Specs & Plans
Fase 39 — nli-runtime Implementation Plan
For agentic workers: REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (
- [ ]) syntax for tracking.
Goal: Build jw_core.fidelity, a runtime NLI (Natural Language Inference) layer that verifies every Finding.summary is semantically entailed by its Finding.excerpt. Wires a triple-target provider stack (api / mlx / nvidia / cpu) following the Fase 33 pattern, plus a @fidelity_wrap decorator that annotates AgentResult.findings[*].metadata with nli_verdict / nli_score / nli_provider and optionally rejects/warns when the verdict falls below a configurable threshold.
Architecture: New subpackage packages/jw-core/src/jw_core/fidelity/ containing verdicts.py (NLIVerdict dataclass + Verdict Literal), nli.py (NLIProvider Protocol + evaluate_entailment helper), factory.py (registry + JW_NLI_PROVIDER env override + JW_PROVIDER_ORDER shared with Fase 33), and nli_providers/ (deberta_mnli, claude_nli, openai_nli, ollama_nli, fakes). The decorator @fidelity_wrap lives in jw-agents/src/jw_agents/fidelity_wrap.py because it needs to know AgentResult. CLI integration adds a --fidelity flag to existing agent commands and a new evaluate_nli MCP tool. Three optional-dependency extras: [nli-anthropic], [nli-openai], [nli-local].
Tech Stack: Python 3.13 · Pydantic-free (we keep jw-core pydantic-light: pure dataclasses) · transformers + torch (extra [nli-local]) · anthropic>=0.40 (extra [nli-anthropic]) · openai>=1.50 (extra [nli-openai]) · httpx (already present for Ollama) · pytest + respx (existing dev-deps) · hypothesis (existing, for property tests).
Spec: docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md.
File map
Creates:
packages/jw-core/src/jw_core/fidelity/__init__.pypackages/jw-core/src/jw_core/fidelity/verdicts.pypackages/jw-core/src/jw_core/fidelity/nli.pypackages/jw-core/src/jw_core/fidelity/factory.pypackages/jw-core/src/jw_core/fidelity/nli_providers/__init__.pypackages/jw-core/src/jw_core/fidelity/nli_providers/fakes.pypackages/jw-core/src/jw_core/fidelity/nli_providers/deberta_mnli.pypackages/jw-core/src/jw_core/fidelity/nli_providers/claude_nli.pypackages/jw-core/src/jw_core/fidelity/nli_providers/openai_nli.pypackages/jw-core/src/jw_core/fidelity/nli_providers/ollama_nli.pypackages/jw-core/tests/test_fidelity_verdicts.pypackages/jw-core/tests/test_fidelity_nli_protocol.pypackages/jw-core/tests/test_fidelity_fakes.pypackages/jw-core/tests/test_fidelity_factory.pypackages/jw-core/tests/test_fidelity_deberta.pypackages/jw-core/tests/test_fidelity_claude.pypackages/jw-core/tests/test_fidelity_openai.pypackages/jw-core/tests/test_fidelity_ollama.pypackages/jw-core/tests/test_fidelity_property.pypackages/jw-agents/src/jw_agents/fidelity_wrap.pypackages/jw-agents/tests/test_fidelity_wrap.pypackages/jw-agents/tests/test_fidelity_integration.pypackages/jw-cli/tests/test_cli_fidelity.pypackages/jw-mcp/tests/test_mcp_nli.pydocs/guias/fidelity-nli.md
Modifies:
packages/jw-core/pyproject.toml— add[nli-anthropic],[nli-openai],[nli-local]extras.packages/jw-cli/src/jw_cli/commands/apologetics.py— add--fidelityflag.packages/jw-cli/src/jw_cli/commands/verse.py— add--fidelityflag.packages/jw-cli/src/jw_cli/commands/research.py— add--fidelityflag.packages/jw-cli/src/jw_cli/commands/meeting.py— add--fidelityflag.packages/jw-mcp/src/jw_mcp/server.py— registerevaluate_nlitool + addfidelityparameter on the four wrapped agent tools.docs/VISION_AUDIT.md— add Fase 39 row.docs/ROADMAP.md— add Fase 39 section.docs/README.md— link the new guide.
Task 1: Scaffold jw_core.fidelity subpackage + Protocol + extras
Files:
-
Create:
packages/jw-core/src/jw_core/fidelity/__init__.py -
Create:
packages/jw-core/src/jw_core/fidelity/nli.py -
Create:
packages/jw-core/tests/test_fidelity_nli_protocol.py -
Modify:
packages/jw-core/pyproject.toml -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_nli_protocol.py
"""Tests that the public NLIProvider Protocol and Target Literal are exported
and structurally typed correctly.
Spec: docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md §"Provider Protocol".
"""
from __future__ import annotations
from typing import get_args
import pytest
from jw_core.fidelity import NLIProvider, Target
def test_target_literal_has_four_values() -> None:
assert set(get_args(Target)) == {"api", "mlx", "nvidia", "cpu"}
def test_nli_provider_is_runtime_checkable() -> None:
class Stub:
name = "stub"
target: Target = "cpu"
def is_available(self) -> bool:
return True
def evaluate(self, claim: str, premise: str, *, language: str = "en"):
raise NotImplementedError
assert isinstance(Stub(), NLIProvider)
def test_nli_provider_rejects_missing_method() -> None:
class Broken:
name = "broken"
target: Target = "cpu"
def is_available(self) -> bool:
return True
# no .evaluate()
assert not isinstance(Broken(), NLIProvider)
def test_public_api_exports_evaluate_entailment_helper() -> None:
from jw_core.fidelity import evaluate_entailment
assert callable(evaluate_entailment)- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-core/tests/test_fidelity_nli_protocol.py -v
Expected: FAIL — ModuleNotFoundError: No module named 'jw_core.fidelity'.
- Step 3: Implement the module
# packages/jw-core/src/jw_core/fidelity/nli.py
"""NLI Provider Protocol — runtime entailment judgement.
Every provider answers a single question: does `claim` semantically
follow from `premise`? The contract is intentionally narrow:
- sync function (no async)
- input: two strings + optional language code
- output: NLIVerdict (verdict label + 0..1 score + provider name + raw)
Rules (heritage of Fase 33):
1. No network at import time. Heavy deps (transformers, anthropic, openai)
are imported lazily inside `is_available()` and `evaluate()`.
2. `is_available()` is cheap — env var checks, package presence, hardware
detection. Called once per `get_default_nli_provider()`.
3. `evaluate()` is sync. API-backed providers should wrap their HTTP call
and block; callers (the @fidelity_wrap decorator) are async-aware and
can offload to threads.
4. `score` is always in [0, 1], normalized by the provider. DeBERTa
returns softmax[entailment]; LLMs return JSON `confidence`.
5. `language` is a hint for LLM providers; transformer NLI models that
are multilingual ignore it.
"""
from __future__ import annotations
from typing import Literal, Protocol, runtime_checkable
from jw_core.fidelity.verdicts import NLIVerdict
Target = Literal["api", "mlx", "nvidia", "cpu"]
@runtime_checkable
class NLIProvider(Protocol):
"""Canonical NLI provider contract.
Implementations declare a stable `name` (used by `JW_NLI_PROVIDER` env
override) and a `target` (used by `JW_PROVIDER_ORDER` ranking, shared
with Fase 33).
"""
name: str
target: Target
def is_available(self) -> bool: ...
def evaluate(
self, claim: str, premise: str, *, language: str = "en"
) -> NLIVerdict: ...
def evaluate_entailment(
claim: str,
premise: str,
*,
language: str = "en",
provider: NLIProvider | None = None,
) -> NLIVerdict:
"""Public helper: evaluate one claim/premise pair.
