustup default 1.88.0
+      - name: Install project and DSS
+        run: |
+          python -m pip install --disable-pip-version-check ".[test,data]"
+          python -m pip install --disable-pip-version-check .ci/dense-subset-sum
+      - name: Verify DSS revision
+        run: python -c 'import dss; assert dss.__rev__ == "d8099bc540c9de7d0f3c9f095e8b2aad8f9d8c8d"'
+      - name: Run offline gate with DSS
+        run: python -m pytest -m "not live and not reproduction"

.github/workflows/reproduction.yml

@@ -0,0 +1,33 @@
+name: Evidence reproduction
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "17 4 * * 1"
+
+permissions:
+  contents: read
+
+concurrency:
+  group: evidence-reproduction
+  cancel-in-progress: false
+
+jobs:
+  bundles:
+    name: Reproduce committed bundles
+    runs-on: macos-14
+    timeout-minutes: 120
+    steps:
+      - name: Check out source
+        uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
+      - name: Set up Python
+        uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+        with:
+          python-version: "3.13.7"
+          cache: pip
+      - name: Install test dependencies
+        run: python -m pip install --disable-pip-version-check ".[test,data]"
+      - name: Verify committed bundle contents
+        run: python -m pytest tests/test_committed_bundles.py::test_every_committed_bundle_is_complete_and_content_verified
+      - name: Reproduce all committed bundles
+        run: python -m pytest -m reproduction tests/test_committed_bundles.py

.gitignore

@@ -3,10 +3,25 @@ __pycache__/
 *.pyc
 .cache/
 .blkcache/
+.blkcache-live/
 *.ndjson
 !catalog/entities.example.ndjson
 slice*.json
+!catalog/runs/slice-*.json
+!results/artifacts/slice-*.json
+!releases/slice-*.bundle.json
 slice*.err
+!catalog/datasets/slice-a-channels-2016-v1.json
 day-*.json
 day-*.err
 decluster.egg-info/
+
+# logs de coleta/análise (efêmeros)
+*.log
+*.done
+
+# épocas comprimidas da coleta multi-epoch (dados locais grandes)
+epoch_*.ndjson.gz
+data/epochs_*/
+# Planning notes and specs: working material, not part of the published record.
+docs/superpowers/

README.md

@@ -1,8 +1,10 @@
 # decluster
 
-Evidence-weighted **de-anonymization of Bitcoin transactions**. It reads how each
-transaction was built (wallet fingerprints) and how its amounts partition (subtransaction
-structure), scores every signal as **bits of evidence**, and clusters coins by owner.
+An adversarial laboratory for testing **de-anonymization failure modes in Bitcoin
+transactions**. It keeps construction fingerprints, amount models, graph structure,
+attribution and provenance as distinct evidence channels, including explicit abstention and
+cannot-link outcomes. Those channels are model-relative and do not establish true ownership or
+a general privacy score.
 
 Two coins spent in the same transaction are normally assumed to share an owner. Because the
 evidence here is *signed*, the clustering can instead **keep them apart** when their
@@ -14,17 +16,27 @@ two people into one).
   how it constructs a transaction — nSequence values, script types, signature grinding, and
   more. Do those quirks actually identify the wallet? Taking address reuse as the same-owner
   label (two transactions spending the same address are the same wallet), the measured
-  fingerprint bits rank a *same-wallet* pair of transactions above a *random* pair **93.3%
-  of the time** (AUC 0.933) on 166k real mainnet transactions. Shuffle the labels and it
-  drops to 0.50 (a coin flip) — so the 0.933 is real signal, not an artifact.
+  fingerprint bits rank a *same-wallet* pair of transactions above a *random* pair **94.6%
+  of the time** on the committed 600-transaction fixture (AUC 0.9459, the manifest-backed run)
+  and **92.4%** on the preserved 22,112-transaction cache (AUC 0.9244). Shuffle the labels and
+  both drop to ~0.50 (a coin flip) — so the separation is real signal, not an artifact. An
+  earlier figure of 0.933 over a 166k-transaction cache is a historical record: that cache was
+  not preserved and the number is not recomputable. Two caveats travel with these: the
+  preserved cache is Taproot-era only (heights 800,000–965,220), and on that cache 0.024 of
+  the AUC is the address-reuse label restating itself — dropping the five axes the shared
+  input address fixes outright takes 0.9244 to 0.9003.
 
