leanscreen 0.1.0__tar.gz

leanscreen-0.1.0/.gitignore

@@ -0,0 +1,10 @@
+__pycache__/
+*.pyc
+.venv/
+.env
+dist/
+build/
+*.egg-info/
+.mypy_cache/
+.ruff_cache/
+.pytest_cache/

leanscreen-0.1.0/LICENSE

@@ -0,0 +1,105 @@
+# Functional Source License, Version 1.1, Apache 2.0 Future License
+
+## Abbreviation
+
+FSL-1.1-Apache-2.0
+
+## Notice
+
+Copyright 2026 Millennium Research
+
+## Terms and Conditions
+
+### Licensor ("We")
+
+The party offering the Software under these Terms and Conditions.
+
+### The Software
+
+The "Software" is each version of the software that we make available under
+these Terms and Conditions, as indicated by our inclusion of these Terms and
+Conditions with the Software.
+
+### License Grant
+
+Subject to your compliance with this License Grant and the Patents,
+Redistribution and Trademark clauses below, we hereby grant you the right to
+use, copy, modify, create derivative works, publicly perform, publicly display
+and redistribute the Software for any Permitted Purpose identified below.
+
+### Permitted Purpose
+
+A Permitted Purpose is any purpose other than a Competing Use. A Competing Use
+means making the Software available to others in a commercial product or
+service that:
+
+1. substitutes for the Software;
+
+2. substitutes for any other product or service we offer using the Software
+   that exists as of the date we make the Software available; or
+
+3. offers the same or substantially similar functionality as the Software.
+
+Permitted Purposes specifically include using the Software:
+
+1. for your internal use and access;
+
+2. for non-commercial education;
+
+3. for non-commercial research; and
+
+4. in connection with professional services that you provide to a licensee
+   using the Software in accordance with these Terms and Conditions.
+
+### Patents
+
+To the extent your use for a Permitted Purpose would necessarily infringe our
+patents, the license grant above includes a license under our patents. If you
+make a claim against any party that the Software infringes or contributes to
+the infringement of any patent, then your patent license to the Software ends
+immediately.
+
+### Redistribution
+
+The Terms and Conditions apply to all copies, modifications and derivatives of
+the Software.
+
+If you redistribute any copies, modifications or derivatives of the Software,
+you must include a copy of or a link to these Terms and Conditions and not
+remove any copyright notices provided in or with the Software.
+
+### Disclaimer
+
+THE SOFTWARE IS PROVIDED "AS IS" AND WITHOUT WARRANTIES OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING WITHOUT LIMITATION WARRANTIES OF FITNESS FOR A PARTICULAR
+PURPOSE, MERCHANTABILITY, TITLE OR NON-INFRINGEMENT.
+
+IN NO EVENT WILL WE HAVE ANY LIABILITY TO YOU ARISING OUT OF OR RELATED TO THE
+SOFTWARE, INCLUDING INDIRECT, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES,
+EVEN IF WE HAVE BEEN INFORMED OF THEIR POSSIBILITY IN ADVANCE.
+
+### Trademarks
+
+Except for displaying the License Details and identifying us as the origin of
+the Software, you have no right under these Terms and Conditions to use our
+trademarks, trade names, service marks or product names.
+
+## Grant of Future License
+
+We hereby irrevocably grant you an additional license to use the Software under
+the Apache License, Version 2.0 that is effective on the second anniversary of
+the date we make the Software available. On or after that date, you may use the
+Software under the Apache License, Version 2.0, in which case the following
+will apply:
+
+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.

leanscreen-0.1.0/PKG-INFO

@@ -0,0 +1,255 @@
+Metadata-Version: 2.4
+Name: leanscreen
+Version: 0.1.0
+Summary: A calibrated faithfulness screen for informal↔Lean 4 statement pairs, served over MCP. Screens only — never certifies.
+Project-URL: Homepage, https://millenniumresearch.ai/leanscreen
+Project-URL: Repository, https://github.com/ibrahimmian36/leanscreen
+Author-email: Millennium Research <ibrahimnmian@gmail.com>
+License: # Functional Source License, Version 1.1, Apache 2.0 Future License
+        
+        ## Abbreviation
+        
+        FSL-1.1-Apache-2.0
+        
+        ## Notice
+        
+        Copyright 2026 Millennium Research
+        
+        ## Terms and Conditions
+        
+        ### Licensor ("We")
+        
+        The party offering the Software under these Terms and Conditions.
