falsifyai 0.1.0__tar.gz

falsifyai-0.1.0/.claude/CLAUDE.md

@@ -0,0 +1,144 @@
+# FalsifyAI — Project Context for Claude
+
+> Project-scoped instructions. Extends, does not replace, user-global `~/.claude/CLAUDE.md`.
+
+## What this project is
+
+**FalsifyAI** is a falsification-first reliability testing framework for AI systems. Status: **active Phase 0 implementation toward `falsifyai==0.1.0`**. Core pipeline is shipped (spec → materialize → execute → judge → save → CLI) with two dogfooded examples; remaining Phase 0 work in [plan.md §22.1](../plan.md).
+
+## Design philosophy (load-bearing)
+
+FalsifyAI optimizes for **evidence density over evidence volume**.
+
+```
+minimal meaningful evidence
++ high evidence quality per cognitive load
++ diverse perturbation categories
++ replayable proof
+= better falsification of AI / LLM systems
+```
+
+The goal is **maximum useful signal**, not maximum data. More evidence is not inherently better evidence.
+
+### Four pillars
+
+- **Minimal meaningful evidence.** Run the smallest experiment that meaningfully increases confidence in a verdict — no more. Adaptive evidence collection is the long-term ideal.
+- **High evidence quality per cognitive load.** Every line / artifact a user sees has to earn its real estate against: *would removing this make the engineer's decision worse?*
+- **Diverse perturbation categories (orthogonal pressure).** The admission criterion for a new perturbation family is *what new failure mode does this expose?* — not breadth. `typo_noise_v2` ≠ a new family; `paraphrase` is.
+- **Replayable proof.** Replay artifacts are the system's promise that claims are inspectable evidence, not anecdotes. CLI compresses; artifact preserves.
+
+### How this shapes decisions
+
+- **CLI output.** One row per case + one-line summary. Not a dashboard.
+- **Verdict design.** Compress evidence into actionable conclusions; don't enumerate it.
+- **Perturbation families.** Each must contribute orthogonal reliability information, not duplicate noise.
+- **Replay artifacts.** Self-contained; carry the full materialized spec so they outlive the YAML file on disk.
+- **MVP scope.** 2 perturbation families, 2 invariants, 5 verdicts — locked in [plan.md §22.1](../plan.md) because *that is enough to tell the story*.
+- **Three-layer architectural separation.** *Evidence generation* (perturbation / materialization / execution) is architecturally distinct from *evidence interpretation* (invariants / verdict resolver / CLI compression), and both are distinct from *evidence preservation* (replay artifacts / stores). New work belongs in exactly one layer; don't let interpretation leak into generation under pressure.
+- **Resolver complexity is bounded.** The verdict resolver is the epistemic authority of the framework; its priority chain must stay compressible and predictable. Expand the consumer surface (replay / diff / future tools) when adding interpretation features, not the verdict logic. The trust test for any resolver change: *a competent user should be able to predict the resolver output from the inputs.*
+
+### Anti-goals / anti-entropy infrastructure
+
+FalsifyAI is **not** optimizing for any of these. When pressure pulls toward them, resist:
+
+- Maximal perturbation volume
+- Maximal telemetry / metrics
+- Dashboard density
+- Benchmark quantity
+- Metric proliferation
+- Exhaustive output verbosity
+- Configuration knobs for every behavior
+- **Resolver inflation** — accreting heuristics, thresholds, verdict types, or confidence semantics into the verdict resolver. Each addition seems reasonable; cumulative effect destroys predictability.
+
+The signal to watch: *does this addition help an engineer make a better decision, or does it crowd the surface where the actual decision lives?* If the latter, defer or rework.
+
+## Naming (locked — do not change without confirmation)
+
+| Layer | Value |
+|---|---|
+| PyPI package | `falsifyai` |
+| Python import | `import falsifyai` |
+| CLI binary | `falsifyai` (e.g. `falsifyai run eval.yaml`) |
+| Brand / prose name | "FalsifyAI" |
+| Repo / folder | `falsifyai` |
+| Plugin entry-point groups | `falsifyai.perturbations`, `falsifyai.invariants`, `falsifyai.oracles`, `falsifyai.adapters`, `falsifyai.reporters`, `falsifyai.stores` |
+| Replay cache dir | `.falsifyai/` (matches CLI name, like `.git` / `.pytest_cache`) |
+
+**Background on the rename**: the original plan used `falsify` for the CLI binary, the `.falsify/` cache dir, and "Falsify" in prose. That collided with the existing `studio-11-co/falsify` project in the AI eval space. Renamed to `falsifyai` / `.falsifyai/` / "FalsifyAI" for full namespace consistency before any public release.
