modelpin 0.1.0__tar.gz

modelpin-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Samarth Puthraya
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

modelpin-0.1.0/PKG-INFO

@@ -0,0 +1,380 @@
+Metadata-Version: 2.4
+Name: modelpin
+Version: 0.1.0
+Summary: Dependabot for AI models — catch model-migration regressions before they hit production.
+Author: Samarth Puthraya
+License: Apache-2.0
+Project-URL: Homepage, https://modelpin.dev
+Project-URL: Repository, https://github.com/samarthputhraya/modelpin
+Project-URL: Issues, https://github.com/samarthputhraya/modelpin/issues
+Keywords: llm,ai,agents,regression-testing,model-pinning,evals,ci,dependabot
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Software Development :: Quality Assurance
+Classifier: Topic :: Software Development :: Testing
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Environment :: Console
+Requires-Python: >=3.12
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: typer>=0.12
+Requires-Dist: rich>=13.7
+Requires-Dist: pydantic>=2.7
+Requires-Dist: pyyaml>=6.0
+Provides-Extra: providers
+Requires-Dist: openai>=1.40; extra == "providers"
+Requires-Dist: google-genai>=1.0; extra == "providers"
+Requires-Dist: anthropic>=0.34; extra == "providers"
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: ruff>=0.5; extra == "dev"
+Requires-Dist: black>=24.0; extra == "dev"
+Dynamic: license-file
+
+# Modelpin
+
+**Dependabot for AI models.** Know before the model breaks you.
+
+A provider ships a new model (or retires the one you depend on). Modelpin **replays your app's
+real behavior** on the new model, decides whether anything *actually* regressed despite model
+randomness, and posts a PR-style report — so you find out in a pull request, not in production.
+
+CLI: `modelpin` (alias `mp`). License: Apache-2.0.
+
+---
+
+## Why this exists (and why you can trust the verdict)
+
+Models are non-deterministic. Run the same prompt twice and the words change. So the naive way to
+"test a new model" — diff the text — cries wolf on every run. An alerter that cries wolf is worse
+than no alerter: you mute it, then it misses the real break.
+
+Modelpin's entire design optimizes for **one north-star metric: false-positive rate.** The promise
+is narrow and falsifiable: *if Modelpin says it broke, it broke.* Everything below is in service of
+that promise — and where the evidence is thin, this README says so plainly.
+
+This is meant to be the **independent, no-BS** tool: it measures behavior change **relative to your
+app**, it never declares one model globally "better," and the whole harness is open source so you
+can reproduce it and disagree.
+
+---
+
+## Quickstart
+
+Install from source (not on PyPI yet — see [Install](#install)):
+
+```bash
+git clone https://github.com/samarthputhraya/modelpin
+cd modelpin
+pip install -e ".[dev,providers]"
+```
+
+### Try it offline, no API key (30 seconds)
+
+Modelpin ships a fake provider that replays canned traces, so you can see the whole pipeline —
+baseline, candidate replay, behavioral diff, report — with zero cost and no key:
+
+```bash
+mp baseline --provider fake --fixtures examples/traces/demo_traces.json \
+  --model claude-opus-4-6 \
+  --scenarios-dir examples/scenarios --config examples/modelpin.yaml
+
+mp check --to claude-opus-4-7 --from claude-opus-4-6 \
+  --provider fake --fixtures examples/traces/demo_traces.json \
+  --scenarios-dir examples/scenarios --config examples/modelpin.yaml
+```
+
+You'll get a per-scenario verdict (`unchanged` / `changed_minor` / `regression`), a confidence
+score, a one-line plain-English explanation per scenario, and a Markdown report written to
+`.modelpin/last-report.md`. `mp check` exits non-zero **only** on a real `regression` — that's the
+CI gate.