+        
+        ### The Software
+        
+        The "Software" is each version of the software that we make available under
+        these Terms and Conditions, as indicated by our inclusion of these Terms and
+        Conditions with the Software.
+        
+        ### License Grant
+        
+        Subject to your compliance with this License Grant and the Patents,
+        Redistribution and Trademark clauses below, we hereby grant you the right to
+        use, copy, modify, create derivative works, publicly perform, publicly display
+        and redistribute the Software for any Permitted Purpose identified below.
+        
+        ### Permitted Purpose
+        
+        A Permitted Purpose is any purpose other than a Competing Use. A Competing Use
+        means making the Software available to others in a commercial product or
+        service that:
+        
+        1. substitutes for the Software;
+        
+        2. substitutes for any other product or service we offer using the Software
+           that exists as of the date we make the Software available; or
+        
+        3. offers the same or substantially similar functionality as the Software.
+        
+        Permitted Purposes specifically include using the Software:
+        
+        1. for your internal use and access;
+        
+        2. for non-commercial education;
+        
+        3. for non-commercial research; and
+        
+        4. in connection with professional services that you provide to a licensee
+           using the Software in accordance with these Terms and Conditions.
+        
+        ### Patents
+        
+        To the extent your use for a Permitted Purpose would necessarily infringe our
+        patents, the license grant above includes a license under our patents. If you
+        make a claim against any party that the Software infringes or contributes to
+        the infringement of any patent, then your patent license to the Software ends
+        immediately.
+        
+        ### Redistribution
+        
+        The Terms and Conditions apply to all copies, modifications and derivatives of
+        the Software.
+        
+        If you redistribute any copies, modifications or derivatives of the Software,
+        you must include a copy of or a link to these Terms and Conditions and not
+        remove any copyright notices provided in or with the Software.
+        
+        ### Disclaimer
+        
+        THE SOFTWARE IS PROVIDED "AS IS" AND WITHOUT WARRANTIES OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING WITHOUT LIMITATION WARRANTIES OF FITNESS FOR A PARTICULAR
+        PURPOSE, MERCHANTABILITY, TITLE OR NON-INFRINGEMENT.
+        
+        IN NO EVENT WILL WE HAVE ANY LIABILITY TO YOU ARISING OUT OF OR RELATED TO THE
+        SOFTWARE, INCLUDING INDIRECT, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES,
+        EVEN IF WE HAVE BEEN INFORMED OF THEIR POSSIBILITY IN ADVANCE.
+        
+        ### Trademarks
+        
+        Except for displaying the License Details and identifying us as the origin of
+        the Software, you have no right under these Terms and Conditions to use our
+        trademarks, trade names, service marks or product names.
+        
+        ## Grant of Future License
+        
+        We hereby irrevocably grant you an additional license to use the Software under
+        the Apache License, Version 2.0 that is effective on the second anniversary of
+        the date we make the Software available. On or after that date, you may use the
+        Software under the Apache License, Version 2.0, in which case the following
+        will apply:
+        
+        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.
+License-File: LICENSE
+Keywords: autoformalization,faithfulness,lean4,mathlib,mcp
+Requires-Python: <3.14,>=3.12
+Requires-Dist: httpx>=0.27.0
+Requires-Dist: mcp<2,>=1.28.0
+Requires-Dist: pydantic-settings>=2.3.0
+Requires-Dist: pydantic>=2.7.0
+Provides-Extra: dev
+Requires-Dist: mypy>=1.10.0; extra == 'dev'
+Requires-Dist: pytest>=8.2.0; extra == 'dev'
+Requires-Dist: ruff>=0.5.0; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# leanscreen
+
+A calibrated faithfulness screen for informal↔Lean 4 statement pairs, served
+over [MCP](https://modelcontextprotocol.io) so Claude (Code, Desktop, or any
+MCP client) can check statements **while you draft them**.
+
+**The one thing to understand before using it:** this screen may only
+*reject*. `passed_screening` means "no defect found by this harness" — it is
+**not** a certification of faithfulness. Measured against 886 frozen human
+verdicts, statements a human reviewer had rejected still passed the full
+screen 17.0% of the time for theorems and 35.6% for definitions; statements
+a human had certified faithful were flagged 15–18% of the time. Every
+response carries this calibration verbatim.
