shipwright-kit 0.7.0__tar.gz → 0.8.1__tar.gz

{shipwright_kit-0.7.0/shipwright_kit.egg-info → shipwright_kit-0.8.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shipwright-kit
-Version: 0.7.0
+Version: 0.8.1
 Summary: Shipwright — AI-agent dev framework + import-light design/eval/security library
 Author: Christian Huhn
 License-Expression: MIT
@@ -51,19 +51,18 @@ The library is consumed today by two real tools: **barb** and **sift** both impo
 
 ## Install
 
-The library is **not on PyPI** — the bare name `shipwright` belongs to an unrelated
-project, so the published distribution is **`shipwright-kit`** and the import name is
-**`shipwright_kit`**. For now, install from git:
+The bare name `shipwright` belongs to an unrelated project on PyPI, so the
+published distribution is **`shipwright-kit`** and the import name is
+**`shipwright_kit`**.
 
 ```bash
-uv pip install "git+https://github.com/duathron/shipwright@main"
+uv pip install "shipwright-kit>=0.7,<0.8"
 # then: import shipwright_kit
 ```
 
 > [!NOTE]
-> Pin a release tag instead of `@main` for reproducible builds once a tagged
-> release of the `shipwright-kit` distribution is cut. Do **not** `pip install
-> shipwright` from PyPI — that is a different, unrelated package.
+> Do **not** `pip install shipwright` from PyPI — that is a different, unrelated
+> package. The correct dist name is `shipwright-kit`.
 
 The security pack needs no extra — it ships with the base install and registers
 through the `shipwright_kit.packs` entry point.

{shipwright_kit-0.7.0 → shipwright_kit-0.8.1}/README.md

@@ -21,19 +21,18 @@ The library is consumed today by two real tools: **barb** and **sift** both impo
 
 ## Install
 
-The library is **not on PyPI** — the bare name `shipwright` belongs to an unrelated
-project, so the published distribution is **`shipwright-kit`** and the import name is
-**`shipwright_kit`**. For now, install from git:
+The bare name `shipwright` belongs to an unrelated project on PyPI, so the
+published distribution is **`shipwright-kit`** and the import name is
+**`shipwright_kit`**.
 
 ```bash
-uv pip install "git+https://github.com/duathron/shipwright@main"
+uv pip install "shipwright-kit>=0.7,<0.8"
 # then: import shipwright_kit
 ```
 
 > [!NOTE]
-> Pin a release tag instead of `@main` for reproducible builds once a tagged
-> release of the `shipwright-kit` distribution is cut. Do **not** `pip install
-> shipwright` from PyPI — that is a different, unrelated package.
+> Do **not** `pip install shipwright` from PyPI — that is a different, unrelated
+> package. The correct dist name is `shipwright-kit`.
 
 The security pack needs no extra — it ships with the base install and registers
 through the `shipwright_kit.packs` entry point.

{shipwright_kit-0.7.0 → shipwright_kit-0.8.1}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 # PyPI distribution name. The bare `shipwright` is taken on PyPI (unrelated 6si
 # tool), so the dist is `shipwright-kit`; the IMPORT name is `shipwright_kit`.
 name = "shipwright-kit"
-version = "0.7.0"
+version = "0.8.1"
 description = "Shipwright — AI-agent dev framework + import-light design/eval/security library"
 readme = "README.md"
 requires-python = ">=3.11"

{shipwright_kit-0.7.0 → shipwright_kit-0.8.1}/shipwright_kit/__init__.py

@@ -1,3 +1,3 @@
 """Shipwright — design-token + tooling library."""
 
-__version__ = "0.7.0"
+__version__ = "0.8.1"

shipwright_kit-0.8.1/shipwright_kit/eval/__init__.py

@@ -0,0 +1,18 @@
+"""Generic classification eval harness: corpus, metrics, evaluate + gate."""
+
+from .corpus import Sample, load_corpus
+from .harness import CorpusDisagreement, CorpusVerifyReport, EvalGateError, evaluate, gate, verify_corpus
+from .metrics import EVAL_SCHEMA_VERSION, EvalResult
+
+__all__ = [
+    "Sample",
+    "load_corpus",
+    "EvalResult",
+    "EVAL_SCHEMA_VERSION",
+    "EvalGateError",
+    "evaluate",
+    "gate",
+    "CorpusDisagreement",
+    "CorpusVerifyReport",
+    "verify_corpus",
+]