Resolves a default provider via `get_default_nli_provider()` if none
is supplied. Used by both `@fidelity_wrap` and Fase 44 (`synth-judge`).
"""
if provider is None:
from jw_core.fidelity.factory import get_default_nli_provider
provider = get_default_nli_provider()
return provider.evaluate(claim, premise, language=language)
__all__ = ["NLIProvider", "Target", "evaluate_entailment"]# packages/jw-core/src/jw_core/fidelity/__init__.py
"""jw_core.fidelity — runtime NLI verification of agent findings.
Public API:
from jw_core.fidelity import (
NLIProvider,
NLIVerdict,
Target,
evaluate_entailment,
get_default_nli_provider,
list_available_nli_providers,
)
Spec: docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md
"""
from __future__ import annotations
from jw_core.fidelity.nli import NLIProvider, Target, evaluate_entailment
from jw_core.fidelity.verdicts import NLIVerdict, Verdict
__all__ = [
"NLIProvider",
"NLIVerdict",
"Target",
"Verdict",
"evaluate_entailment",
"get_default_nli_provider",
"list_available_nli_providers",
]
def __getattr__(name: str): # noqa: D401
# Lazy re-exports of factory functions to avoid importing providers at
# import time (keeps `import jw_core` cheap on hosts without transformers).
if name == "get_default_nli_provider":
from jw_core.fidelity.factory import get_default_nli_provider as fn
return fn
if name == "list_available_nli_providers":
from jw_core.fidelity.factory import list_available_nli_providers as fn
return fn
raise AttributeError(name)Note: verdicts.py is added in Task 2 — for now create a minimal stub so the test in Step 1 imports cleanly:
# packages/jw-core/src/jw_core/fidelity/verdicts.py (stub — overwritten in Task 2)
from __future__ import annotations
from dataclasses import dataclass
from typing import Literal
Verdict = Literal["entails", "neutral", "contradicts"]
@dataclass(frozen=True)
class NLIVerdict:
verdict: Verdict
score: float
provider: str
raw: dict- Step 4: Add the extras to pyproject.toml
Edit packages/jw-core/pyproject.toml. Inside [project.optional-dependencies] (create the table if absent), add:
[project.optional-dependencies]
nli-anthropic = ["anthropic>=0.40,<1.0"]
nli-openai = ["openai>=1.50,<2.0"]
nli-local = [
"transformers>=4.45,<5.0",
"torch>=2.4",
]
nli-all = [
"anthropic>=0.40,<1.0",
"openai>=1.50,<2.0",
"transformers>=4.45,<5.0",
"torch>=2.4",
](If [project.optional-dependencies] already exists with other entries, append these keys without removing existing ones.)
- Step 5: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_nli_protocol.py -v
Expected: 4 passed.
- Step 6: Commit
git add packages/jw-core/src/jw_core/fidelity packages/jw-core/tests/test_fidelity_nli_protocol.py packages/jw-core/pyproject.toml
git commit -m "feat(jw-core): scaffold jw_core.fidelity Protocol and NLI extras"Task 2: NLIVerdict dataclass + Verdict Literal
Files:
-
Modify (overwrite):
packages/jw-core/src/jw_core/fidelity/verdicts.py -
Create:
packages/jw-core/tests/test_fidelity_verdicts.py -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_verdicts.py
"""Tests for jw_core.fidelity.verdicts.
The dataclass is frozen (hashable) and serializable via `asdict`. The
Verdict Literal is exhaustive: only three labels are legal, anything else
must trip a runtime guard.
"""
from __future__ import annotations
from dataclasses import asdict, is_dataclass
from typing import get_args
import pytest
from jw_core.fidelity.verdicts import NLIVerdict, Verdict, ensure_verdict
def test_verdict_literal_has_three_values() -> None:
assert set(get_args(Verdict)) == {"entails", "neutral", "contradicts"}
def test_nli_verdict_is_frozen_dataclass() -> None:
v = NLIVerdict(verdict="entails", score=0.92, provider="fake-nli", raw={})
assert is_dataclass(v)
with pytest.raises(Exception): # FrozenInstanceError subclass of AttributeError
v.score = 0.5 # type: ignore[misc]
def test_nli_verdict_asdict_roundtrips() -> None:
v = NLIVerdict(
verdict="contradicts",
score=0.71,
provider="claude-nli",
raw={"reason": "negation"},
)
d = asdict(v)
assert d == {
"verdict": "contradicts",
"score": 0.71,
"provider": "claude-nli",
"raw": {"reason": "negation"},
}
def test_nli_verdict_clamps_score_in_constructor_via_ensure() -> None:
# ensure_verdict is the canonical safe constructor used by providers
v = ensure_verdict(verdict="entails", score=1.7, provider="x")
assert v.score == 1.0
v2 = ensure_verdict(verdict="entails", score=-0.3, provider="x")
assert v2.score == 0.0
def test_ensure_verdict_rejects_bad_label() -> None:
with pytest.raises(ValueError, match="invalid verdict"):
ensure_verdict(verdict="maybe", score=0.5, provider="x") # type: ignore[arg-type]
def test_ensure_verdict_default_raw_is_empty_dict() -> None:
v = ensure_verdict(verdict="neutral", score=0.5, provider="x")
assert v.raw == {}- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-core/tests/test_fidelity_verdicts.py -v
Expected: FAIL — ImportError: cannot import name 'ensure_verdict'.
- Step 3: Implement the dataclass + helper
# packages/jw-core/src/jw_core/fidelity/verdicts.py
"""NLIVerdict — the canonical output of every NLIProvider.
We use a frozen dataclass (not Pydantic) because `jw-core` deliberately
avoids Pydantic dependencies at the leaf layer — Pydantic lives one level
up in `jw-eval` / `jw-agents`. Frozen dataclasses are hashable, fast, and
sufficient for our needs.
`ensure_verdict` is the safe constructor every provider should funnel
through — it clamps `score` to [0, 1] and validates the verdict label.
This is the single chokepoint that protects downstream consumers from
provider bugs (LLM hallucinated `score=1.7`, etc.).
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any, Literal, get_args
Verdict = Literal["entails", "neutral", "contradicts"]
_VALID_VERDICTS: frozenset[str] = frozenset(get_args(Verdict))
@dataclass(frozen=True)
class NLIVerdict:
"""One NLI judgement, suitable for `Finding.metadata["nli_*"]`.
Fields:
verdict — discrete label (entails / neutral / contradicts).
score — confidence in [0, 1]. For multi-class providers, this is
the probability of the chosen verdict; for LLM judges,
the JSON-returned confidence.
provider — provider.name for traceability ("claude-nli", "deberta-v3-mnli",
"fake-nli", ...). The decorator stamps this into metadata.
raw — provider-specific debug payload. Optional. May be persisted
to traces (Fase 43) but is NEVER displayed in CLI output.
"""
verdict: Verdict
score: float
provider: str
raw: dict[str, Any] = field(default_factory=dict)
def ensure_verdict(
*,
verdict: str,
score: float,
provider: str,
raw: dict[str, Any] | None = None,
) -> NLIVerdict:
"""Canonical constructor — clamp score, validate verdict label."""
if verdict not in _VALID_VERDICTS:
raise ValueError(
f"invalid verdict {verdict!r}; expected one of {sorted(_VALID_VERDICTS)}"
)
clamped = max(0.0, min(1.0, float(score)))
return NLIVerdict(
verdict=verdict, # type: ignore[arg-type]
score=clamped,
provider=provider,
raw=dict(raw) if raw else {},
)
__all__ = ["NLIVerdict", "Verdict", "ensure_verdict"]Update packages/jw-core/src/jw_core/fidelity/__init__.py to re-export ensure_verdict:
# packages/jw-core/src/jw_core/fidelity/__init__.py (full overwrite)
"""jw_core.fidelity — runtime NLI verification of agent findings.