+
+## Two tools
+
+**`check_fast`** — deterministic only: lints (unused binders, trivially
+satisfiable existentials, pinned `∃!` witnesses, suspicious ℕ-arithmetic,
+…), vacuity checks (reflexive goals, `True` goals, withheld declarations),
+and Lean 4 elaboration against your own mathlib environment. **Zero API
+calls, no key needed, ~0.1s per statement once the REPL is warm.** Call it
+constantly while drafting.
+
+**`check_deep`** — everything in `check_fast`, plus two independent LLM
+judges under strict consensus (a back-translation judge and a
+clause-by-clause checklist judge on separate models) and an adversarial
+counterexample probe. Runs on **your** `ANTHROPIC_API_KEY`; measured cost is
+roughly $0.17–0.27 per statement (the response reports actual spend as
+`actual_cost_usd`), 30–60 seconds. Call it deliberately, before something
+ships.
+
+Both take `informal` (the natural-language statement), `lean` (the Lean 4
+statement), and an optional `kind` (`theorem` | `definition`, inferred from
+the declaration head when omitted). Responses rank their evidence —
+`counterexample` > `deterministic` > `two-judge-consensus` > `single-judge`
+— and a single-judge flag is explicitly labeled as below the reporting bar.
+
+## Install
+
+```bash
+pip install git+https://github.com/ibrahimmian36/leanscreen
+```
+
+Requires Python ≥3.12. Runtime dependencies: `httpx`, `pydantic`,
+`pydantic-settings`, `mcp` — nothing else.
+
+## Lean setup (optional but recommended)
+
+Without a Lean project, the server still runs — `check_fast` does lints +
+vacuity and says plainly that elaboration was skipped. With one, statements
+are elaborated for real:
+
+1. A Lean 4 project with mathlib, built: `lake build` inside it.
+2. The [community REPL](https://github.com/leanprover-community/repl),
+   built against the **same toolchain**: `lake build` inside the repl repo
+   gives you `.lake/build/bin/repl`.
+3. `lake` on the server's PATH.
+
+mathlib imports once at server startup (~100 seconds, in the background —
+calls arriving mid-warm-up answer immediately with a "still warming" note),
+then each check takes ~0.1s.
+
+## Configuration
+
+Environment variables (or a `.env` in the working directory), all
+`LEANSCREEN_`-prefixed:
+
+| Variable | Default | Meaning |
+|---|---|---|
+| `LEANSCREEN_LEAN_PROJECT_PATH` | unset | Lean 4 + mathlib project (elaboration off when unset) |
+| `LEANSCREEN_LEAN_REPL_PATH` | unset | community REPL binary; without it every check pays a full `lake env lean` |
+| `LEANSCREEN_LEAN_TIMEOUT_SECONDS` | `180` | per-statement Lean budget |
+| `LEANSCREEN_ANTHROPIC_MODEL` | `claude-opus-4-8` | judge A + probe (the calibrated default) |
+| `LEANSCREEN_JUDGE_B_MODEL` | `claude-fable-5` | checklist judge (calibrated default; locked-surface models get a 32k token budget automatically) |
+| `LEANSCREEN_MAX_TOKENS` | `4096` | judge A response budget |
+| `ANTHROPIC_API_KEY` | unset | needed for `check_deep` only |
+
+Claude Code (`.mcp.json` in your project) or Claude Desktop
+(`claude_desktop_config.json`):
+
+```json
+{
+  "mcpServers": {
+    "lean-faithfulness-screen": {
+      "command": "leanscreen",
+      "env": {
+        "LEANSCREEN_LEAN_PROJECT_PATH": "/path/to/your/lean-mathlib-project",
+        "LEANSCREEN_LEAN_REPL_PATH": "/path/to/repl/.lake/build/bin/repl",
+        "ANTHROPIC_API_KEY": "sk-ant-…"
+      }
+    }
+  }
+}
+```
+
+## What this does not guarantee
+
+The judge configuration was calibrated 2026-07-15 against 886 frozen human
+verdicts (595 faithful / 291 unfaithful) from a production research-math
+corpus. Under strict two-judge consensus, human-rejected pairs still passed
+17.0% (theorems) / 35.6% (definitions) of the time, and human-certified
+pairs were flagged 15–18% of the time. Both judges are Anthropic-family
+models, so correlated blind spots cannot be ruled out. The counterexample
+probe confabulates: on one PutnamBench sample its counterexamples were wrong
+4 times out of 5. **Treat every flag as a candidate for human confirmation
+and every pass as "nothing found", never "faithful."**
+
+Human certification — an expert reviewer confirming that the Lean means the
+informal statement — is what this screen deliberately does not automate. We
+offer it as a service: **contact ibrahimnmian@gmail.com**.
