pragmatiq 0.1.0b1__tar.gz

pragmatiq-0.1.0b1/.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,54 @@
+name: Bug report
+description: Report a reproducible problem in pragmatiq.
+title: "[Bug]: "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for helping improve pragmatiq.
+
+        pragmatiq is an independent implementation inspired by the PRAGMA paper
+        (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.
+
+        Please do not include real customer data, secrets, or private banking records.
+  - type: textarea
+    id: summary
+    attributes:
+      label: Summary
+      description: What happened?
+    validations:
+      required: true
+  - type: textarea
+    id: reproduce
+    attributes:
+      label: Steps to reproduce
+      description: Include the smallest command or code snippet that reproduces the issue.
+      placeholder: |
+        1. Run ...
+        2. Observe ...
+    validations:
+      required: true
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected behavior
+    validations:
+      required: true
+  - type: textarea
+    id: environment
+    attributes:
+      label: Environment
+      description: OS, Python version, PyTorch version, install extras, CPU/GPU.
+      placeholder: |
+        - OS:
+        - Python:
+        - PyTorch:
+        - Install command:
+    validations:
+      required: true
+  - type: textarea
+    id: logs
+    attributes:
+      label: Logs or traceback
+      render: shell

pragmatiq-0.1.0b1/.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,45 @@
+name: Feature request
+description: Suggest an improvement or extension for pragmatiq.
+title: "[Feature]: "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for sharing an idea.
+
+        pragmatiq is an independent implementation inspired by the PRAGMA paper
+        (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.
+  - type: textarea
+    id: problem
+    attributes:
+      label: Problem
+      description: What workflow, research question, or user need would this solve?
+    validations:
+      required: true
+  - type: textarea
+    id: proposal
+    attributes:
+      label: Proposed solution
+      description: Describe the smallest useful version of the feature.
+    validations:
+      required: true
+  - type: dropdown
+    id: area
+    attributes:
+      label: Area
+      options:
+        - Synthetic data
+        - Tokenizer
+        - Model architecture
+        - Training or finetuning
+        - Inference or serving
+        - AML graph
+        - Documentation
+        - Other
+    validations:
+      required: true
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives considered

pragmatiq-0.1.0b1/.github/dependabot.yml

@@ -0,0 +1,14 @@
+version: 2
+updates:
+  - package-ecosystem: "pip"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+    groups:
+      python-dependencies:
+        patterns:
+          - "*"
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"

pragmatiq-0.1.0b1/.github/pull_request_template.md

@@ -0,0 +1,19 @@
+> pragmatiq is an independent implementation inspired by the PRAGMA paper
+> (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.
+
+## Summary
+
+- 
+
+## Verification
+
+- [ ] `pytest tests/ -x -q`
+- [ ] `ruff check . && mypy pragmatiq`
+- [ ] Relevant gate: `bash scripts/gates/gate_<N>.sh`
+
+## Notes
+
+- [ ] I updated docs/tests for user-visible behavior changes.
+- [ ] I did not add logic to `pragmatiq/cli.py`; CLI changes delegate to `pragmatiq/api.py`.
+- [ ] Randomized behavior has explicit seed handling.
+- [ ] New docs/assets have clear provenance and do not use Revolut trademarks, Revolut trade dress, or paper figures without attribution.