Public API:
from jw_core.fidelity import (
NLIProvider,
NLIVerdict,
Target,
Verdict,
ensure_verdict,
evaluate_entailment,
get_default_nli_provider,
list_available_nli_providers,
)
Spec: docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md
"""
from __future__ import annotations
from jw_core.fidelity.nli import NLIProvider, Target, evaluate_entailment
from jw_core.fidelity.verdicts import NLIVerdict, Verdict, ensure_verdict
__all__ = [
"NLIProvider",
"NLIVerdict",
"Target",
"Verdict",
"ensure_verdict",
"evaluate_entailment",
"get_default_nli_provider",
"list_available_nli_providers",
]
def __getattr__(name: str):
if name == "get_default_nli_provider":
from jw_core.fidelity.factory import get_default_nli_provider as fn
return fn
if name == "list_available_nli_providers":
from jw_core.fidelity.factory import list_available_nli_providers as fn
return fn
raise AttributeError(name)- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_verdicts.py packages/jw-core/tests/test_fidelity_nli_protocol.py -v
Expected: 6 + 4 = 10 passed.
- Step 5: Commit
git add packages/jw-core/src/jw_core/fidelity packages/jw-core/tests/test_fidelity_verdicts.py
git commit -m "feat(jw-core): NLIVerdict frozen dataclass + ensure_verdict safe constructor"Task 3: FakeNLI deterministic provider
Files:
-
Create:
packages/jw-core/src/jw_core/fidelity/nli_providers/__init__.py -
Create:
packages/jw-core/src/jw_core/fidelity/nli_providers/fakes.py -
Create:
packages/jw-core/tests/test_fidelity_fakes.py -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_fakes.py
"""Tests for FakeNLI — the always-available deterministic provider.
Algorithm (per spec §"FakeNLI"):
- verdict = "entails" iff Jaccard(words(claim), words(premise)) >= 0.8
- verdict = "contradicts" iff a negation token appears in EXACTLY one of
{claim, premise}: "no es" / "is not" / "não é"
- else verdict = "neutral"
- score = round(jaccard, 2)
The provider must be 100% pure (no network, no model files) and stable
across processes — `evaluate("a", "b")` returns the same NLIVerdict
forever.
"""
from __future__ import annotations
import pytest
from jw_core.fidelity import NLIProvider
from jw_core.fidelity.nli_providers.fakes import FakeNLI
@pytest.fixture()
def provider() -> FakeNLI:
return FakeNLI()
def test_fake_implements_protocol(provider: FakeNLI) -> None:
assert isinstance(provider, NLIProvider)
assert provider.name == "fake-nli"
assert provider.target == "cpu"
assert provider.is_available() is True
def test_entails_when_claim_is_subset(provider: FakeNLI) -> None:
v = provider.evaluate(
claim="God loves the world",
premise="God so loved the world that he gave his only Son",
)
assert v.verdict == "entails"
assert v.score >= 0.5
assert v.provider == "fake-nli"
def test_contradicts_on_asymmetric_negation_en(provider: FakeNLI) -> None:
v = provider.evaluate(
claim="The Trinity is biblical",
premise="The Trinity is not biblical",
)
assert v.verdict == "contradicts"
def test_contradicts_on_asymmetric_negation_es(provider: FakeNLI) -> None:
v = provider.evaluate(
claim="el alma muere",
premise="el alma no es inmortal",
)
assert v.verdict == "contradicts"
def test_contradicts_on_asymmetric_negation_pt(provider: FakeNLI) -> None:
v = provider.evaluate(
claim="Jesus é Deus",
premise="Jesus não é Deus",
)
assert v.verdict == "contradicts"
def test_neutral_when_disjoint(provider: FakeNLI) -> None:
v = provider.evaluate(
claim="bananas are yellow",
premise="the sky was blue today",
)
assert v.verdict == "neutral"
assert v.score < 0.3
def test_deterministic_same_input_same_output(provider: FakeNLI) -> None:
a = provider.evaluate(claim="hello world", premise="hello world today")
b = provider.evaluate(claim="hello world", premise="hello world today")
assert a == b
def test_score_is_clamped_in_unit_interval(provider: FakeNLI) -> None:
v = provider.evaluate(claim="x", premise="x")
assert 0.0 <= v.score <= 1.0
def test_empty_inputs_do_not_crash(provider: FakeNLI) -> None:
v = provider.evaluate(claim="", premise="")
assert v.verdict in {"entails", "neutral", "contradicts"}
assert 0.0 <= v.score <= 1.0
def test_negation_in_both_does_not_count_as_contradiction(provider: FakeNLI) -> None:
v = provider.evaluate(
claim="el alma no es eterna",
premise="el alma no es inmortal",
)
# both contain a negation → cancels out → verdict driven by jaccard only
assert v.verdict in {"entails", "neutral"}- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-core/tests/test_fidelity_fakes.py -v
Expected: FAIL — ModuleNotFoundError: No module named 'jw_core.fidelity.nli_providers'.
- Step 3: Implement FakeNLI
# packages/jw-core/src/jw_core/fidelity/nli_providers/__init__.py
"""Concrete NLIProvider implementations.
Each provider lives in its own module so optional deps (transformers,
anthropic, openai) can be imported lazily and CI hosts without those
deps still install `jw-core` cleanly.
"""# packages/jw-core/src/jw_core/fidelity/nli_providers/fakes.py
"""Deterministic Fake NLI provider — no network, no model weights.
Algorithm:
1. Tokenize both inputs (Unicode word chars, lowercased).
2. Compute Jaccard similarity J = |A ∩ B| / |A ∪ B| (0 if both empty).
3. Detect explicit negation in each input (regex per language).
4. If negation appears in exactly one input → verdict = "contradicts".
If J >= 0.8 → verdict = "entails".
Else → verdict = "neutral".
5. score = round(J, 2), clamped to [0, 1] by ensure_verdict.
This is what every test in the test suite reaches for by default — the
factory falls back to it when no real provider is configured. It must
never raise on legal inputs and must be byte-identical across processes.
"""
from __future__ import annotations
import re
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict, ensure_verdict
# Regexes for explicit negation phrases. Conservative on purpose — false
# positives are worse than false negatives for a stub.
_NEGATION_PATTERNS: tuple[re.Pattern[str], ...] = (
re.compile(r"\bis\s+not\b", re.IGNORECASE),
re.compile(r"\bare\s+not\b", re.IGNORECASE),
re.compile(r"\bnever\b", re.IGNORECASE),
re.compile(r"\bno\s+es\b", re.IGNORECASE),
re.compile(r"\bno\s+son\b", re.IGNORECASE),
re.compile(r"\bnunca\b", re.IGNORECASE),
re.compile(r"\bnão\s+é\b", re.IGNORECASE),
re.compile(r"\bnão\s+são\b", re.IGNORECASE),
)
_TOKEN_RE = re.compile(r"\w+", re.UNICODE)
def _words(text: str) -> frozenset[str]:
return frozenset(_TOKEN_RE.findall(text.lower()))
def _jaccard(a: frozenset[str], b: frozenset[str]) -> float:
if not a and not b:
return 0.0
union = a | b
if not union:
return 0.0
return len(a & b) / len(union)
def _has_negation(text: str) -> bool:
return any(p.search(text) for p in _NEGATION_PATTERNS)
class FakeNLI:
"""Pure-Python deterministic NLI. Always available."""
name = "fake-nli"
target: Target = "cpu"
def is_available(self) -> bool:
return True
def evaluate(
self, claim: str, premise: str, *, language: str = "en"
) -> NLIVerdict:
wa, wb = _words(claim), _words(premise)
jacc = _jaccard(wa, wb)
neg_claim = _has_negation(claim)
neg_premise = _has_negation(premise)
asymmetric_negation = neg_claim ^ neg_premise
if asymmetric_negation:
verdict = "contradicts"
elif jacc >= 0.8:
verdict = "entails"
else:
verdict = "neutral"
return ensure_verdict(
verdict=verdict,
score=round(jacc, 2),
provider=self.name,
raw={
"jaccard": round(jacc, 4),
"neg_claim": neg_claim,
"neg_premise": neg_premise,
"lang": language,
},
)
__all__ = ["FakeNLI"]- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_fakes.py -v
Expected: 10 passed.