-- **The shape of the payment graph reveals owners too.** Independently of who-spent-with-whom,
-  the *structure* of the graph (who pays whom) betrays common ownership — the same effect
-  that de-anonymized social networks (Narayanan–Shmatikov). Across five eras (2012–2024),
+- **The shape of the payment graph provides another model-relative signal.** Independently of
+  who-spent-with-whom, recurring counterparty structure can support linkage. The local
+  common-neighbour experiment is not the seeded cross-view Narayanan–Shmatikov attack. Across five eras (2012–2024),
   payment-graph structure *alone* predicts whether two addresses share an owner, ranking
   same-owner pairs correctly **0.95–0.97 of the time at one hop on the clean eras**, and
   **0.97–1.00 across all five eras by four hops** (1.0 = perfect, 0.5 = chance; the churny
-  2013 slice starts near chance at one hop and needs the deeper hops).
+  2013 slice starts near chance at one hop and needs the deeper hops). That is a *pairwise
+  ranking* score, and it is the premise the attack needs rather than the attack: run to
+  completion on two real Bitcoin views, the faithful 2009 propagation kernel reaches
+  **0.41%–1.28% precision**, against a shuffled-seed control that gets nothing right
+  (`results/RESULTS-ns-bitcoin.md`).
 
 - **It survives a transaction built to fool it.** On a real transaction deliberately
   constructed to merge two owners into one (the false link from above), the method keeps
@@ -34,16 +46,64 @@ two people into one).
 
 ## Layout
 
-- `decluster/` — the attacker (the measurement half that runs): extractors, library, combiner, cluster (engine: `cluster_refined`), propagate (entity-level N-S seed-and-propagate via provenance signatures; synthetic evaluation), graph_deanon
-- the construction/cost half (deferred — PAPER §9): `cost` (leak / amount-cut / topology leaf terms + the deferred `construction_cost`), `ancestry` (the absorber-model provenance target; feeds propagate), `report` (fuses the terms on a real tx), `subsetsum`/`coinjoin_demix` (the amount de-mix channel); consumes the `dense-subset-sum` engine (build: `maturin develop`); `cluster_refined` optionally refuses links when provenance and fingerprints diverge
-- chain-analysis channels that select what to ask and read the answer, all outside the engine:
-  `conservation` (what the other participants could not have funded — arithmetic on one transaction,
-  no client model), `provenance` (which inputs descend from known transactions), `monitor` (watches
-  tracked coins for the co-spend an intersection argument needs), `intersect` (the N-ary origin
-  intersection, handed to `cluster_refined` to score rather than asserted)
-- `PAPER.md` — the manuscript; `results/` — reproducible outputs; `catalog/`, `bigquery/`
+**The package.** `decluster/` holds the measurement half. Modules are named for the evidence channel
+they carry — `extractors`/`library`/`combiner` for construction fingerprints, `subsetsum`/`counting`/
+`subtransaction` for amounts, `ancestry`/`provenance`/`intersect` for provenance, `views`/
+`view_partition`/`contraction` for the coin graph — and they converge on two objects that travel in
+opposite directions across the partition lattice: `cluster.cluster_refined` ascends, declining a
+co-spend the merge-only heuristic would take, and `declustering.decluster` descends, cutting a
+partition it did not build. `decluster/baselines/` reimplements published algorithms against their papers,
+`decluster/experiments/` holds one module per canonical run, and `decluster/adaptations/` holds the pieces that are
+named adaptations rather than reproductions.
+
+**The evidence chain.** Every published number travels the same path, and each step is checkable:
+
+```
+catalog/datasets/   the data, pinned by digest
+        ↓
+decluster/experiments/   one module per run
+        ↓
+catalog/runs/   the manifest: command, parameters, environment, claim ids, verification
+        ↓
+results/artifacts/   the machine-written result
+        ↓
+results/generated/   the document rendered from it, never edited by hand
+        ↓
+releases/   the evidence bundle, listing every blob by name, size and digest
+        ↓
+artifacts/sha256/   the content-addressed store holding those bytes
+```
+
+`reproduction/` carries one environment and lockfile per bundle, and `sources/` a deterministic
+archive of the code at the revision a run names. `decluster-bundle sync` keeps tree, index and store
+in step; a gate fails if they drift. A run's `verify` recomputes its artifact and compares it byte
+for byte.
+
+**The prose.** `PAPER.md` is the manuscript. `results/` holds both halves of the record: documents
+generated from artifacts, and hand-written `RESULTS-*.md` reports that predate the chain.
+`results/REPRODUCIBILITY.md` files every one of them under an evidence state, and each document
+carries a footer naming its state or saying plainly that none has been assigned.
+
+**The rest.** `bigquery/` holds the extraction recipes, including the ones for collections the
+policy names as still missing. `catalog/ctp-claims.json` and `ctp-sources.json` bind each claim to
+the literature and to the code that implements it. `tests/` is written as gates rather than
+coverage: each file pins a property that has broken once, and says which in its docstring.
 