shipwright_kit-0.8.1/shipwright_kit/eval/harness.py

@@ -0,0 +1,129 @@
+"""Run a predict function over a corpus and gate the result. Count-and-skip on a
+predict-time exception (faithful to barb — a bad row must not abort the run)."""
+
+from __future__ import annotations
+
+from collections.abc import Callable
+from dataclasses import dataclass
+
+from .corpus import Sample
+from .metrics import EvalResult
+
+
+class EvalGateError(AssertionError):
+    """Raised when an eval result is below the required thresholds."""
+
+
+@dataclass(frozen=True)
+class CorpusDisagreement:
+    """A single row where the predictor's output disagrees with the human label."""
+
+    value: str  # the sample input
+    label: str  # the human-assigned label
+    predicted: str  # what the predictor returned
+
+
+@dataclass(frozen=True)
+class CorpusVerifyReport:
+    """Result of :func:`verify_corpus`. Stdlib-only, no rich/pyfiglet import."""
+
+    disagreements: list[CorpusDisagreement]
+    total: int
+    disagreement_count: int
+
+    @property
+    def clean(self) -> bool:
+        """True when every row agrees — safe to proceed to floor-setting."""
+        return self.disagreement_count == 0
+
+    def summary(self) -> str:
+        """Single-line human-readable summary, suitable for stderr."""
+        if self.clean:
+            return f"corpus-verify: OK — {self.total} rows, 0 disagreements"
+        return f"corpus-verify: FAIL — {self.disagreement_count}/{self.total} rows disagree (label vs prediction)"
+
+
+def verify_corpus(
+    corpus: list[Sample],
+    predictor: Callable[[str], str],
+    *,
+    eq: Callable[[str, str], bool] | None = None,
+) -> CorpusVerifyReport:
+    """Run *predictor* over every labeled row and report label-vs-prediction disagreements.
+
+    Use this **before** setting a precision/recall floor to catch mislabeled or
+    dishonest corpus rows — a predictor that is believed correct is compared
+    directly to the human label; rows that differ are flagged.
+
+    Parameters
+    ----------
+    corpus:
+        List of :class:`~shipwright_kit.eval.Sample` objects (input + label pairs).
+    predictor:
+        Callable that maps an input string to a prediction string.  Must be the
+        same callable you intend to gate — usually the production classifier.
+    eq:
+        Optional equality function ``(label, predicted) -> bool``.  Defaults to
+        plain string equality (``label == predicted``).  Supply a custom function
+        when the label space differs from the prediction space (e.g. case-folding,
+        synonyms, or a mapping dict).
+
+    Returns
+    -------
+    CorpusVerifyReport
+        Structured report with the full list of disagreements plus a summary count.
+        Predictor exceptions on a row are treated as a disagreement (predicted value
+        is set to ``"<error>"``).
+    """
+    _eq: Callable[[str, str], bool] = eq if eq is not None else (lambda a, b: a == b)
+    disagreements: list[CorpusDisagreement] = []
+    for sample in corpus:
+        try:
+            pred = predictor(sample.input)
+        except Exception:
+            disagreements.append(CorpusDisagreement(sample.input, sample.label, "<error>"))
+            continue
+        if not _eq(sample.label, pred):
+            disagreements.append(CorpusDisagreement(sample.input, sample.label, pred))
+    return CorpusVerifyReport(
+        disagreements=disagreements,
+        total=len(corpus),
+        disagreement_count=len(disagreements),
+    )
+
+
+def evaluate(
+    predict_fn: Callable[[str], str],
+    corpus: list[Sample],
+    *,
+    positive_pred: Callable[[str], bool],
+    positive_expected: Callable[[str], bool] | None = None,
+) -> EvalResult:
+    binarize_expected = positive_expected or positive_pred  # default = same-space (Phase B)
+    tp = fp = tn = fn = errors = 0
+    for sample in corpus:
+        try:
+            pred = predict_fn(sample.input)
+        except Exception:  # count-and-skip, surfaced via errors
+            errors += 1
+            continue
+        exp = binarize_expected(sample.label)
+        got = positive_pred(pred)
+        if exp and got:
+            tp += 1
+        elif got and not exp:
+            fp += 1
+        elif exp and not got:
+            fn += 1
+        else:
+            tn += 1
+    return EvalResult(tp, fp, tn, fn, errors)
+
+
+def gate(result: EvalResult, *, min_precision: float, min_recall: float) -> None:
+    if result.precision < min_precision:
+        raise EvalGateError(f"precision {result.precision:.3f} < {min_precision}")
+    if result.recall < min_recall:
+        raise EvalGateError(f"recall {result.recall:.3f} < {min_recall}")
+    if (result.tp + result.fn) > 0 and result.recall == 0.0:
+        raise EvalGateError("zero recall with positives present")

{shipwright_kit-0.7.0 → shipwright_kit-0.8.1/shipwright_kit.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shipwright-kit
-Version: 0.7.0
+Version: 0.8.1
 Summary: Shipwright — AI-agent dev framework + import-light design/eval/security library
 Author: Christian Huhn
 License-Expression: MIT
@@ -51,19 +51,18 @@ The library is consumed today by two real tools: **barb** and **sift** both impo
 