+
+## License
+
+[FSL-1.1-Apache-2.0](LICENSE) (the Functional Source License): free to
+use, copy, modify, and redistribute — including internal commercial use,
+non-commercial education and research, and professional services — but not
+to offer as a competing commercial product or service. **Each version
+automatically becomes Apache 2.0 two years after its release** (the same
+license as mathlib). Not OSI-approved until the conversion — read it
+before building on it commercially.
+
+## Provenance
+
+Extracted from Millennium Research's private formalization platform
+(2026-07-28); the detector stack, judge prompts, and calibration figures are
+the ones behind our benchmark audits — the miniF2F and ProofNet# filings are
+public, and the PutnamBench, ProofNetVerif, and CLEVER audits have been
+shared with their maintainers. The calibration *data* is not included.
+
+Project page: [millenniumresearch.ai/leanscreen](https://millenniumresearch.ai/leanscreen)

leanscreen-0.1.0/README.md

@@ -0,0 +1,130 @@
+# leanscreen
+
+A calibrated faithfulness screen for informal↔Lean 4 statement pairs, served
+over [MCP](https://modelcontextprotocol.io) so Claude (Code, Desktop, or any
+MCP client) can check statements **while you draft them**.
+
+**The one thing to understand before using it:** this screen may only
+*reject*. `passed_screening` means "no defect found by this harness" — it is
+**not** a certification of faithfulness. Measured against 886 frozen human
+verdicts, statements a human reviewer had rejected still passed the full
+screen 17.0% of the time for theorems and 35.6% for definitions; statements
+a human had certified faithful were flagged 15–18% of the time. Every
+response carries this calibration verbatim.
+
+## Two tools
+
+**`check_fast`** — deterministic only: lints (unused binders, trivially
+satisfiable existentials, pinned `∃!` witnesses, suspicious ℕ-arithmetic,
+…), vacuity checks (reflexive goals, `True` goals, withheld declarations),
+and Lean 4 elaboration against your own mathlib environment. **Zero API
+calls, no key needed, ~0.1s per statement once the REPL is warm.** Call it
+constantly while drafting.
+
+**`check_deep`** — everything in `check_fast`, plus two independent LLM
+judges under strict consensus (a back-translation judge and a
+clause-by-clause checklist judge on separate models) and an adversarial
+counterexample probe. Runs on **your** `ANTHROPIC_API_KEY`; measured cost is
+roughly $0.17–0.27 per statement (the response reports actual spend as
+`actual_cost_usd`), 30–60 seconds. Call it deliberately, before something
+ships.
+
+Both take `informal` (the natural-language statement), `lean` (the Lean 4
+statement), and an optional `kind` (`theorem` | `definition`, inferred from
+the declaration head when omitted). Responses rank their evidence —
+`counterexample` > `deterministic` > `two-judge-consensus` > `single-judge`
+— and a single-judge flag is explicitly labeled as below the reporting bar.
+
+## Install
+
+```bash
+pip install git+https://github.com/ibrahimmian36/leanscreen
+```
+
+Requires Python ≥3.12. Runtime dependencies: `httpx`, `pydantic`,
+`pydantic-settings`, `mcp` — nothing else.
+
+## Lean setup (optional but recommended)
+
+Without a Lean project, the server still runs — `check_fast` does lints +
+vacuity and says plainly that elaboration was skipped. With one, statements
+are elaborated for real:
+
+1. A Lean 4 project with mathlib, built: `lake build` inside it.
+2. The [community REPL](https://github.com/leanprover-community/repl),
+   built against the **same toolchain**: `lake build` inside the repl repo
+   gives you `.lake/build/bin/repl`.
+3. `lake` on the server's PATH.
+
+mathlib imports once at server startup (~100 seconds, in the background —
+calls arriving mid-warm-up answer immediately with a "still warming" note),
+then each check takes ~0.1s.