- Step 5: Commit
git add packages/jw-core/src/jw_core/fidelity/nli_providers packages/jw-core/tests/test_fidelity_fakes.py
git commit -m "feat(jw-core): FakeNLI deterministic provider (jaccard + negation heuristic)"Task 4: factory.py with JW_NLI_PROVIDER env override
Files:
-
Create:
packages/jw-core/src/jw_core/fidelity/factory.py -
Create:
packages/jw-core/tests/test_fidelity_factory.py -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_factory.py
"""Tests for the NLI factory.
Contracts:
1. `get_default_nli_provider()` always returns something (FakeNLI is
the last-resort fallback).
2. `JW_NLI_PROVIDER=fake-nli` selects FakeNLI explicitly.
3. `JW_NLI_PROVIDER=claude` selects ClaudeNLI when available, else raises
(we do NOT silently degrade — the user asked for a specific provider).
4. `JW_NLI_PROVIDER=bogus` raises ValueError.
5. `JW_PROVIDER_ORDER` reorders the registry (shared with Fase 33).
6. `list_available_nli_providers()` excludes fakes from the public listing
but `_named_lookup("fake-deberta-v3-mnli")` finds the fake variant.
"""
from __future__ import annotations
import pytest
from jw_core.fidelity.factory import (
ENV_NLI,
ENV_PROVIDER_ORDER,
get_default_nli_provider,
list_available_nli_providers,
)
def test_default_returns_a_provider(monkeypatch) -> None:
monkeypatch.delenv(ENV_NLI, raising=False)
p = get_default_nli_provider()
assert p is not None
assert hasattr(p, "evaluate")
assert hasattr(p, "name")
def test_env_override_selects_fake(monkeypatch) -> None:
monkeypatch.setenv(ENV_NLI, "fake-nli")
p = get_default_nli_provider()
assert p.name == "fake-nli"
def test_env_override_unknown_name_raises(monkeypatch) -> None:
monkeypatch.setenv(ENV_NLI, "bogus-provider")
with pytest.raises(ValueError, match="unknown JW_NLI_PROVIDER"):
get_default_nli_provider()
def test_env_override_claude_when_unavailable_raises(monkeypatch) -> None:
monkeypatch.setenv(ENV_NLI, "claude-nli")
monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False)
# ClaudeNLI without API key is_available() == False → factory must raise
# because the user explicitly named it.
with pytest.raises(RuntimeError, match="not available"):
get_default_nli_provider()
def test_fallback_to_fake_when_nothing_available(monkeypatch) -> None:
monkeypatch.delenv(ENV_NLI, raising=False)
monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False)
monkeypatch.delenv("OPENAI_API_KEY", raising=False)
p = get_default_nli_provider()
# On CI hosts without GPUs and without API keys, fake-nli is the floor.
assert p.name in {
"fake-nli",
"deberta-v3-mnli",
"ollama-nli",
"claude-nli",
"openai-nli",
}
def test_list_available_excludes_fake(monkeypatch) -> None:
monkeypatch.delenv(ENV_NLI, raising=False)
listed = list_available_nli_providers()
names = {p.name for p in listed}
assert "fake-nli" not in names
def test_provider_order_env_reorders(monkeypatch) -> None:
monkeypatch.delenv(ENV_NLI, raising=False)
monkeypatch.setenv(ENV_PROVIDER_ORDER, "cpu,api,mlx,nvidia")
# Just check the call doesn't crash and still returns something.
p = get_default_nli_provider()
assert p is not None
def test_named_lookup_can_select_fake_explicitly(monkeypatch) -> None:
monkeypatch.setenv(ENV_NLI, "fake-nli")
p = get_default_nli_provider()
assert p.name == "fake-nli"- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-core/tests/test_fidelity_factory.py -v
Expected: FAIL — ModuleNotFoundError: No module named 'jw_core.fidelity.factory'.
- Step 3: Implement factory.py
# packages/jw-core/src/jw_core/fidelity/factory.py
"""NLIProvider registry + factory.
Mirrors `jw_rag.rerank_providers.factory` so the operational model is
identical to Fase 33: an ordered list of provider instances + env override
+ shared `JW_PROVIDER_ORDER` env for target ranking.
Order of resolution:
1. If `JW_NLI_PROVIDER` is set:
- look up by exact `provider.name`.
- if `is_available()` → return.
- if not available → raise RuntimeError (do not silently fall through).
- if name unknown → raise ValueError.
2. Else: iterate `_instantiate_registry()`, return the first `is_available()`,
skipping FakeNLI (it's the last-resort floor).
3. If nothing is available → return FakeNLI.
Registry order (priority): Claude > OpenAI > DeBERTa(mlx) > DeBERTa(nvidia)
> DeBERTa(cpu) > Ollama > FakeNLI.
"""
from __future__ import annotations
import logging
import os
from typing import Literal
from jw_core.fidelity.nli import NLIProvider, Target
logger = logging.getLogger(__name__)
PROVIDER_ORDER_DEFAULT: list[Target] = ["api", "mlx", "nvidia", "cpu"]
ENV_NLI = "JW_NLI_PROVIDER"
ENV_PROVIDER_ORDER = "JW_PROVIDER_ORDER"
def _provider_order() -> list[Target]:
raw = os.getenv(ENV_PROVIDER_ORDER, "")
if not raw.strip():
return PROVIDER_ORDER_DEFAULT
parts: list[Target] = []
for piece in raw.split(","):
piece = piece.strip()
if piece in {"api", "mlx", "nvidia", "cpu"}:
parts.append(piece) # type: ignore[arg-type]
return parts or PROVIDER_ORDER_DEFAULT
def _instantiate_registry() -> list[NLIProvider]:
"""Build the canonical registry of all NLI providers.
Constructors are cheap (no model load, no network). The heavy work is
deferred to `is_available()` and the first `evaluate()` call.
"""
from jw_core.fidelity.nli_providers.claude_nli import ClaudeNLI
from jw_core.fidelity.nli_providers.deberta_mnli import DeBERTaV3MNLI
from jw_core.fidelity.nli_providers.fakes import FakeNLI
from jw_core.fidelity.nli_providers.ollama_nli import OllamaNLI
from jw_core.fidelity.nli_providers.openai_nli import OpenAINLI
return [
ClaudeNLI(),
OpenAINLI(),
DeBERTaV3MNLI(target="mlx"),
DeBERTaV3MNLI(target="nvidia"),
DeBERTaV3MNLI(target="cpu"),
OllamaNLI(),
FakeNLI(),
]
def _named_lookup(name: str) -> NLIProvider | None:
"""Find a provider in the registry by exact `.name` match."""
for r in _instantiate_registry():
if r.name == name:
return r
return None
def list_available_nli_providers() -> list[NLIProvider]:
"""Public listing: every available provider EXCEPT fakes.