 Installable (`pip install -e .`), so a protocol-specific client model can consume these primitives
 from above without this repository knowing the protocol exists.
 
-Every number is reproducible. MIT — see `LICENSE`.
+Reproducibility is tracked per result. Every canonical run is bitwise reproducible from the committed store; the exact-oracle audit was the first, and bootstrapping any bundle looks the same:
+
+```console
+decluster-bundle --index releases/exact-oracle-evidence-v1.bundle.json \
+  --store artifacts --root /tmp/decluster-bundle bootstrap
+decluster-bundle --index releases/exact-oracle-evidence-v1.bundle.json \
+  --root /tmp/decluster-bundle --work /tmp/decluster-run reproduce
+```
+
+That environment currently requires CPython 3.13 on macOS arm64. The committed content-addressed
+store makes local cold-start execution possible. Public bootstrap remains pending until every blob
+has a canonical HTTPS location and an independent mirror. Other results retain the guarantees and
+limitations declared in `results/REPRODUCIBILITY.md` and their manifests.
+
+MIT. See `LICENSE`.

artifacts/sha256/00/9538122a9c35006d6ac47a1129317d504fd15ac6861dffaa3c115cc3edc872

@@ -0,0 +1,19 @@
+# N-S 2011 executable coverage
+
+Generated from the canonical experiment artifact. Do not edit manually.
+
+| component | status | limitation |
+|---|---|---|
+| `algorithm1_similarity` | `implemented` | — |
+| `algorithm2_positive_weights` | `implemented` | — |
+| `algorithm2_dummy_weights` | `implemented` | a dummy-incident node term is fixed at zero; the paper states this in prose, not in the printed formula |
+| `annealing` | `partial` | iteration budget, RNG and best-state return are explicit local controls |
+| `two_stage_mapping` | `implemented` | the paper picks an arbitrary unmapped node and stage 1 feeds back, so the ``repr`` order this driver imposes is a result-bearing local choice; injectivity comes from Algorithm 3's 1-1 mapping, not from the propagation section |
+| `confidence_pruning` | `not_reproduced` | an implementation-complete pruning policy is not available |
+| `accepted_mapping_correction` | `not_reproduced` | schedule and conflict resolution are not specified sufficiently |
+| `algorithm3_cascade` | `implemented` | — |
+| `learned_25_feature_model` | `not_reproduced` | the caller supplies an ML score; the published feature producer is absent |
+
+Every non-implemented component is exercised through the refusal gate. The implemented components pass that gate, but the end-to-end paper pipeline remains unreproduced.
+
+This deterministic contract does not reproduce the Flickr/Kaggle corpus, reported metrics, dummy-node convention, pruning policy, correction schedule, or learned model.

artifacts/sha256/04/f9670b4969de9df7ef526818503be374a6b71fa6c2df178db03e7d1c688bb3