+
+## Configuration
+
+Environment variables (or a `.env` in the working directory), all
+`LEANSCREEN_`-prefixed:
+
+| Variable | Default | Meaning |
+|---|---|---|
+| `LEANSCREEN_LEAN_PROJECT_PATH` | unset | Lean 4 + mathlib project (elaboration off when unset) |
+| `LEANSCREEN_LEAN_REPL_PATH` | unset | community REPL binary; without it every check pays a full `lake env lean` |
+| `LEANSCREEN_LEAN_TIMEOUT_SECONDS` | `180` | per-statement Lean budget |
+| `LEANSCREEN_ANTHROPIC_MODEL` | `claude-opus-4-8` | judge A + probe (the calibrated default) |
+| `LEANSCREEN_JUDGE_B_MODEL` | `claude-fable-5` | checklist judge (calibrated default; locked-surface models get a 32k token budget automatically) |
+| `LEANSCREEN_MAX_TOKENS` | `4096` | judge A response budget |
+| `ANTHROPIC_API_KEY` | unset | needed for `check_deep` only |
+
+Claude Code (`.mcp.json` in your project) or Claude Desktop
+(`claude_desktop_config.json`):
+
+```json
+{
+  "mcpServers": {
+    "lean-faithfulness-screen": {
+      "command": "leanscreen",
+      "env": {
+        "LEANSCREEN_LEAN_PROJECT_PATH": "/path/to/your/lean-mathlib-project",
+        "LEANSCREEN_LEAN_REPL_PATH": "/path/to/repl/.lake/build/bin/repl",
+        "ANTHROPIC_API_KEY": "sk-ant-…"
+      }
+    }
+  }
+}
+```
+
+## What this does not guarantee
+
+The judge configuration was calibrated 2026-07-15 against 886 frozen human
+verdicts (595 faithful / 291 unfaithful) from a production research-math
+corpus. Under strict two-judge consensus, human-rejected pairs still passed
+17.0% (theorems) / 35.6% (definitions) of the time, and human-certified
+pairs were flagged 15–18% of the time. Both judges are Anthropic-family
+models, so correlated blind spots cannot be ruled out. The counterexample
+probe confabulates: on one PutnamBench sample its counterexamples were wrong
+4 times out of 5. **Treat every flag as a candidate for human confirmation
+and every pass as "nothing found", never "faithful."**
+
+Human certification — an expert reviewer confirming that the Lean means the
+informal statement — is what this screen deliberately does not automate. We
+offer it as a service: **contact ibrahimnmian@gmail.com**.
+
+## License
+
+[FSL-1.1-Apache-2.0](LICENSE) (the Functional Source License): free to
+use, copy, modify, and redistribute — including internal commercial use,
+non-commercial education and research, and professional services — but not
+to offer as a competing commercial product or service. **Each version
+automatically becomes Apache 2.0 two years after its release** (the same
+license as mathlib). Not OSI-approved until the conversion — read it
+before building on it commercially.
+
+## Provenance
+
+Extracted from Millennium Research's private formalization platform
+(2026-07-28); the detector stack, judge prompts, and calibration figures are
+the ones behind our benchmark audits — the miniF2F and ProofNet# filings are
+public, and the PutnamBench, ProofNetVerif, and CLEVER audits have been
+shared with their maintainers. The calibration *data* is not included.
+
+Project page: [millenniumresearch.ai/leanscreen](https://millenniumresearch.ai/leanscreen)

leanscreen-0.1.0/docs/PUBLISH_BRIEF.md

@@ -0,0 +1,112 @@
+# Build prompt — extract the screen into `leanscreen`, ready for public release
+
+> This is the brief the extraction was executed against (2026-07-28). It is
+> written so a fresh Claude Code session could run it end to end; it also
+> records every decision made, so future maintainers know what was deliberate.
+
+---
+
+I want the MCP faithfulness screener extracted from the private
+`arxiv-math-platform` repo into a standalone, publicly publishable package.
+The private repo stays the product; the public package is the funnel.
+
+## What the product is
+
+`leanscreen`: one pip-installable package, one console script (`leanscreen`),
+serving the two MCP tools that already exist — `check_fast` (free,
+deterministic: lints + vacuity + Lean elaboration) and `check_deep` (paid:
+dual-judge strict consensus + counterexample probe, on the user's own
+`ANTHROPIC_API_KEY`). Local-first, single-user, no hosting, no auth, no
+telemetry.
+
+## What moves, what stays private
+
+**Moves** (the screen subtree — it depends only on `httpx` at runtime):
+`llm.py`, `backtranslate.py` (judge prompts), `lints.py`, `quality.py`,
+`falsify.py`, `verify.py`, `lean_server.py`, `logging.py`, the single-pair
+`score_pair` core of `score_external.py`, and the MCP server. Unit tests for
+each module come along.