Fakes are reachable via explicit `JW_NLI_PROVIDER=fake-nli` but never
surface in the default listing — otherwise the auto-fallback would silently
use them on hosts that also have real providers.
"""
order = _provider_order()
available = [
r
for r in _instantiate_registry()
if r.is_available() and r.name != "fake-nli"
]
return sorted(
available,
key=lambda r: order.index(r.target) if r.target in order else len(order),
)
def get_default_nli_provider() -> NLIProvider:
"""Resolve the provider to use in the current process."""
env_name = os.getenv(ENV_NLI, "").strip()
if env_name:
p = _named_lookup(env_name)
if p is None:
raise ValueError(f"unknown JW_NLI_PROVIDER={env_name!r}")
if not p.is_available():
raise RuntimeError(
f"JW_NLI_PROVIDER={env_name!r} not available "
f"(target={p.target}, missing deps or env vars)"
)
return p
for r in list_available_nli_providers():
return r
# Last-resort floor — always works.
from jw_core.fidelity.nli_providers.fakes import FakeNLI
logger.info("No real NLI provider available; falling back to FakeNLI")
return FakeNLI()
__all__ = [
"ENV_NLI",
"ENV_PROVIDER_ORDER",
"PROVIDER_ORDER_DEFAULT",
"get_default_nli_provider",
"list_available_nli_providers",
]Note: the factory imports ClaudeNLI, OpenAINLI, DeBERTaV3MNLI, OllamaNLI. Add minimal stubs now so imports succeed; Tasks 5–7 fill them in. Each stub must declare name, target, is_available() -> False, and a evaluate that raises NotImplementedError:
# packages/jw-core/src/jw_core/fidelity/nli_providers/claude_nli.py (stub for now)
from __future__ import annotations
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict
class ClaudeNLI:
name = "claude-nli"
target: Target = "api"
def is_available(self) -> bool:
return False
def evaluate(self, claim: str, premise: str, *, language: str = "en") -> NLIVerdict:
raise NotImplementedError("ClaudeNLI not yet wired (Task 5)")# packages/jw-core/src/jw_core/fidelity/nli_providers/openai_nli.py (stub)
from __future__ import annotations
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict
class OpenAINLI:
name = "openai-nli"
target: Target = "api"
def is_available(self) -> bool:
return False
def evaluate(self, claim: str, premise: str, *, language: str = "en") -> NLIVerdict:
raise NotImplementedError("OpenAINLI not yet wired (Task 6)")# packages/jw-core/src/jw_core/fidelity/nli_providers/deberta_mnli.py (stub)
from __future__ import annotations
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict
class DeBERTaV3MNLI:
name = "deberta-v3-mnli"
def __init__(self, *, target: Target = "cpu") -> None:
self.target = target
def is_available(self) -> bool:
return False
def evaluate(self, claim: str, premise: str, *, language: str = "en") -> NLIVerdict:
raise NotImplementedError("DeBERTaV3MNLI not yet wired (Task 7)")# packages/jw-core/src/jw_core/fidelity/nli_providers/ollama_nli.py (stub)
from __future__ import annotations
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict
class OllamaNLI:
name = "ollama-nli"
target: Target = "cpu"
def is_available(self) -> bool:
return False
def evaluate(self, claim: str, premise: str, *, language: str = "en") -> NLIVerdict:
raise NotImplementedError("OllamaNLI not yet wired (Task 7)")- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_factory.py -v
Expected: 8 passed.
- Step 5: Commit
git add packages/jw-core/src/jw_core/fidelity packages/jw-core/tests/test_fidelity_factory.py
git commit -m "feat(jw-core): NLI factory with JW_NLI_PROVIDER env + ordered registry"Task 5: ClaudeNLI provider (anthropic)
Files:
-
Modify (overwrite stub):
packages/jw-core/src/jw_core/fidelity/nli_providers/claude_nli.py -
Create:
packages/jw-core/tests/test_fidelity_claude.py -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_claude.py
"""Tests for ClaudeNLI provider.
We never hit the real API: the test injects a FakeAnthropicClient that
returns canned JSON. This keeps CI offline + deterministic.
"""
from __future__ import annotations
import json
from typing import Any
import pytest
from jw_core.fidelity.nli_providers.claude_nli import ClaudeNLI
class _FakeMessage:
def __init__(self, text: str) -> None:
self.content = [type("Block", (), {"text": text})()]
class _FakeMessages:
def __init__(self, response_text: str) -> None:
self.response_text = response_text
self.calls: list[dict[str, Any]] = []
def create(self, **kwargs: Any) -> _FakeMessage:
self.calls.append(kwargs)
return _FakeMessage(self.response_text)
class _FakeAnthropicClient:
def __init__(self, response_text: str) -> None:
self.messages = _FakeMessages(response_text)
def test_claude_unavailable_without_api_key(monkeypatch) -> None:
monkeypatch.delenv("ANTHROPIC_API_KEY", raising=False)
p = ClaudeNLI()
assert p.is_available() is False
def test_claude_available_with_api_key(monkeypatch) -> None:
# Skip if anthropic SDK isn't installed in the dev env
pytest.importorskip("anthropic")
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
p = ClaudeNLI()
assert p.is_available() is True
def test_claude_parses_entails(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient(
json.dumps({"verdict": "entails", "score": 0.91, "reason": "supported"})
)
p = ClaudeNLI(client=client)
v = p.evaluate(claim="A", premise="B", language="es")
assert v.verdict == "entails"
assert v.score == 0.91
assert v.provider == "claude-nli"
assert v.raw["reason"] == "supported"
def test_claude_parses_contradicts(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient(
json.dumps({"verdict": "contradicts", "score": 0.83, "reason": "negation"})
)
p = ClaudeNLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "contradicts"
assert v.score == 0.83
def test_claude_parses_neutral_default(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient(json.dumps({"verdict": "neutral", "score": 0.5}))
p = ClaudeNLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "neutral"
def test_claude_fallback_on_invalid_json(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient("not even json at all")
p = ClaudeNLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "neutral"
assert v.score == 0.5
assert "parse_error" in v.raw
def test_claude_fallback_on_invalid_verdict(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient(json.dumps({"verdict": "maybe", "score": 0.9}))
p = ClaudeNLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "neutral"
def test_claude_truncates_long_premise(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient(json.dumps({"verdict": "entails", "score": 0.8}))
p = ClaudeNLI(client=client)
very_long_premise = "x" * 20000
p.evaluate(claim="short", premise=very_long_premise)
sent = client.messages.calls[0]
# The user message body must contain a TRUNCATED premise (<= 6000 chars)
user_msg = sent["messages"][0]["content"]
assert "x" * 6000 in user_msg
assert "x" * 7000 not in user_msg
def test_claude_uses_env_model(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
monkeypatch.setenv("JW_NLI_CLAUDE_MODEL", "claude-haiku-4-5-20251001")
client = _FakeAnthropicClient(json.dumps({"verdict": "entails", "score": 0.9}))
p = ClaudeNLI(client=client)
p.evaluate(claim="A", premise="B")
assert client.messages.calls[0]["model"] == "claude-haiku-4-5-20251001"
def test_claude_sets_prompt_caching(monkeypatch) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "sk-fake")
client = _FakeAnthropicClient(json.dumps({"verdict": "entails", "score": 0.9}))
p = ClaudeNLI(client=client)
p.evaluate(claim="A", premise="B")
sent = client.messages.calls[0]
# system prompt sent as a list-of-blocks with cache_control on the last block
system = sent["system"]
assert isinstance(system, list)
assert any(
block.get("cache_control", {}).get("type") == "ephemeral"
for block in system
if isinstance(block, dict)
)- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-core/tests/test_fidelity_claude.py -v
Expected: FAIL — current ClaudeNLI raises NotImplementedError and does not accept client=.