+
+### The real flow, on your own app
+
+```bash
+# 1. Scaffold modelpin.yaml + scenarios/ (never overwrites existing files)
+mp init
+
+# 2. See which models your repo already depends on, and where
+mp scan
+
+# 3. Add a scenario or two (a JSON file per representative case — see below),
+#    then record how your current model behaves, N times
+export OPENAI_API_KEY=sk-...        # your key, read from the env — never stored
+mp baseline                         # uses models[0] + providers[0] from modelpin.yaml
+
+# 4. Replay your scenarios on a candidate model and diff the behavior
+mp check --to gpt-5.5
+```
+
+A scenario is a small JSON file (one per case) under `scenarios/`. The one `mp init` writes:
+
+```json
+{
+  "id": "greeting",
+  "name": "Simple greeting",
+  "kind": "single",
+  "input": {"messages": [{"role": "user", "content": "Say hello in one short sentence."}]},
+  "assertions": {"must_contain": ["hello"]}
+}
+```
+
+Scenarios can also be agent runs: set `"kind": "agent"`, add `"tools"` (and canned `"tool_results"`)
+to `input`, and Modelpin drives a multi-turn model↔tool loop so trajectories like
+`lookup_order → issue_refund` actually emerge during replay. Eight worked examples spanning tool
+trajectories, semantic equivalence, refusals, and output format live in
+[`examples/suite/`](examples/suite/).
+
+---
+
+## See it in your PR (GitHub Action)
+
+The point of Modelpin is that the answer shows up **at review time.** It ships a real composite
+GitHub Action: it installs Modelpin, optionally records a baseline, runs `mp check`, posts a
+**sticky PR comment** (found-and-updated in place via a hidden marker — no comment spam), and
+**fails the job on a regression**. Drop this at `.github/workflows/modelpin.yml`:
+
+```yaml
+name: Modelpin
+
+on:
+  pull_request:
+  workflow_dispatch:        # trigger by hand the day a provider ships a new model
+
+permissions:
+  contents: read
+  pull-requests: write      # so the action can post/update the PR comment
+
+jobs:
+  model-check:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: samarthputhraya/modelpin/actions@v1
+        with:
+          from: gpt-4o-mini       # the model you depend on today (committed baseline)
+          to: gpt-5.5             # the candidate to vet before adopting
+          provider: openai
+          runs: "5"
+        env:
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}     # BYO-key from repo secrets — never inline a key
+          # If your judge_model lives on another provider, add its key too:
+          # GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
+          # GROQ_API_KEY:   ${{ secrets.GROQ_API_KEY }}
+```
+
+Action inputs: `to` (required), `from`, `provider`, `config`, `scenarios-dir`, `runs`, `match`,
+`baseline`, `comment`, `fail-on-regression`, `github-token`, `modelpin-spec`, `python-version`,
+`working-directory`. Outputs: `verdict-exit-code` and `report-path`. The usual pattern is to
+**commit your baseline** so CI only replays the candidate; flip `baseline: "true"` to record fresh
+(needs the old model still reachable). Copy-paste workflow:
+[`examples/github-workflow.yml`](examples/github-workflow.yml).
+
+---
+
+## How the behavioral diff works (the moat)
+
+Modelpin decides "did it really change?" from **multiple signals over multiple runs**, then gates
+every regression behind a **distributional significance test plus an effect-size floor**. A single
+odd run never trips it; a majority that merely flips between two equally-likely behaviors never
+trips it. Here is the whole decision rule, no hand-waving:
+
+**1. Multi-run, not single-shot.** Each scenario runs N times (`runs:` in config; default 5,
+minimum 2 — a single run can't form a distribution, so `--runs 1` is rejected outright). Baseline
+*and* candidate are both sampled, so the comparison is distribution-vs-distribution.
+
+**2. Structural signals** (per run, no network, deterministic):
+- **Tool-call trajectory match** with four modes — `strict | unordered | subset | superset`
+  (`--match`) — so you choose how strict "same plan" means for your agent.
+- **Output format / assertion validity** — your scenario's `must_contain` / `must_not_contain`
+  text assertions, checked as a rate across runs.
+- **Refusal detection** — did the model start declining requests it used to answer?
+- **Latency / token deltas** — captured and reported, but **informational only**; they never gate
+  the verdict (latency is jittery; a token bump isn't a behavior regression).
+
+**3. Semantic signal** (optional LLM-as-judge): a low-temperature judge answers the only question
+that matters — *do these answers mean / accomplish the same thing?* This catches the structural
+blind spot: two answers that are textually different but identical in meaning ("The total is $5."
+vs "5 dollars."). The judge is **injected and optional** — with no `judge_model` set (and always on
+the offline `fake` path) the diff stays purely structural and makes zero network calls, so CI can
+run for $0. The judge is independent of the two models being compared, so it can arbitrate a
+cross-vendor check.