+
+**Stays private** (the commercial harness and everything upstream of it):
+the batch layer (`load_pairs` / `run` / `render_report` / budget rails /
+crash-safe resume), the numeric cross-evaluation probe and its sandbox, the
+whole data platform (ingestion, extraction, DB, review UI, export,
+benchmark, grader), the 886-verdict calibration set, and every census
+artifact. The public package quotes the calibration *numbers*; the data that
+produced them is the moat and never ships.
+
+## Decisions (made, revisit deliberately)
+
+- **Name**: `leanscreen`; package `leanscreen`, script `leanscreen`, MCP
+  server name stays `lean-faithfulness-screen`.
+- **License**: FSL-1.1-Apache-2.0 (revised from ELv2 on 2026-07-28 after
+  license research, on Ibby's request). Why FSL won: (1) its Competing Use
+  clause covers "any other product or service we offer using the Software" —
+  protecting the certification/screening *service*, where ELv2 only blocks
+  hosted versions of the software itself; (2) Permitted Purposes explicitly
+  allow internal use, non-commercial education, and research — the mathlib
+  audience; (3) each version converts to Apache 2.0 (mathlib's license)
+  after 2 years, the best community optics available, and a 2-year-old
+  screener without current calibration is not a competitive threat; (4) it
+  is simpler than BUSL, whose Additional Use Grant ambiguity is exactly what
+  Sentry created FSL to fix. (Not legal advice; have a lawyer confirm before
+  anything heavy rides on it.)
+- **Env prefix**: `LEANSCREEN_` (`LEAN_PROJECT_PATH`, `LEAN_REPL_PATH`,
+  `ANTHROPIC_MODEL`, `JUDGE_B_MODEL`, `MAX_TOKENS`, …). A fresh, slim
+  `Settings` — none of the platform's 50 knobs.
+- **Judge B is now a setting** (`LEANSCREEN_JUDGE_B_MODEL`, default
+  `claude-fable-5` = the calibrated configuration) instead of a hard-coded
+  constant; the locked-surface 32k max_tokens rule keys off the value.
+- **Numeric probe dropped** from the public `score_pair` (the MCP tools never
+  used it; smaller surface, smaller audit).
+- **`infer_kind` made public** (was `_infer_kind`).
+- **Fork policy**: this is a fork-with-provenance, not a shared library. The
+  private repo does NOT depend on the public package. When a detector
+  improves in one repo, port it to the other deliberately; the commit
+  message must name the source commit.
+
+## Non-negotiable constraints (carried over verbatim)
+
+- The screen may only reject — never certify. `CALIBRATION_DISCLOSURE` ships
+  verbatim in every payload; both tool descriptions state passing is not
+  certification (test-pinned).
+- Evidence tiers stay ranked, never flattened: counterexample >
+  deterministic > two-judge-consensus > single-judge (below the reporting
+  bar).
+- Every flag is a candidate, not a verdict.
+- No secrets, no internal paths, no customer or reviewer names anywhere in
+  the public tree — audit before every push.
+
+## Quality gate
+
+```bash
+.venv/bin/ruff check . && .venv/bin/ruff format --check . && .venv/bin/mypy && .venv/bin/pytest
+```
+
+Strict mypy, zero errors, same bar as the private repo.
+
+## Acceptance criteria
+
+1. `pip install -e .` in a fresh venv pulls only httpx, pydantic(-settings),
+   and mcp (plus their transitive deps) — none of the platform's heavy deps.
+2. The full test suite passes; the MCP-server tests are the same ones that
+   pass in the private repo, re-pointed.
+3. `leanscreen` starts over stdio with no Lean project and no API key and
+   answers `check_fast`, exactly as the private server does.
+4. README works for a stranger: REPL build walkthrough, generic paths,
+   config snippets for Claude Code / Claude Desktop, the calibration table,
+   the ELv2 summary, and a contact for certification engagements.
+5. A pre-publish audit pass over the whole tree: secret scan, internal-path
+   scan, license header presence, README claims match measured numbers.
+
+## Steps reserved for Ibby (do not automate)
+
+1. Confirm the license choice and the public contact address.
+2. `gh repo create` (public) + first push — publishing is a send.
+3. Any PyPI upload, registry listing, or announcement post (10am rule).
+
+---
+
+*Executed 2026-07-28. Origin: arxiv-math-platform @ 0f57167 (commits e39d2f9
++ 0f57167 are the MCP server and the proof-stub fix this package forked
+from).*