+
+## Toolchain
+
+- **Python:** 3.13+ (locked in `.python-version` and `pyproject.toml`)
+- **Package manager:** `uv` (not pip directly)
+- **Build backend:** `hatchling`
+- **Test:** `pytest` + `pytest-cov`
+- **Lint/format:** `ruff` (line-length 100, target py313)
+- **License:** Apache-2.0
+
+The `uv` binary lives at `C:\Users\Eric\AppData\Roaming\Python\Python313\Scripts\uv.exe`. PATH is configured. If a shell can't find `uv`, prepend that directory to `$env:PATH`.
+
+## Branch workflow
+
+- **Active development branch is `dev`.** Do not commit directly to `main`.
+- `main` is reserved for tagged releases and merged work. CI is gated on PRs to `main`.
+- Feature commits land on `dev` (or topic branches off `dev`); promote to `main` via PR when a milestone ships.
+- If you find yourself on `main` mid-session, switch to `dev` before staging changes.
+
+## Common commands
+
+```bash
+uv sync --extra dev          # install runtime + dev deps into .venv
+uv run pytest                # run tests
+uv run ruff check .          # lint
+uv run ruff format .         # format
+uv run python -c "import falsifyai; print(falsifyai.__version__)"
+```
+
+## Layout (flat, not src/)
+
+Package directory is at repo root, not under `src/`. See [plan.md §4](../plan.md). When the plan says `falsifyai/cli/main.py`, that means `<repo>/falsifyai/cli/main.py`.
+
+```
+falsifyai/                    ← repo root
+├── pyproject.toml
+├── falsifyai/                ← Python package
+│   ├── cli/  spec/  session/  perturbation/  execution/
+│   ├── invariants/  oracles/  statistical/  falsifiability/
+│   ├── verdict/  replay/  differential/  reporting/
+├── tests/
+│   ├── unit/  integration/  fixtures/  meta/
+└── examples/
+```
+
+All subpackages have empty `__init__.py` files only — no implementation yet.
+
+## Design anchors (when implementing, do not reinvent)
+
+- **8 verdicts in 2D space:** `STABLE`, `INFORMATION_PRESENT`, `CONSISTENTLY_WRONG`, `ADVERSARIALLY_VULNERABLE`, `FRAGILE`, `INFORMATION_NULL`, `AMBIGUOUS`, `INVALID_EVAL` — see [plan.md §2](../plan.md).
+- **Worst-case stratified stability**, not aggregate — see [plan.md §12](../plan.md).
+- **Spec materialization** separates intention (YAML) from instance (realized perturbations) — see [plan.md §8](../plan.md).
+- **Meta-oracle is the sole source of `INVALID_EVAL`** — see [plan.md §11.2](../plan.md).
+- **Perturbation validity is required** (bidirectional NLI default) — see [plan.md §9.3](../plan.md).
+- **`falsifyai diff` is a Phase 1 deliverable**, not Phase 2 — see [plan.md §14](../plan.md).
+- **Storage behind `ReplayStore` protocol** — SQLite default, no SQLite-specific code in core — see [plan.md §18](../plan.md).
+- **Falsifiability scoring is required** for every invariant — see [plan.md §15](../plan.md).
+
+## Scope discipline
+
+- **Phase 0 MVP is locked**: 3 weeks, single launch as `falsifyai==0.1.0`. See [plan.md §22.1](../plan.md). Includes `falsifyai diff`, `CONSISTENTLY_WRONG`, falsifiability scoring, and dogfooding from Week 1. Compression around the differentiator, not expansion of timeline.
+- **MVP verdict set**: `STABLE`, `FRAGILE`, `CONSISTENTLY_WRONG`, `INSUFFICIENT`, `INVALID_EVAL` (5 verdicts; full 8 in Phase 1).
+- **MVP perturbations**: `typo_noise` + `casing_variant` only (2 families — required for honest bootstrap CI).
+- **MVP invariants**: `contains` + `semantic_equivalence`.
+- **8-item acceptance gate** ([plan.md §22.1.1](../plan.md)) must pass before tagging 0.1.0. PyPI publication is deployment, not validation.
+- Do not add features beyond what the spec demands. Do not invent abstractions for hypothetical extensions.
+- Do not change naming without explicit user confirmation.
+- Do not deviate from the flat package layout without asking.
+- Cuts from MVP that may feel tempting: rich/colored terminal output (defer), heavyweight NLI for ConsistencyOracle (use embeddings for MVP, NLI in Phase 1), full 8-verdict resolver (5 verdicts for MVP).
+
+## What to NOT do
+
+- Don't add `src/` layout.
+- Don't add a `setup.py` or `setup.cfg`. `pyproject.toml` is the only build config.
+- Don't install pytest/ruff via pip directly — use `uv add --dev`.
+- Don't pre-create files for sections of the plan that aren't being implemented yet. Empty `__init__.py` is the current correct state.
+- Don't enable the CLI script entry-point in `pyproject.toml` until `falsifyai/cli/main.py` actually exists.