+
+**4. The statistics that kill false alarms.** Every gating signal goes through an **exact
+two-sample permutation test** (`modelpin/diff/stats.py` — no SciPy, deterministic, so golden tests
+stay reproducible). A signal counts as a regression only when **both**:
+- the candidate distribution differs from baseline at **p ≤ 0.05** (`ALPHA`), **and**
+- the effect clears a conservative **size floor** — tool-call shift ≥ 0.5 total-variation distance
+  (`MIN_TOOL_TVD`), refusal-rate rise ≥ 0.34 (`MIN_REFUSAL_DELTA`), or semantic-divergence rate
+  ≥ 0.5 over baseline (`MIN_SEMANTIC_DELTA`).
+
+The size floor is what stops a *statistically* significant but *practically* trivial jitter from
+firing once N grows large. These floors are intentionally conservative — biased toward missing a
+borderline change rather than inventing one — because a miss is a false *negative* (the safe
+direction for a trust product), while a false alarm erodes trust permanently.
+
+**Output:** each scenario gets a verdict, a confidence score, the underlying signals, and a one-line
+explanation. A structural tool-call / refusal break or a calibrated semantic divergence is a
+**CI-failing `regression`**; format/assertion drift alone is `changed_minor` (reported, doesn't fail
+the build).
+
+### The false-positive evidence — and its limits, stated plainly
+
+**Result: 0/8 false positives** on a held-out 8-scenario suite (a model judged against *itself*,
+judge on, all `unchanged` at confidence 1.00). On that same run, the two perturbations that
+genuinely changed behavior were caught, and one prompt-injection the model resisted was correctly
+left `unchanged` (not a false negative). Corroborated by additional same-model and cross-model
+splits. Full writeup: [`docs/fp-measurement.md`](docs/fp-measurement.md).
+
+The semantic judge's escalation threshold is **calibrated** on a labeled set in
+[`examples/calibration/`](examples/calibration/) that is **deliberately distinct from the held-out
+suite** (so it can't leak into the 0/8 number): equivalent-but-reworded pairs land at divergence
+0.0, real meaning changes at ≥ 0.8, leaving an empty gap around the 0.5 floor — 0 false positives.
+FP-safety was re-checked with an **independent judge** (a different model arbitrating) and
+re-validated on the held-out suite after promoting semantic divergence from `changed_minor` to a
+CI-failing `regression` (still 0/8).
+
+**This is a first calibration. Do not over-trust it.** The honest limitations, documented in
+[`docs/STATUS.md`](docs/STATUS.md):
+- the calibration set is **small** (≈6+6 pairs) and the perturbations are **synthetic**, not
+  harvested from real migrations;
+- recall on subtle changes was 4/6 — it can *miss* a subtle real change (again, the safe direction);
+- the judge is **OpenAI-only** so far;
+- the structural floors are FP-validated by the held-out suite but **not yet swept** on a labeled
+  set.
+
+Planned before any high-stakes reliance: ≥30 pairs including real migration traces, and a
+non-OpenAI judge. We'd rather you know this than discover it.
+
+---
+
+## Cross-vendor (including a free third vendor)
+
+A model migration isn't always within one lab. Modelpin diffs **across vendors** through one engine;
+a separate judge model arbitrates meaning-equivalence.
+
+| Provider | Status |
+|---|---|
+| **OpenAI** | Live (Chat Completions), multi-turn tool loops |
+| **Google / Gemini** | Live (`google-genai`), multi-turn tool loops, cross-vendor proven |
+| **OpenAI-compatible hosts** — `groq`, `openrouter`, `together`, `cerebras` | Live (the OpenAI adapter pointed at the host's `base_url`) |
+| **Anthropic** | **Stub** — raises `NotImplementedError` (deferred until a paid key is in play) |
+
+**What we observed (open suite, our settings):**
+- `gpt-4o-mini` vs `gemini-3.1-flash-lite`, 5 runs × 8 scenarios, OpenAI judge on → **8/8
+  `unchanged`**: the cross-vendor judge genuinely fired and found the two vendors behaviorally
+  equivalent on this suite.
+- `gpt-4o-mini` vs `llama-3.3-70b-versatile` on **Groq**, same suite → **8/8 `unchanged`**.
+
+**Free third vendor, zero cost:** [Groq](https://console.groq.com) serves Llama models over the
+OpenAI-compatible API and has a free tier, so a free key makes a zero-cost cross-vendor check:
+
+```bash
+export GROQ_API_KEY=...     # free at console.groq.com
+mp check --provider groq --from gpt-4o-mini --to llama-3.3-70b-versatile
+```
+
+A caveat worth stating: open-model *hosts* rotate ids but don't retire on a lab's fixed schedule the
+way the big providers do, so Groq/OpenRouter/etc. are a genuine cross-vendor bonus and an
+architecture proof — not the core migration wedge.