- Step 3: Implement ClaudeNLI
# packages/jw-core/src/jw_core/fidelity/nli_providers/claude_nli.py
"""ClaudeNLI — entailment via Anthropic's Claude.
Design (per spec §"ClaudeNLI"):
- System prompt (cached): "You are an NLI judge. Decide if the CONCLUSION
strictly entails from the PREMISE. Reply JSON-only: {verdict, score, reason}."
- User prompt: "PREMISE:\n{premise}\n\nCONCLUSION:\n{claim}\n\nLanguage: {language}"
- Parse JSON; on failure → verdict="neutral", score=0.5, raw["parse_error"]=raw.
- Cost guard: truncate premise to 6000 chars when (premise + claim) > 8000.
- Prompt caching: `cache_control: {type: "ephemeral"}` on the system block.
- Model default: `claude-sonnet-4-5-20250929`, override via `JW_NLI_CLAUDE_MODEL`.
The optional `client=` kwarg in the constructor exists for testing —
production code passes nothing and we lazily instantiate `Anthropic()`.
"""
from __future__ import annotations
import json
import logging
import os
from typing import Any
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict, ensure_verdict
logger = logging.getLogger(__name__)
_DEFAULT_MODEL = "claude-sonnet-4-5-20250929"
_SYSTEM_PROMPT = (
"You are an NLI judge. Decide if the CONCLUSION strictly entails from "
"the PREMISE. Reply JSON-only with this exact shape: "
'{"verdict": "entails"|"neutral"|"contradicts", '
'"score": 0.0-1.0, "reason": "short explanation"}. '
"Output nothing else."
)
_MAX_PREMISE_CHARS = 6000
_MAX_TOTAL_CHARS = 8000
class ClaudeNLI:
name = "claude-nli"
target: Target = "api"
def __init__(self, *, client: Any | None = None) -> None:
self._client = client # injectable for tests
def is_available(self) -> bool:
if not os.getenv("ANTHROPIC_API_KEY"):
return False
try:
import anthropic # noqa: F401
except ImportError:
return False
return True
def _ensure_client(self) -> Any:
if self._client is not None:
return self._client
from anthropic import Anthropic
self._client = Anthropic()
return self._client
def _truncate(self, premise: str, claim: str) -> str:
if len(premise) + len(claim) <= _MAX_TOTAL_CHARS:
return premise
return premise[:_MAX_PREMISE_CHARS]
def evaluate(
self, claim: str, premise: str, *, language: str = "en"
) -> NLIVerdict:
client = self._ensure_client()
model = os.getenv("JW_NLI_CLAUDE_MODEL", _DEFAULT_MODEL)
truncated_premise = self._truncate(premise, claim)
user_body = (
f"PREMISE:\n{truncated_premise}\n\n"
f"CONCLUSION:\n{claim}\n\n"
f"Language: {language}"
)
system_blocks = [
{
"type": "text",
"text": _SYSTEM_PROMPT,
"cache_control": {"type": "ephemeral"},
}
]
try:
msg = client.messages.create(
model=model,
max_tokens=256,
system=system_blocks,
messages=[{"role": "user", "content": user_body}],
)
text = msg.content[0].text # type: ignore[union-attr,attr-defined]
except Exception as exc: # noqa: BLE001
logger.warning("ClaudeNLI call failed: %r", exc)
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"api_error": repr(exc)},
)
try:
data = json.loads(text)
verdict = str(data.get("verdict", "")).lower()
score = float(data.get("score", 0.5))
reason = str(data.get("reason", ""))
except Exception as exc: # noqa: BLE001
logger.warning(
"ClaudeNLI JSON parse failed: %r (raw=%s)", exc, text[:200]
)
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"parse_error": str(exc), "raw_text": text[:500]},
)
if verdict not in {"entails", "neutral", "contradicts"}:
logger.warning("ClaudeNLI unexpected verdict %r → neutral/0.5", verdict)
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"unexpected_verdict": verdict, "reason": reason},
)
return ensure_verdict(
verdict=verdict,
score=score,
provider=self.name,
raw={"reason": reason, "model": model, "lang": language},
)
__all__ = ["ClaudeNLI"]- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_claude.py -v
Expected: 10 passed (1 of which may show skipped if anthropic isn’t installed — that’s fine).
- Step 5: Commit
git add packages/jw-core/src/jw_core/fidelity/nli_providers/claude_nli.py packages/jw-core/tests/test_fidelity_claude.py
git commit -m "feat(jw-core): ClaudeNLI provider with prompt caching + JSON fallback"Task 6: OpenAINLI provider
Files:
-
Modify (overwrite stub):
packages/jw-core/src/jw_core/fidelity/nli_providers/openai_nli.py -
Create:
packages/jw-core/tests/test_fidelity_openai.py -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_openai.py
"""Tests for OpenAINLI provider.
Uses a FakeOpenAIClient that emulates `client.chat.completions.create` with
`response_format={"type": "json_schema", ...}` and returns canned JSON.
"""
from __future__ import annotations
import json
from typing import Any
import pytest
from jw_core.fidelity.nli_providers.openai_nli import OpenAINLI
class _FakeMessage:
def __init__(self, content: str) -> None:
self.content = content
class _FakeChoice:
def __init__(self, content: str) -> None:
self.message = _FakeMessage(content)
class _FakeResponse:
def __init__(self, content: str) -> None:
self.choices = [_FakeChoice(content)]
class _FakeCompletions:
def __init__(self, content: str) -> None:
self.content = content
self.calls: list[dict[str, Any]] = []
def create(self, **kwargs: Any) -> _FakeResponse:
self.calls.append(kwargs)
return _FakeResponse(self.content)
class _FakeChat:
def __init__(self, content: str) -> None:
self.completions = _FakeCompletions(content)
class _FakeOpenAIClient:
def __init__(self, content: str) -> None:
self.chat = _FakeChat(content)
def test_openai_unavailable_without_api_key(monkeypatch) -> None:
monkeypatch.delenv("OPENAI_API_KEY", raising=False)
p = OpenAINLI()
assert p.is_available() is False
def test_openai_parses_entails(monkeypatch) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-fake")
client = _FakeOpenAIClient(
json.dumps({"verdict": "entails", "score": 0.88, "reason": "ok"})
)
p = OpenAINLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "entails"
assert v.score == 0.88
assert v.provider == "openai-nli"
def test_openai_uses_structured_output(monkeypatch) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-fake")
client = _FakeOpenAIClient(json.dumps({"verdict": "neutral", "score": 0.5}))
p = OpenAINLI(client=client)
p.evaluate(claim="A", premise="B")
sent = client.chat.completions.calls[0]
rf = sent["response_format"]
assert rf["type"] == "json_schema"
assert "json_schema" in rf
schema = rf["json_schema"]["schema"]
assert "verdict" in schema["properties"]
assert "score" in schema["properties"]
def test_openai_uses_env_model(monkeypatch) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-fake")
monkeypatch.setenv("JW_NLI_OPENAI_MODEL", "gpt-4o")
client = _FakeOpenAIClient(json.dumps({"verdict": "entails", "score": 0.9}))
p = OpenAINLI(client=client)
p.evaluate(claim="A", premise="B")
assert client.chat.completions.calls[0]["model"] == "gpt-4o"
def test_openai_fallback_on_garbage(monkeypatch) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-fake")
client = _FakeOpenAIClient("not json")
p = OpenAINLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "neutral"
assert v.score == 0.5
def test_openai_truncates_long_premise(monkeypatch) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-fake")
client = _FakeOpenAIClient(json.dumps({"verdict": "entails", "score": 0.8}))
p = OpenAINLI(client=client)
p.evaluate(claim="short", premise="y" * 20000)
sent = client.chat.completions.calls[0]
user_msg = sent["messages"][-1]["content"]
assert "y" * 6000 in user_msg
assert "y" * 7000 not in user_msg
def test_openai_fallback_on_invalid_verdict(monkeypatch) -> None:
monkeypatch.setenv("OPENAI_API_KEY", "sk-fake")
client = _FakeOpenAIClient(json.dumps({"verdict": "??", "score": 1.0}))
p = OpenAINLI(client=client)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "neutral"- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-core/tests/test_fidelity_openai.py -v
Expected: FAIL — stub raises NotImplementedError.