falsifyai-0.1.0/.claude/settings.json

@@ -0,0 +1,23 @@
+{
+  "$schema": "https://json.schemastore.org/claude-code-settings.json",
+  "permissions": {
+    "allow": [
+      "PowerShell(uv:*)",
+      "PowerShell(uv sync*)",
+      "PowerShell(uv run:*)",
+      "PowerShell(uv add:*)",
+      "PowerShell(uv lock*)",
+      "PowerShell(uv tree*)",
+      "PowerShell(uv pip list*)",
+      "PowerShell(py -m uv:*)",
+      "PowerShell(py --version)",
+      "PowerShell(py -c:*)",
+      "Bash(uv:*)",
+      "Bash(uv sync*)",
+      "Bash(uv run:*)",
+      "Bash(uv add:*)",
+      "Bash(uv lock*)",
+      "Bash(uv tree*)"
+    ]
+  }
+}

falsifyai-0.1.0/.claude/skills/pr-review/SKILL.md

@@ -0,0 +1,93 @@
+---
+name: pr-review
+description: Use this skill before committing, pushing, opening a PR, merging, or starting implementation from a locked plan in the FalsifyAI repo. It performs a pre-flight self-review against the three-layer architecture, evidence-density principle, resolver-inflation guardrail, replay-preservation expectations, dogfood/example requirements, and release/readiness gates.
+---
+
+# pr-review — FalsifyAI pre-flight self-review
+
+This skill runs **before** a destination-bound action: a commit, a push, opening a PR, merging a PR, or starting implementation from a locked plan. It is not a code-review pass on someone else's work — it is *your* checklist for whether the change you're about to ship clears FalsifyAI's architectural gates.
+
+## STOP clause (load-bearing)
+
+**If any gate below fails, stop.** Do not commit. Do not push. Do not continue implementation. Surface the failing gate to the user verbatim and ask whether to:
+
+1. **Split** the change (most common — usually means the PR is touching multiple layers),
+2. **Revise** the change to clear the gate, or
+3. **Explicitly accept the risk** (rare; requires the user to name what they're accepting).
+
+A skill that does not stop on failure is decoration. This one stops.
+
+## When to invoke
+
+Auto-invoke when the immediate intent is clearly:
+
+- about to `git commit` or `git push`
+- about to open a PR (`gh pr create`)
+- about to merge a PR
+- about to start implementation from a plan the user has approved
+
+Do **not** auto-invoke for:
+
+- casual discussion, brainstorming, or design exploration
+- README copy edits, typo fixes, doc-only formatting
+- exploratory reads / Q&A about the codebase
+
+## The six gates
+
+For each gate: state the answer in one sentence. If unclear or "no," that is a failure — stop and surface it.
+
+### Gate 1 — Which layer does this touch?
+
+FalsifyAI separates **evidence generation** (perturbation / materialization / execution) from **evidence interpretation** (invariants / verdict resolver / CLI compression) from **evidence preservation** (replay artifacts / stores). See [`docs/ARCHITECTURE.md`](../../../docs/ARCHITECTURE.md) and [`.claude/CLAUDE.md`](../../CLAUDE.md#design-philosophy-load-bearing).
+
+**Answer in one of**: generation, interpretation, preservation, consumer surface (CLI / diff / replay), contributor infrastructure, maintainer infrastructure.
+
+**Precision note** — these classifications are load-bearing; don't blur them:
+
+- *Preservation* is the **replay artifact / store** system only. It is not "things that get committed and persist."
+- README / CHANGELOG / CONTRIBUTING / `.github/` templates → **consumer surface** (what a user or contributor reads on arrival).
+- `docs/ARCHITECTURE.md` / `docs/RELEASE.md` / `docs/DEMO.md` / `dev_notes/*` → **contributor infrastructure** (read by people who change the code).
+- `.claude/skills/*` / `.claude/CLAUDE.md` / maintainer tooling → **maintainer infrastructure** (read by future-you while operating the project).
+