 ## Install
 
-The library is **not on PyPI** — the bare name `shipwright` belongs to an unrelated
-project, so the published distribution is **`shipwright-kit`** and the import name is
-**`shipwright_kit`**. For now, install from git:
+The bare name `shipwright` belongs to an unrelated project on PyPI, so the
+published distribution is **`shipwright-kit`** and the import name is
+**`shipwright_kit`**.
 
 ```bash
-uv pip install "git+https://github.com/duathron/shipwright@main"
+uv pip install "shipwright-kit>=0.7,<0.8"
 # then: import shipwright_kit
 ```
 
 > [!NOTE]
-> Pin a release tag instead of `@main` for reproducible builds once a tagged
-> release of the `shipwright-kit` distribution is cut. Do **not** `pip install
-> shipwright` from PyPI — that is a different, unrelated package.
+> Do **not** `pip install shipwright` from PyPI — that is a different, unrelated
+> package. The correct dist name is `shipwright-kit`.
 
 The security pack needs no extra — it ships with the base install and registers
 through the `shipwright_kit.packs` entry point.

{shipwright_kit-0.7.0 → shipwright_kit-0.8.1}/tests/test_template_wiring.py

@@ -56,7 +56,10 @@ def test_security_preset_installs_security_extra(tmp_path):
     text = (proj / "pyproject.toml").read_text()
     req = _shipwright_req(text)
     assert req.extras == set()  # security pack ships with base (entry-point); no [security] extra exists
-    assert "git+https://github.com/duathron/shipwright" in str(req.url)
+    # W2: PyPI range pin (>=0.7,<0.8) — url is None, specifier encodes the range
+    assert req.url is None
+    assert ">=0.7" in str(req.specifier)
+    assert "<0.8" in str(req.specifier)
     assert 'preset = "security"' in text
     banner = proj / "acme" / "banner.py"
     assert banner.exists()
@@ -68,5 +71,8 @@ def test_none_preset_core_only(tmp_path):
     text = (proj / "pyproject.toml").read_text()
     req = _shipwright_req(text)
     assert req.extras == set()  # no security extra
-    assert "git+https://github.com/duathron/shipwright" in str(req.url)
+    # W2: PyPI range pin (>=0.7,<0.8) — url is None, specifier encodes the range
+    assert req.url is None
+    assert ">=0.7" in str(req.specifier)
+    assert "<0.8" in str(req.specifier)
     assert 'preset = "none"' in text

shipwright_kit-0.7.0/shipwright_kit/eval/__init__.py

@@ -1,7 +0,0 @@
-"""Generic classification eval harness: corpus, metrics, evaluate + gate."""
-
-from .corpus import Sample, load_corpus
-from .harness import EvalGateError, evaluate, gate
-from .metrics import EVAL_SCHEMA_VERSION, EvalResult
-
-__all__ = ["Sample", "load_corpus", "EvalResult", "EVAL_SCHEMA_VERSION", "EvalGateError", "evaluate", "gate"]

shipwright_kit-0.7.0/shipwright_kit/eval/harness.py

@@ -1,50 +0,0 @@
-"""Run a predict function over a corpus and gate the result. Count-and-skip on a
-predict-time exception (faithful to barb — a bad row must not abort the run)."""
-
-from __future__ import annotations
-
-from collections.abc import Callable
-
-from .corpus import Sample
-from .metrics import EvalResult
-
-
-class EvalGateError(AssertionError):
-    """Raised when an eval result is below the required thresholds."""
-
-
-def evaluate(
-    predict_fn: Callable[[str], str],
-    corpus: list[Sample],
-    *,
-    positive_pred: Callable[[str], bool],
-    positive_expected: Callable[[str], bool] | None = None,
-) -> EvalResult:
-    binarize_expected = positive_expected or positive_pred  # default = same-space (Phase B)
-    tp = fp = tn = fn = errors = 0
-    for sample in corpus:
-        try:
-            pred = predict_fn(sample.input)
-        except Exception:  # count-and-skip, surfaced via errors
-            errors += 1
-            continue
-        exp = binarize_expected(sample.label)
-        got = positive_pred(pred)
-        if exp and got:
-            tp += 1
-        elif got and not exp:
-            fp += 1
-        elif exp and not got:
-            fn += 1
-        else:
-            tn += 1
-    return EvalResult(tp, fp, tn, fn, errors)
-
-
-def gate(result: EvalResult, *, min_precision: float, min_recall: float) -> None:
-    if result.precision < min_precision:
-        raise EvalGateError(f"precision {result.precision:.3f} < {min_precision}")
-    if result.recall < min_recall:
-        raise EvalGateError(f"recall {result.recall:.3f} < {min_recall}")
-    if (result.tp + result.fn) > 0 and result.recall == 0.0:
-        raise EvalGateError("zero recall with positives present")