- Step 3: Implement OpenAINLI
# packages/jw-core/src/jw_core/fidelity/nli_providers/openai_nli.py
"""OpenAINLI — entailment via OpenAI Chat Completions with structured output.
Uses `response_format={"type": "json_schema", "json_schema": {...}}` so the
SDK guarantees we receive a JSON-shaped string matching our schema. Default
model `gpt-4o-mini`, overridable via `JW_NLI_OPENAI_MODEL`.
Same defensive parsing as ClaudeNLI: bad JSON / bad verdict label → fallback
to verdict="neutral", score=0.5, raw["parse_error"].
"""
from __future__ import annotations
import json
import logging
import os
from typing import Any
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict, ensure_verdict
logger = logging.getLogger(__name__)
_DEFAULT_MODEL = "gpt-4o-mini"
_SYSTEM_PROMPT = (
"You are an NLI judge. Decide if the CONCLUSION strictly entails from "
"the PREMISE. Reply JSON-only with this exact shape: "
'{"verdict": "entails"|"neutral"|"contradicts", '
'"score": 0.0-1.0, "reason": "short explanation"}.'
)
_JSON_SCHEMA = {
"name": "nli_verdict",
"schema": {
"type": "object",
"properties": {
"verdict": {
"type": "string",
"enum": ["entails", "neutral", "contradicts"],
},
"score": {"type": "number", "minimum": 0.0, "maximum": 1.0},
"reason": {"type": "string"},
},
"required": ["verdict", "score"],
"additionalProperties": False,
},
}
_MAX_PREMISE_CHARS = 6000
_MAX_TOTAL_CHARS = 8000
class OpenAINLI:
name = "openai-nli"
target: Target = "api"
def __init__(self, *, client: Any | None = None) -> None:
self._client = client
def is_available(self) -> bool:
if not os.getenv("OPENAI_API_KEY"):
return False
try:
import openai # noqa: F401
except ImportError:
return False
return True
def _ensure_client(self) -> Any:
if self._client is not None:
return self._client
from openai import OpenAI
self._client = OpenAI()
return self._client
def _truncate(self, premise: str, claim: str) -> str:
if len(premise) + len(claim) <= _MAX_TOTAL_CHARS:
return premise
return premise[:_MAX_PREMISE_CHARS]
def evaluate(
self, claim: str, premise: str, *, language: str = "en"
) -> NLIVerdict:
client = self._ensure_client()
model = os.getenv("JW_NLI_OPENAI_MODEL", _DEFAULT_MODEL)
truncated = self._truncate(premise, claim)
user_body = (
f"PREMISE:\n{truncated}\n\n"
f"CONCLUSION:\n{claim}\n\n"
f"Language: {language}"
)
try:
resp = client.chat.completions.create(
model=model,
response_format={"type": "json_schema", "json_schema": _JSON_SCHEMA},
messages=[
{"role": "system", "content": _SYSTEM_PROMPT},
{"role": "user", "content": user_body},
],
)
text = resp.choices[0].message.content or ""
except Exception as exc: # noqa: BLE001
logger.warning("OpenAINLI call failed: %r", exc)
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"api_error": repr(exc)},
)
try:
data = json.loads(text)
verdict = str(data.get("verdict", "")).lower()
score = float(data.get("score", 0.5))
reason = str(data.get("reason", ""))
except Exception as exc: # noqa: BLE001
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"parse_error": str(exc), "raw_text": text[:500]},
)
if verdict not in {"entails", "neutral", "contradicts"}:
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"unexpected_verdict": verdict, "reason": reason},
)
return ensure_verdict(
verdict=verdict,
score=score,
provider=self.name,
raw={"reason": reason, "model": model, "lang": language},
)
__all__ = ["OpenAINLI"]- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_openai.py -v
Expected: 7 passed.
- Step 5: Commit
git add packages/jw-core/src/jw_core/fidelity/nli_providers/openai_nli.py packages/jw-core/tests/test_fidelity_openai.py
git commit -m "feat(jw-core): OpenAINLI provider with json_schema response format"Task 7: DeBERTaV3MNLI (local, lazy torch) + OllamaNLI
Files:
-
Modify (overwrite stub):
packages/jw-core/src/jw_core/fidelity/nli_providers/deberta_mnli.py -
Modify (overwrite stub):
packages/jw-core/src/jw_core/fidelity/nli_providers/ollama_nli.py -
Create:
packages/jw-core/tests/test_fidelity_deberta.py -
Create:
packages/jw-core/tests/test_fidelity_ollama.py -
Step 1: Write the failing tests
# packages/jw-core/tests/test_fidelity_deberta.py
"""Tests for DeBERTaV3MNLI.
We do NOT download model weights in CI. Tests inject a FakePipeline that
exposes the same `.tokenizer`/`.model` shape via duck typing. The integration
test that hits the real HuggingFace model is gated by the `nli-local`
extra and only runs in the nightly job.
"""
from __future__ import annotations
import pytest
from jw_core.fidelity.nli_providers.deberta_mnli import DeBERTaV3MNLI
class _FakePipelineOutput:
"""Mimics transformers.AutoModelForSequenceClassification output."""