pragmatiq-0.1.0b1/.github/workflows/ci.yml

@@ -0,0 +1,113 @@
+name: ci
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+  workflow_dispatch:
+
+concurrency:
+  group: ci-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  quality:
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+          cache: pip
+      - name: Install (CPU torch)
+        run: |
+          python -m pip install --upgrade pip
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
+          pip install -e ".[dev]"
+      - name: Ruff
+        run: ruff check pragmatiq tests
+      - name: Mypy
+        run: mypy pragmatiq
+      - name: Pytest smoke
+        run: >
+          pytest
+          tests/test_models.py
+          tests/test_registry_cli.py
+          tests/test_tokenizer.py
+          tests/test_validate.py
+          tests/test_synthetic_world.py
+          tests/test_synthetic_labels.py
+          tests/test_synthetic_generate.py
+          tests/test_sharding_dataset.py
+          tests/test_gnn.py::TestGraphBuilder
+          tests/test_gnn.py::TestHandcrafted
+          tests/test_training.py::TestOptim
+          tests/test_training.py::TestPretrainer::test_checkpoint_contains_required_state
+          tests/test_training.py::TestPretrainer::test_tokenizer_mismatch_refused
+          tests/test_training.py::TestProbe::test_embed_records_notebook_api
+          -q
+          -x
+          --durations=10
+
+  python-compat:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    strategy:
+      matrix:
+        python-version: ["3.12"]
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - name: Install (CPU torch)
+        run: |
+          python -m pip install --upgrade pip
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
+          pip install -e ".[dev]"
+      - name: Compatibility smoke tests
+        run: >
+          pytest
+          tests/test_models.py
+          tests/test_registry_cli.py
+          tests/test_tokenizer.py
+          tests/test_validate.py
+          tests/test_synthetic_world.py
+          -q
+
+  gates:
+    # Run the CI-scale acceptance gates on pushes and pull requests, so a PR can
+    # never merge a red gate.
+    runs-on: ubuntu-latest
+    needs: [quality, python-compat]
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+          cache: pip
+      - name: Install
+        run: |
+          python -m pip install --upgrade pip
+          pip install torch --index-url https://download.pytorch.org/whl/cpu
+          pip install -e ".[dev,gnn,serve]"
+      - name: Gate 1 (synthetic data)
+        run: bash scripts/gates/gate_1.sh
+      - name: Gate 2 (tokenizer)
+        run: bash scripts/gates/gate_2.sh
+      - name: Gate 3 (sharding / padding-equivalence)
+        run: bash scripts/gates/gate_3.sh
+      - name: Gate 4 (model)
+        run: bash scripts/gates/gate_4.sh
+      - name: Gate 5 (training / resume / probe)
+        run: bash scripts/gates/gate_5.sh
+      - name: Gate 6 (AML GNN ablation)
+        run: bash scripts/gates/gate_6.sh
+      - name: Gate 7 (inference / serving)
+        run: bash scripts/gates/gate_7.sh
+      - name: Gate 8 (validate / nano e2e / packaging)
+        run: bash scripts/gates/gate_8.sh

pragmatiq-0.1.0b1/.github/workflows/docs.yml

@@ -0,0 +1,40 @@
+name: docs-site
+
+# Build the documentation site and verify its facts have not drifted from the code.
+on:
+  push:
+    branches: [main]
+    paths: ['website/**', 'scripts/docs_facts.py', 'scripts/docs_drift_check.py', 'pragmatiq/**']
+  pull_request:
+    paths: ['website/**', 'scripts/docs_facts.py', 'scripts/docs_drift_check.py', 'pragmatiq/**']
+
+jobs:
+  drift-check:
+    # facts.json must match what the library emits today (model sizes, config defaults,
+    # gate list, API surface), so the docs cannot silently fall out of sync with the code.
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+      - run: pip install -e ".[dev]"
+      - run: python scripts/docs_drift_check.py
+
+  build:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: website
+    steps:
+      - uses: actions/checkout@v4
+      - uses: pnpm/action-setup@v4
+        with:
+          version: 10
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: pnpm
+          cache-dependency-path: website/pnpm-lock.yaml
+      - run: pnpm install --frozen-lockfile
+      - run: pnpm build

pragmatiq-0.1.0b1/.github/workflows/release.yml

@@ -0,0 +1,33 @@
+name: publish
+
+# Build and publish pragmatiq to PyPI when a GitHub Release is published.
+#
+# Uses PyPI Trusted Publishing (OIDC) — no API token/secret required. One-time setup:
+# on https://pypi.org/manage/project/pragmatiq/settings/publishing/, add a trusted
+# publisher for owner `dynamiq-ai`, repo `pragmatiq`, workflow `release.yml`,
+# environment `pypi`. A pre-release version (e.g. 0.1.0b1) publishes as a pre-release.
+on:
+  release:
+    types: [published]
+
+jobs:
+  pypi:
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write # required for Trusted Publishing
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+      - name: Build sdist + wheel
+        run: |
+          python -m pip install --upgrade build
+          python -m build
+      - name: Verify metadata
+        run: |
+          python -m pip install --upgrade twine
+          twine check dist/*
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1