+
+---
+
+## Bring your own key
+
+Modelpin replays with **the end user's own API key**, always read from the environment, **never**
+hardcoded, shipped, or stored (cost stays yours; provider ToS stays clean):
+
+- `OPENAI_API_KEY`
+- `GEMINI_API_KEY` (or `GOOGLE_API_KEY`)
+- `GROQ_API_KEY` (and the equivalents for other OpenAI-compatible hosts)
+
+In CI, supply these as repo secrets (see the workflow above). Error text is scrubbed of
+`sk-` / `Bearer` tokens, so a failed call never leaks your key into a log, traceback, or PR comment.
+
+---
+
+## CLI reference
+
+| Command | What it does |
+|---|---|
+| `mp init [dir]` | Scaffold `modelpin.yaml` + `scenarios/` (never overwrites). |
+| `mp scan [path]` | Detect which AI models the repo depends on, and where. |
+| `mp baseline` | Record current model behavior for your scenarios (N runs). |
+| `mp check --to <model>` | Replay scenarios on a new model, diff vs baseline, write the PR-style report, fail CI on a regression. |
+| `mp version` | Print the Modelpin version. |
+| `mp report` | **Not built yet** — prints a "coming soon" notice. The public Modelpin Report suite is a stub. |
+
+Shared flags on `baseline` / `check`: `--from` / `--model`, `--provider`, `--runs`, `--match`
+(`strict\|unordered\|subset\|superset`), `--config`, `--scenarios-dir`, `--store-dir`, and
+`--fixtures` (with `--provider fake`).
+
+---
+
+## Install
+
+**Not on PyPI yet** (publishing is pending). For now, from source:
+
+```bash
+git clone https://github.com/samarthputhraya/modelpin
+cd modelpin
+python -m venv .venv && source .venv/bin/activate
+pip install -e ".[dev,providers]"     # Python 3.12+; 'providers' pulls the live SDKs
+mp version
+```
+
+Once published, the intended install is:
+
+```bash
+pip install "modelpin[providers]"     # or: pipx install "modelpin[providers]"
+```
+
+The `providers` extra pulls in the `openai`, `google-genai`, and `anthropic` SDKs. The base install
+runs the offline `fake` path with no provider SDKs at all.
+
+---
+
+## What this is *not* (non-goals, on purpose)
+
+Modelpin is a **migration tool**, and stays one. It is **not**:
+
+- a general eval / observability platform,
+- prompt management,
+- a model gateway or host,
+- an absolute "which model is best" leaderboard.
+
+It measures **behavior change relative to your app** — not abstract quality. Saying no to that scope
+is what keeps the false-positive promise honest and the tool small enough to trust.
+
+### Honest-framing rules (this is a trust product)
+
+- Any public / measurement claim is phrased as *"on our open suite, under these settings, we
+  observed…"* — **never** "Model X is worse." The harness and scenarios are open source so anyone
+  can rerun and disagree. That's the whole point of being the independent voice.
+- We don't overclaim and we don't falsely undersell. The engine is real and cross-vendor proven;
+  *and* Anthropic is a stub, `mp report` isn't built, the judge calibration is a documented first
+  pass, and it's pre-PyPI. All four are true at once.
+
+---
+
+## Status
+
+**Phase 0 (core engine MVP) — essentially complete.** Live-validated cross-vendor (OpenAI ↔ Google
+↔ Groq/Llama); held-out false-positive rate **0/8**; multi-turn replay; a real GitHub Action;
+PyPI-ready packaging (0.1.0, builds a clean wheel, `mp` + `modelpin` entry points); 132 tests
+passing, `ruff` + `black` clean. Stubs/TODOs (Anthropic adapter, `mp report`) are called out above,
+and it is **not yet published** to PyPI or the GitHub Marketplace.
+
+The full engineering record and roadmap live in [`docs/STATUS.md`](docs/STATUS.md). Next up:
+publish, the first public Modelpin Report (gated on a real model launch), and the Anthropic adapter.
+
+## License
+
+**Apache-2.0.** See [`LICENSE`](LICENSE). The open-source core (CLI, engine, Action) is and stays
+open; any future hosted tier lives in a separate, proprietary package.
+
+Repo: <https://github.com/samarthputhraya/modelpin>