+### Gate 2 — Does it touch more than one layer?
+
+If yes: should it be split into separate commits or PRs? Cross-layer changes are the most common source of architectural drift. The default answer is *split it*; the exception requires a one-line justification.
+
+### Gate 3 — Does it inflate the resolver?
+
+The verdict resolver is the epistemic authority of the framework. Its priority chain must stay compressible and predictable.
+
+**Trust test** (authoritative copy in [`CONTRIBUTING.md`](../../../CONTRIBUTING.md)): *A competent user should be able to predict the resolver output from the inputs.*
+
+If this change adds heuristics, thresholds, new verdict types, new confidence semantics, new knobs, or new metrics that the resolver consults — the trust test must still pass after the change. If it does not, the work belongs in the **consumer surface** (replay, diff, future tools), not the resolver.
+
+### Gate 4 — Evidence density or evidence volume?
+
+FalsifyAI optimizes for **evidence density**, not volume. See the four pillars in [`.claude/CLAUDE.md`](../../CLAUDE.md#four-pillars).
+
+Two sub-checks, in order:
+
+1. **Would removing this output / field / row make the engineer's decision worse?** If no, the addition is volume — cut it.
+2. **Does this addition crowd the decision surface, or does it sit behind it?** The *decision surface* is where the user actually makes a call (CLI output, exit codes, the README hook, the verdict table). Volume that lives *behind* the decision surface (architecture docs, release runbook, internal notes) is acceptable when each item earns its keep alone. Volume that *crowds* the decision surface is not. When a new doc / row / field is borderline, ask which side it lives on.
+
+### Gate 5 — Are replay artifacts preserved, not recomputed?
+
+Replay is read-only. Verdicts shown by `falsifyai replay` are the ones assigned at run time and never re-resolved. If this change reads from a stored artifact and then re-judges, re-resolves, or recomputes a verdict — that is a preservation violation. Move the logic to *run-time* (write path) or to a new consumer surface command that is explicitly not `replay`.
+
+### Gate 6 — Examples dogfooded if user-facing behavior changed?
+
+If this change alters CLI behavior, spec language, verdict semantics, or output shape: at least one example under [`examples/`](../../../examples/) and the matching dogfood test in [`tests/integration/test_examples.py`](../../../tests/integration/test_examples.py) must demonstrate the new behavior end-to-end. Examples are the canonical user-facing spec surface — if the parser or resolver no longer accepts something the examples use, CI must fail immediately.
+
+## How to surface a failing gate
+
+When stopping, write one short paragraph in this shape:
+
+> **Stopping before [commit/push/PR/implementation].** Gate N (*one-line gate name*) failed: *one sentence on why*. Options: split / revise / accept-risk. Which?
+
+Do not enumerate every gate that passed. Surface only the failing one. The user already knows what the gates are.
+
+## Scope notes
+
+- This skill is a pre-flight self-check, not a substitute for the `code-reviewer` agent on substantive changes — invoke that separately when the change is non-trivial.
+- This skill is project-specific to FalsifyAI. Generic code-quality concerns (file size, naming, error handling) are covered by user-global rules and are not duplicated here.
+- Authoritative philosophy lives in [`.claude/CLAUDE.md`](../../CLAUDE.md), authoritative architecture in [`docs/ARCHITECTURE.md`](../../../docs/ARCHITECTURE.md), authoritative resolver trust test in [`CONTRIBUTING.md`](../../../CONTRIBUTING.md). If a gate here drifts from those docs, the docs win — update the skill, not the doc.