pragmatiq-0.1.0b1/.gitignore

@@ -0,0 +1,45 @@
+# Python
+__pycache__/
+*.py[cod]
+*.egg-info/
+.eggs/
+build/
+dist/
+.venv/
+venv/
+
+# Tooling caches
+.claude/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+
+# Project artifacts (anchored to repo root so they never shadow the
+# pragmatiq/data/ source package or pragmatiq/.../runs modules)
+/runs/
+/data/
+/outputs/
+*.parquet
+*.lmdb
+realism_report.html
+/wandb/
+
+# Notebooks
+.ipynb_checkpoints/
+
+# Secrets / local env — never commit
+.env
+.env.*
+!.env.example
+*.key
+secrets/
+
+# OS / editors
+.DS_Store
+.idea/
+.vscode/
+
+# Internal spec — kept locally, not published
+docs/SPEC.md

pragmatiq-0.1.0b1/AGENTS.md

@@ -0,0 +1,79 @@
+# pragmatiq — project memory
+
+pragmatiq is an independent open-source implementation inspired by PRAGMA
+(arXiv 2604.08649) by Dynamiq.
+The ONLY spec is **docs/SPEC.md** — read it from disk; never trust chat context
+over the file. Subagents must read docs/SPEC.md before reviewing anything.
+
+## Current phase
+
+> **current phase: 8 (polish). All gates 1-8 green at CI scale; gate-6 also
+> confirmed at full scale on GPU (12k accounts × 3 seeds: a≈0.50, c≈0.85). The
+> phase-6 AML ablation demonstrates relational recovery: money mules are ordinary
+> accounts whose only signal is ring membership in the transfer graph, so an
+> isolated embedding is at chance (a≈0.50) while a graph-aware model recovers them
+> (c≈0.85). On these structurally-distinctive synthetic rings the signal is
+> largely transfer-graph degree, so hand-crafted degree is a strong baseline and
+> (b)>(c) is not claimed. Gate 8 is the CPU pipeline smoke test; gate 5 is the
+> performance gate. See MODEL_CARD.md and notebooks/04.**
+
+Phases & gates: see docs/SPEC.md. Gate scripts live in `scripts/gates/gate_N.sh`.
+Never start phase N+1 with a red gate N.
+
+## Public code hygiene (non-negotiable)
+
+This repository is public, production-grade open source. Everything that ships in
+git — code comments, docstrings, commit messages, README, MODEL_CARD, CHANGELOG,
+notebooks, gate scripts — must read as a clean first version written by its
+authors, not as a record of how it was fixed:
+
+- Comments and docstrings explain WHAT the code does and its forward-looking design
+  rationale. They never describe prior behavior, bug history, or any review/audit
+  process — avoid "was X, now Y", "previously/old/used to", "fixed", "byte-identical
+  to before", an AUC a removed bug scored, "the audit/verification found …", etc.
+- Commit messages and the CHANGELOG describe v1 capabilities, not defect fixes.
+- This is v1: there is no prior release. Do NOT add backwards-compatibility shims,
+  legacy fallbacks, or "for older …" branches.
+- Keep working notes, edit history, and learnings under `outputs/` (gitignored) —
+  never in tracked files.
+
+## Commands
+
+```bash
+pip install -e ".[dev]"            # editable install + test deps
+pytest tests/ -x -q                # full suite
+pytest tests/test_<module>.py -q   # one module
+bash scripts/gates/gate_1.sh       # phase-1 acceptance gate (etc.)
+ruff check . && mypy pragmatiq     # lint + types
+```
+
+Gate scripts honor `PRAGMATIQ_GATE_FULL=1` for full-scale runs (100k users,
+8 cores); default is CI scale (small N, throughput extrapolation).
+
+## Non-negotiable global rules
+
+1. All logic lives in the library (`pragmatiq/`); `cli.py` (Typer) only parses args
+   and calls `pragmatiq/api.py` functions.
+1. Every randomized component takes an explicit seed; same seed → byte-identical
+   output (CI-enforced for the generator).
+1. Checkpoints store model + optimizer + LR scheduler + sampler position + RNG
+   states, and embed the tokenizer hash + resolved config. `from_pretrained()`
+   refuses to run with a mismatched tokenizer (clear error message).
+1. Unseen keys/values at inference map to [UNK] with a logged warning — never a
+   KeyError.
+1. Everything must run on CPU (slow but correct). CUDA paths are accelerations,
+   not requirements. Use flash-attn varlen if available, else fall back to SDPA
+   with an attention mask built from cu_seqlens.
+1. Write tests alongside each module; type-hint the public API; docstrings on
+   every public function.
+1. Public API layering: `pragmatiq/api.py` exposes `synthesize / tokenize / pretrain / finetune / embed / probe`; notebooks get
+   `PragmaModel.from_pretrained(run)` and `model.embed_records(list_of_dicts)`
+   (plain dicts, no shard pipeline needed interactively).
+1. Extension points via `pragmatiq/registry.py` decorators —
+   `@register_head(name)`, `@register_masker(name)`, `@register_value_encoder(name)`
+   — configs reference components by name so engineers customize without forking.
+
+## Attribution (must appear in every README/doc)
+
+> pragmatiq is an independent implementation inspired by the PRAGMA paper
+> (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.