@@ -0,0 +1,93 @@
+{
+  "schema_version": 1,
+  "id": "amount-channel-survey-v1",
+  "claim_ids": [
+    "ctp.payjoin.unnecessary_input",
+    "ctp.amount.subtransaction_ambiguity"
+  ],
+  "code": {
+    "revision": "bcd9a3e511860b1b0ff5f3ee6338a2df29ecbfbf",
+    "dirty": false
+  },
+  "command": {
+    "argv": [
+      "python",
+      "-m",
+      "decluster.experiments.amount_channel_summary",
+      "reproduce",
+      "--artifact",
+      "results/artifacts/amount-channel-survey-v1.json",
+      "--markdown",
+      "results/generated/amount-channel-survey-v1.md"
+    ]
+  },
+  "environment": {
+    "python": "3.13.7",
+    "lock_digest": "e0d6a8cada5d383c0aca2bd2a37f28c15721392469c94084128d369c29c5544d",
+    "platform": "macOS-26.6.2-arm64-arm-64bit-Mach-O",
+    "dependencies": []
+  },
+  "datasets": [],
+  "parameters": {
+    "composition": "identity-checked upstream artifacts only",
+    "components": {
+      "amount-local-channels-v1": "210ecb06ecaa32cc4cabe03ced8b85b0882a824dca2c361bd2c5dede7dcccf02",
+      "amount-subset-sum-v1": "240592b5af814fbcd28b06221805b934424170119da5e62e9da0281944377a9d",
+      "amount-per-coin-v1": "7be46858e221d0c942e7fa3ccc93b5eba28cf3a494fd00f07c2d50f2caa473df",
+      "amount-mapping-family-v1": "11f0ec645ea1be334dc7d24fbfc31047b234f335f48b5a514960472c7444b7d6",
+      "amount-pairwise-family-v1": "83f6a9e8501ee4df45f08b8c8ffc2f7265d40bf9add6d3f5093d712657c9dbf5"
+    }
+  },
+  "rng": {
+    "algorithm": null,
+    "seeds": []
+  },
+  "outputs": [
+    {
+      "path": "results/artifacts/amount-channel-survey-v1.json",
+      "bytes": 2589,
+      "sha256": "fcd4ad596d735b7c5a8dba43bc96481946ce5e78c3806932b60c10c865d1dcf8"
+    },
+    {
+      "path": "results/generated/amount-channel-survey-v1.md",
+      "bytes": 1082,
+      "sha256": "2f74b40fb24558a23f162d09da46a5f76a4434c170a78e0deb75eabba5d2ca28"
+    }
+  ],
+  "reproducibility": {
+    "level": "bitwise_reproducible",
+    "availability": "complete"
+  },
+  "verification": {
+    "mode": "exact",
+    "argv": [
+      "python",
+      "-m",
+      "decluster.experiments.amount_channel_summary",
+      "verify",
+      "--artifact",
+      "results/artifacts/amount-channel-survey-v1.json",
+      "--markdown",
+      "results/generated/amount-channel-survey-v1.md"
+    ],
+    "tests": [
+      "tests/test_amount_channel_summary_experiment.py"
+    ],
+    "properties": [
+      "all five component identities are checked before composition",
+      "the summary performs no mechanism calculation",
+      "channels remain separate rather than becoming a scalar score",
+      "DSS claims retain their restricted-family qualification",
+      "the generated Markdown equals the canonical rendering"
+    ],
+    "tolerance": null
+  },
+  "limitations": [
+    "this artifact composes upstream results and performs no mechanism calculation",
+    "component SHA-256 identities are mandatory",
+    "the five components measure different objects and are not combined into one score",
+    "DSS mapping and pairwise claims remain restricted to DSS's mapping family",
+    "the Bitcoin snapshot covers one 137-block interval from 2023",
+    "this summary is not a privacy certificate or CoinScore"
+  ]
+}

artifacts/sha256/06/ff8b5e8cbcbe3c4c47f6e5b6e5510b591526e40c31802240a33a05ef99ec4a

@@ -0,0 +1,13 @@
+{
+  "schema_version": 1,
+  "python": {
+    "implementation": "cpython",
+    "version": "3.13",
+    "system": "Darwin",
+    "machine": "arm64"
+  },
+  "source_archive": "sources/decluster-bcd9a3e-runtime.tar",
+  "project_root": "decluster",
+  "requirements": "reproduction/deterministic-partition/requirements.lock",
+  "dependency_directory": "reproduction/deterministic-partition"
+}

artifacts/sha256/08/376f7020b9d4d70c689748196cb83f7695f9b4e7b526702535e7644901039a

@@ -0,0 +1,13 @@
+{
+  "schema_version": 1,
+  "python": {
+    "implementation": "cpython",
+    "version": "3.13",
+    "system": "Darwin",
+    "machine": "arm64"
+  },
+  "source_archive": "sources/decluster-bcd9a3e-runtime.tar",
+  "project_root": "decluster",
+  "requirements": "reproduction/partition-schemes/requirements.lock",
+  "dependency_directory": "reproduction/partition-schemes"
+}