falsifyai-0.1.0/.env.example

@@ -0,0 +1,56 @@
+# FalsifyAI environment variables (template)
+#
+# Copy this file to `.env` or `.env.local` and fill in the values for the
+# provider(s) you actually use. Both `.env` and `.env.local` are gitignored.
+#
+# FalsifyAI does NOT auto-load these files. LiteLLM (the model adapter
+# layer) reads from process environment variables directly. To use this
+# template, load the values into your shell environment first:
+#
+# ─── bash / zsh ─────────────────────────────────────────────────────────
+#
+#   cp .env.example .env.local
+#   # edit .env.local with real values
+#   set -a; source .env.local; set +a
+#   falsifyai run examples/model_migration.yaml
+#
+# ─── PowerShell ─────────────────────────────────────────────────────────
+#
+#   cp .env.example .env.local
+#   # edit .env.local with real values
+#   Get-Content .env.local | ForEach-Object {
+#     if ($_ -match '^([A-Z_][A-Z0-9_]*)=(.*)$') {
+#       Set-Item "env:$($Matches[1])" $Matches[2]
+#     }
+#   }
+#   falsifyai run examples/model_migration.yaml
+#
+# ─── Or just set inline (no .env.local needed) ──────────────────────────
+#
+#   # bash/zsh
+#   OPENAI_API_KEY=sk-... falsifyai run examples/model_migration.yaml
+#
+#   # PowerShell
+#   $env:OPENAI_API_KEY = "sk-..."; falsifyai run examples/model_migration.yaml
+#
+# Or use `direnv` (Unix) for automatic per-directory loading.
+
+# ─────────────────────────────────────────────────────────────────────────
+# OpenAI (used by the README walkthrough + most examples)
+# ─────────────────────────────────────────────────────────────────────────
+OPENAI_API_KEY=
+
+# Optional: override the default model used in spec files.
+# OPENAI_MODEL=gpt-4o-mini
+
+# ─────────────────────────────────────────────────────────────────────────
+# Anthropic
+# ─────────────────────────────────────────────────────────────────────────
+# ANTHROPIC_API_KEY=
+
+# ─────────────────────────────────────────────────────────────────────────
+# Other providers
+# ─────────────────────────────────────────────────────────────────────────
+# LiteLLM supports 100+ providers. See https://docs.litellm.ai/docs/providers
+# for the full list and the env var name each one expects (e.g.,
+# GOOGLE_API_KEY for Gemini, COHERE_API_KEY for Cohere, GROQ_API_KEY, etc.).

falsifyai-0.1.0/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,52 @@
+---
+name: Bug report
+about: Something FalsifyAI did unexpectedly. Include a replay session id if possible.
+title: '[bug] '
+labels: bug
+---
+
+## What happened
+
+<!-- Brief description of the unexpected behavior. -->
+
+## Expected behavior
+
+<!-- What you thought should happen instead. -->
+
+## Reproduction
+
+<!-- Minimal spec / command sequence that triggers the issue.
+     If possible, paste the YAML spec inline. -->
+
+```yaml
+# your spec here
+```
+
+```bash
+$ falsifyai run ...
+# observed output
+```
+
+## Replay session id (high-signal!)
+
+<!-- The unique value-add: if the bug shows up in a real run, the saved
+     replay session contains EVERYTHING needed to reproduce. -->
+
+- **Session id:** `<paste the session_id printed at the end of `falsifyai run`>`
+- **Store path:** `<usually .falsifyai/replays.db>`
+- Confirm you're OK sharing the artifact contents (model outputs may
+  contain sensitive prompts/responses).
+
+If you can attach the `.falsifyai/replays.db` file (or a sanitized copy),
+add it to the issue. That's the deepest reproduction we can ask for.
+
+## Environment
+
+- **FalsifyAI version:** <`falsifyai --version` or `python -c "import falsifyai; print(falsifyai.__version__)"`>
+- **Python version:** <`python --version`>
+- **OS:** <macOS / Linux / Windows + version>
+- **Model provider + model:** <e.g., openai/gpt-4o-mini>
+
+## Additional context
+
+<!-- Anything else: workarounds you tried, related issues, etc. -->