pragmatiq-0.1.0b1/CHANGELOG.md

@@ -0,0 +1,49 @@
+# Changelog
+
+> pragmatiq is an independent implementation inspired by the PRAGMA paper
+> (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.
+
+All notable changes to pragmatiq are documented in this file. This project
+follows [Semantic Versioning](https://semver.org); 0.x releases are pre-1.0 and
+the public API may change.
+
+## [0.1.0b1] — Public beta
+
+First public (beta) release: an end-to-end, CPU-first toolkit for behavioral
+banking foundation models.
+
+### Added
+- **Synthetic data** — a deterministic, agent-based banking simulator (events,
+  profiles, transfers, and causal fraud / credit-default / AML / churn / LTV
+  labels with strict eval-point truncation), an HTML realism report, and
+  `synth calibrate` for fitting the generator to aggregate statistics.
+- **Tokenizer & data pipeline** — a key–value–time tokenizer (percentile-binned
+  numerics, categorical/BPE values, calendar + log-second time features, `[UNK]`
+  fallbacks), parquet sharding with an LMDB user index, and padding-free varlen
+  batching under a token budget.
+- **Model** — PRAGMA-style encoders with shared/tied key–value embeddings,
+  TimeRoPE continuous-time positions, per-event independent encoding, profile and
+  history encoders, and a tied MLM head; `small`/`medium`/`large` presets plus a
+  CPU-friendly `nano` size.
+- **Training & adaptation** — pretraining on Lightning Fabric (Muon + AdamW,
+  cosine schedule, fully resumable checkpoints), a gradient-boosting probe
+  (ROC-AUC + PR-AUC vs a same-classifier raw-count baseline; logistic/LightGBM
+  selectable), LoRA fine-tuning, and a registry for swappable heads, maskers, and
+  value encoders.
+- **Hands-off scale** — gradient accumulation, multi-node DDP, and a `config="auto"`
+  sizer that picks the batch and schedule from the data + device, so a run scales
+  from 1M to 26M records without tuning.
+- **PRAGMA+Nemotron variant** — an optional, switchable text pathway that embeds
+  high-cardinality text fields with a frozen text encoder and reconstructs them
+  with MSE; off by default, so the BPE path is byte-identical.
+- **AML over the transfer graph** — a GraphSAGE node classifier and the
+  three-way relational-recovery ablation (isolated embedding vs graph-aware node
+  features); see MODEL_CARD.md and `notebooks/04`.
+- **Inference & serving** — a batch embedder, integrated-gradients event
+  attribution, ONNX export, a Triton serving image that installs pragmatiq, a
+  turnkey `deploy_serving.sh`, monitoring (Prometheus + Grafana), and a Streamlit demo.
+- **Documentation** — a modern docs/educational site (Next.js + Fumadocs) at
+  pragmatiq.getdynamiq.ai, with interactive visualizers and a facts drift-check.
+- **Engineering** — CPU-first throughout (CUDA and flash-attn are accelerations),
+  a typed public API, notebooks, a model card, and CI (ruff + mypy + pytest +
+  acceptance gates).

pragmatiq-0.1.0b1/CITATION.cff

@@ -0,0 +1,34 @@
+cff-version: 1.2.0
+message: "If you use pragmatiq, please cite the software and the PRAGMA paper that inspired it."
+type: software
+title: "pragmatiq: an open-source PRAGMA-inspired banking foundation model stack"
+version: "0.1.0"
+license: Apache-2.0
+authors:
+  - name: "Dynamiq"
+    email: "support@getdynamiq.ai"
+repository-code: "https://github.com/dynamiq-ai/pragmatiq"
+url: "https://github.com/dynamiq-ai/pragmatiq"
+abstract: >-
+  pragmatiq is an independent implementation inspired by the PRAGMA paper
+  (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.
+references:
+  - type: article
+    title: "PRAGMA: Revolut Foundation Model"
+    authors:
+      - name: "Maxim Ostroukhov"
+      - name: "Ruslan Mikhailov"
+      - name: "Vladimir Iashin"
+      - name: "Artem Sokolov"
+      - name: "Andrei Akshonov"
+      - name: "Vitaly Protasov"
+      - name: "Dmitrii Beloborodov"
+      - name: "Vince Mullin"
+      - name: "Roman Yokunda Enzmann"
+      - name: "Georgios Kolovos"
+      - name: "Jason Renders"
+      - name: "Pavel Nesterov"
+      - name: "Anton Repushko"
+    doi: "10.48550/arXiv.2604.08649"
+    url: "https://arxiv.org/abs/2604.08649"
+    year: 2026