falsifyai-0.1.0/.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,47 @@
+---
+name: Feature request
+about: Suggest a new feature, perturbation, invariant, or workflow.
+title: '[feature] '
+labels: enhancement
+---
+
+## Use case
+
+<!-- What you're trying to do that FalsifyAI doesn't currently support.
+     Concrete scenarios beat abstract requests. -->
+
+## Why current FalsifyAI doesn't cover it
+
+<!-- Briefly: which existing feature is closest, and why doesn't it fit? -->
+
+## Proposed surface (if you have one in mind)
+
+<!-- CLI command, spec field, output format, etc. Rough is fine. -->
+
+```bash
+# example invocation
+falsifyai ...
+```
+
+```yaml
+# example spec extension
+```
+
+## Alternatives considered
+
+<!-- Other approaches you thought about and rejected, with brief reasons.
+     This is especially useful for resolver / verdict changes (see
+     CONTRIBUTING.md on why the resolver complexity is bounded). -->
+
+## Layer
+
+<!-- Which architectural layer would this touch?
+     - generation (perturbation / materialize / execute)
+     - interpretation (invariants / verdict / falsifiability / render)
+     - preservation (replay / artifact / store)
+     - consumer (new CLI subcommand reading existing data)
+     If it touches more than one, decomposition might be in order. -->
+
+## Additional context
+
+<!-- Links to similar features in other tools, prior discussion, etc. -->

falsifyai-0.1.0/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,60 @@
+<!--
+Thanks for the PR! This template mirrors FalsifyAI's local dev_notes
+summary format. Fill out what's relevant; delete what's not.
+
+For non-trivial changes, see CONTRIBUTING.md for the architectural
+constraints (especially: resolver complexity is bounded; three-layer
+separation is non-negotiable).
+-->
+
+## Headline
+
+<!-- One sentence: what does this PR do? -->
+
+## Problem pressure
+
+<!-- 1-2 sentences: what gap does this close? Why now? -->
+
+## Abstraction shipped
+
+<!-- The new contract / Protocol / module / behavior, named explicitly. -->
+
+## Alternatives rejected
+
+<!-- Bullet list, one line each, with one-line reasoning per alternative.
+     High-signal for future engineers who hit the same decision fork. -->
+
+-
+
+## Architectural invariants
+
+<!-- System-level contracts this PR establishes or preserves. NOT coding
+     style. If this PR touches the verdict resolver, include an
+     explicit answer to the trust test from CONTRIBUTING.md:
+     "Can a competent user still predict the resolver output from the
+     inputs?" -->
+
+-
+
+## Test plan
+
+<!-- - [x] specific tests added
+     - [ ] manual smoke
+     - [ ] `uv run pytest` passes
+     - [ ] `uv run ruff check . && uv run ruff format --check .` clean
+     - [ ] CI green on `dev`
+     - [ ] CI green on PR target `main`
+-->
+
+- [ ] `uv run pytest` passes
+- [ ] `uv run ruff check .` clean
+- [ ] `uv run ruff format --check .` clean
+
+## Architectural fit (self-check)
+
+- [ ] Touches exactly **one** of the three layers (generation /
+  interpretation / preservation), or is a pure consumer.
+- [ ] If touching `falsifyai/verdict/resolver.py`: the trust test still
+  passes (a competent user can predict the output from the inputs).
+- [ ] Does not introduce new spec language fields, verdict types, or
+  configurable thresholds without a separate architectural conversation.

falsifyai-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,42 @@
+name: CI
+
+on:
+  push:
+    branches: [main, dev]
+  pull_request:
+    branches: [main, dev]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  lint-and-test:
+    name: Lint + test (Python 3.13, Linux)
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+        with:
+          version: "0.11.15"
+          enable-cache: true
+          cache-dependency-glob: "uv.lock"
+
+      - name: Set up Python 3.13
+        run: uv python install 3.13
+
+      - name: Sync dependencies
+        run: uv sync --extra dev --frozen
+
+      - name: Ruff lint
+        run: uv run ruff check .
+
+      - name: Ruff format check
+        run: uv run ruff format --check .
+
+      - name: Pytest
+        run: uv run pytest -v

falsifyai-0.1.0/.gitignore

@@ -0,0 +1,72 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# Virtual environments
+.venv/
+venv/
+env/
+ENV/
+
+# Testing / coverage
+.pytest_cache/
+.coverage
+.coverage.*
+htmlcov/
+.tox/
+.nox/
+coverage.xml
+*.cover
+*.py,cover
+
+# Type checking
+.mypy_cache/
+.pyright/
+.pytype/
+
+# Linting
+.ruff_cache/
+
+# FalsifyAI replay artifacts
+.falsifyai/
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo
+*~
+
+# OS
+.DS_Store
+Thumbs.db
+desktop.ini
+
+# Secrets / env
+.env
+.env.local
+*.pem
+*.key
+
+# Notebooks
+.ipynb_checkpoints/

falsifyai-0.1.0/.python-version

@@ -0,0 +1 @@
+3.13