pragmatiq-0.1.0b1/CLAUDE.md

@@ -0,0 +1,79 @@
+# pragmatiq — project memory
+
+pragmatiq is an independent open-source implementation inspired by PRAGMA
+(arXiv 2604.08649) by Dynamiq.
+The ONLY spec is **docs/SPEC.md** — read it from disk; never trust chat context
+over the file. Subagents must read docs/SPEC.md before reviewing anything.
+
+## Current phase
+
+> **current phase: 8 (polish). All gates 1-8 green at CI scale; gate-6 also
+> confirmed at full scale on GPU (12k accounts × 3 seeds: a≈0.50, c≈0.85). The
+> phase-6 AML ablation demonstrates relational recovery: money mules are ordinary
+> accounts whose only signal is ring membership in the transfer graph, so an
+> isolated embedding is at chance (a≈0.50) while a graph-aware model recovers them
+> (c≈0.85). On these structurally-distinctive synthetic rings the signal is
+> largely transfer-graph degree, so hand-crafted degree is a strong baseline and
+> (b)>(c) is not claimed. Gate 8 is the CPU pipeline smoke test; gate 5 is the
+> performance gate. See MODEL_CARD.md and notebooks/04.**
+
+Phases & gates: see docs/SPEC.md. Gate scripts live in `scripts/gates/gate_N.sh`.
+Never start phase N+1 with a red gate N.
+
+## Public code hygiene (non-negotiable)
+
+This repository is public, production-grade open source. Everything that ships in
+git — code comments, docstrings, commit messages, README, MODEL_CARD, CHANGELOG,
+notebooks, gate scripts — must read as a clean first version written by its
+authors, not as a record of how it was fixed:
+
+- Comments and docstrings explain WHAT the code does and its forward-looking design
+  rationale. They never describe prior behavior, bug history, or any review/audit
+  process — avoid "was X, now Y", "previously/old/used to", "fixed", "byte-identical
+  to before", an AUC a removed bug scored, "the audit/verification found …", etc.
+- Commit messages and the CHANGELOG describe v1 capabilities, not defect fixes.
+- This is v1: there is no prior release. Do NOT add backwards-compatibility shims,
+  legacy fallbacks, or "for older …" branches.
+- Keep working notes, edit history, and learnings under `outputs/` (gitignored) —
+  never in tracked files.
+
+## Commands
+
+```bash
+pip install -e ".[dev]"            # editable install + test deps
+pytest tests/ -x -q                # full suite
+pytest tests/test_<module>.py -q   # one module
+bash scripts/gates/gate_1.sh       # phase-1 acceptance gate (etc.)
+ruff check . && mypy pragmatiq     # lint + types
+```
+
+Gate scripts honor `PRAGMATIQ_GATE_FULL=1` for full-scale runs (100k users,
+8 cores); default is CI scale (small N, throughput extrapolation).
+
+## Non-negotiable global rules
+
+1. All logic lives in the library (`pragmatiq/`); `cli.py` (Typer) only parses args
+   and calls `pragmatiq/api.py` functions.
+1. Every randomized component takes an explicit seed; same seed → byte-identical
+   output (CI-enforced for the generator).
+1. Checkpoints store model + optimizer + LR scheduler + sampler position + RNG
+   states, and embed the tokenizer hash + resolved config. `from_pretrained()`
+   refuses to run with a mismatched tokenizer (clear error message).
+1. Unseen keys/values at inference map to [UNK] with a logged warning — never a
+   KeyError.
+1. Everything must run on CPU (slow but correct). CUDA paths are accelerations,
+   not requirements. Use flash-attn varlen if available, else fall back to SDPA
+   with an attention mask built from cu_seqlens.
+1. Write tests alongside each module; type-hint the public API; docstrings on
+   every public function.
+1. Public API layering: `pragmatiq/api.py` exposes `synthesize / tokenize / pretrain / finetune / embed / probe`; notebooks get
+   `PragmaModel.from_pretrained(run)` and `model.embed_records(list_of_dicts)`
+   (plain dicts, no shard pipeline needed interactively).
+1. Extension points via `pragmatiq/registry.py` decorators —
+   `@register_head(name)`, `@register_masker(name)`, `@register_value_encoder(name)`
+   — configs reference components by name so engineers customize without forking.
+
+## Attribution (must appear in every README/doc)
+
+> pragmatiq is an independent implementation inspired by the PRAGMA paper
+> (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.

pragmatiq-0.1.0b1/CODE_OF_CONDUCT.md

@@ -0,0 +1,34 @@
+# Code of Conduct
+
+> pragmatiq is an independent implementation inspired by the PRAGMA paper
+> (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.
+
+## Our pledge
+
+We pledge to make participation in pragmatiq welcoming, respectful, and
+constructive for contributors and users regardless of background, identity,
+experience level, or affiliation.
+
+## Expected behavior
+
+- Be respectful and assume good intent.
+- Keep technical debate focused on evidence, reproducibility, and project goals.
+- Avoid sharing private customer data, credentials, or proprietary banking
+  records in issues, pull requests, discussions, or examples.
+- Credit prior work clearly, including the PRAGMA paper when discussing the
+  methods that inspired this implementation.
+
+## Unacceptable behavior
+
+- Harassment, threats, discriminatory language, or personal attacks.
+- Publishing private information without explicit permission.
+- Misrepresenting pragmatiq as affiliated with, endorsed by, sponsored by, or
+  officially connected to Revolut.
+- Uploading copyrighted third-party assets, Revolut trademarks, Revolut trade
+  dress, or copied paper figures without permission and attribution.
+
+## Enforcement
+
+Report concerns to `support@getdynamiq.ai`. Maintainers may remove content, close
+threads, restrict participation, or take other reasonable action to protect the
+project and its community.