def __init__(self, logits) -> None:
import torch
self.logits = torch.tensor(logits)
class _FakeTokenizer:
def __call__(self, premise, hypothesis, return_tensors, truncation, max_length):
# Echo so we can inspect truncation behavior
import torch
return {"input_ids": torch.tensor([[1, 2, 3]])}
class _FakeModel:
def __init__(self, logits) -> None:
self.logits = logits
def __call__(self, **kwargs):
return _FakePipelineOutput(self.logits)
def eval(self):
return self
def to(self, device): # noqa: ARG002
return self
def test_deberta_unavailable_without_transformers(monkeypatch) -> None:
# Pretend transformers is missing
import sys
monkeypatch.setitem(sys.modules, "transformers", None)
p = DeBERTaV3MNLI(target="cpu")
assert p.is_available() is False
def test_deberta_cpu_available_when_transformers_installed() -> None:
pytest.importorskip("transformers")
pytest.importorskip("torch")
p = DeBERTaV3MNLI(target="cpu")
assert p.is_available() is True
def test_deberta_nvidia_requires_cuda(monkeypatch) -> None:
pytest.importorskip("torch")
import torch
monkeypatch.setattr(torch.cuda, "is_available", lambda: False)
p = DeBERTaV3MNLI(target="nvidia")
assert p.is_available() is False
def test_deberta_evaluate_entails_via_injected_model() -> None:
pytest.importorskip("torch")
p = DeBERTaV3MNLI(target="cpu")
# logits[contradiction, neutral, entailment] = [0.1, 0.2, 5.0] → softmax ≈ entailment
p._tokenizer = _FakeTokenizer()
p._model = _FakeModel([[0.1, 0.2, 5.0]])
v = p.evaluate(claim="claim", premise="premise")
assert v.verdict == "entails"
assert v.score > 0.9
assert v.provider == "deberta-v3-mnli"
def test_deberta_evaluate_neutral_via_injected_model() -> None:
pytest.importorskip("torch")
p = DeBERTaV3MNLI(target="cpu")
p._tokenizer = _FakeTokenizer()
p._model = _FakeModel([[0.1, 5.0, 0.2]])
v = p.evaluate(claim="claim", premise="premise")
assert v.verdict == "neutral"
def test_deberta_evaluate_contradicts_via_injected_model() -> None:
pytest.importorskip("torch")
p = DeBERTaV3MNLI(target="cpu")
p._tokenizer = _FakeTokenizer()
p._model = _FakeModel([[5.0, 0.1, 0.2]])
v = p.evaluate(claim="claim", premise="premise")
assert v.verdict == "contradicts"
def test_deberta_lazy_load_caches_singleton(monkeypatch) -> None:
pytest.importorskip("torch")
p = DeBERTaV3MNLI(target="cpu")
p._tokenizer = _FakeTokenizer()
p._model = _FakeModel([[0.1, 0.2, 5.0]])
# Second call should NOT reload model — check the same instance is reused.
p.evaluate(claim="a", premise="b")
same_tokenizer = p._tokenizer
same_model = p._model
p.evaluate(claim="c", premise="d")
assert p._tokenizer is same_tokenizer
assert p._model is same_model# packages/jw-core/tests/test_fidelity_ollama.py
"""Tests for OllamaNLI — local LLM judge via Ollama HTTP API.
Uses `respx` to mock the HTTP endpoints. CI never actually contacts Ollama.
"""
from __future__ import annotations
import json
import httpx
import pytest
import respx
from jw_core.fidelity.nli_providers.ollama_nli import OllamaNLI
def test_ollama_unavailable_when_server_down() -> None:
p = OllamaNLI()
with respx.mock(assert_all_called=False) as router:
router.get("http://localhost:11434/api/tags").mock(
side_effect=httpx.ConnectError("ECONNREFUSED")
)
assert p.is_available() is False
def test_ollama_unavailable_when_model_missing() -> None:
p = OllamaNLI()
with respx.mock() as router:
router.get("http://localhost:11434/api/tags").mock(
return_value=httpx.Response(
200, json={"models": [{"name": "qwen2.5:7b"}]}
)
)
assert p.is_available() is False
def test_ollama_available_when_model_present() -> None:
p = OllamaNLI()
with respx.mock() as router:
router.get("http://localhost:11434/api/tags").mock(
return_value=httpx.Response(
200, json={"models": [{"name": "llama3.1:8b-instruct"}]}
)
)
assert p.is_available() is True
def test_ollama_evaluate_parses_entails() -> None:
p = OllamaNLI()
with respx.mock() as router:
router.get("http://localhost:11434/api/tags").mock(
return_value=httpx.Response(
200, json={"models": [{"name": "llama3.1:8b-instruct"}]}
)
)
router.post("http://localhost:11434/api/chat").mock(
return_value=httpx.Response(
200,
json={
"message": {
"content": json.dumps(
{"verdict": "entails", "score": 0.87, "reason": "ok"}
)
}
},
)
)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "entails"
assert v.score == 0.87
assert v.provider == "ollama-nli"
def test_ollama_fallback_on_garbage_response() -> None:
p = OllamaNLI()
with respx.mock() as router:
router.get("http://localhost:11434/api/tags").mock(
return_value=httpx.Response(
200, json={"models": [{"name": "llama3.1:8b-instruct"}]}
)
)
router.post("http://localhost:11434/api/chat").mock(
return_value=httpx.Response(
200, json={"message": {"content": "not even json"}}
)
)
v = p.evaluate(claim="A", premise="B")
assert v.verdict == "neutral"
assert v.score == 0.5
def test_ollama_uses_env_host(monkeypatch) -> None:
monkeypatch.setenv("OLLAMA_HOST", "http://example.local:9999")
p = OllamaNLI()
with respx.mock() as router:
router.get("http://example.local:9999/api/tags").mock(
return_value=httpx.Response(
200, json={"models": [{"name": "llama3.1:8b-instruct"}]}
)
)
assert p.is_available() is True
def test_ollama_uses_env_model(monkeypatch) -> None:
monkeypatch.setenv("JW_NLI_OLLAMA_MODEL", "qwen2.5:7b")
p = OllamaNLI()
with respx.mock() as router:
router.get("http://localhost:11434/api/tags").mock(
return_value=httpx.Response(
200, json={"models": [{"name": "qwen2.5:7b"}]}
)
)
chat = router.post("http://localhost:11434/api/chat").mock(
return_value=httpx.Response(
200,
json={
"message": {
"content": json.dumps(
{"verdict": "entails", "score": 0.9}
)
}
},
)
)
p.evaluate(claim="A", premise="B")
body = json.loads(chat.calls.last.request.content)
assert body["model"] == "qwen2.5:7b"
assert body["format"] == "json"- Step 2: Run tests to verify they fail
Run:
uv run pytest packages/jw-core/tests/test_fidelity_deberta.py packages/jw-core/tests/test_fidelity_ollama.py -vExpected: FAIL — both stub providers raise NotImplementedError.
- Step 3: Implement DeBERTaV3MNLI and OllamaNLI
# packages/jw-core/src/jw_core/fidelity/nli_providers/deberta_mnli.py
"""DeBERTaV3MNLI — local transformer NLI via HuggingFace.
Model: `MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli` (Apache-2.0,
~440MB). Multilingual fallback `MoritzLaurer/mDeBERTa-v3-base-mnli-xnli` is
selectable via env `JW_NLI_DEBERTA_MODEL`.
Three targets — auto-detected via `is_available()`:
- target="mlx" : requires `mlx-transformers` (Apple Silicon).
- target="nvidia" : requires `torch.cuda.is_available()`.
- target="cpu" : always works when `transformers + torch` installed.
Lazy load + singleton: the model is downloaded/loaded on the FIRST
`evaluate()` call, not at `__init__` (instantiation must stay cheap so the
factory can probe all three targets without loading anything).
Inference:
- tokenize as a pair-sequence (premise, claim).
- softmax 3 logits: [contradiction=0, neutral=1, entailment=2].
- verdict = argmax label; score = probability of that label.
- truncation: `max_length=512`, `truncation="only_first"` (preserves the
shorter `claim`, recovers room by trimming the `premise`).
"""
from __future__ import annotations
import logging
import os
import threading
from typing import Any
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict, ensure_verdict
logger = logging.getLogger(__name__)
_DEFAULT_MODEL = "MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli"
_LABELS: tuple[str, str, str] = ("contradicts", "neutral", "entails")
class DeBERTaV3MNLI:
name = "deberta-v3-mnli"
def __init__(self, *, target: Target = "cpu") -> None:
self.target: Target = target
self._model: Any | None = None
self._tokenizer: Any | None = None
self._device: str | None = None
self._lock = threading.Lock()
def is_available(self) -> bool:
# Common: need transformers + torch present.
try:
import torch # noqa: F401
import transformers # noqa: F401
except ImportError:
return False
if self.target == "cpu":
return True
if self.target == "nvidia":
try:
import torch
return bool(torch.cuda.is_available())
except Exception:
return False
if self.target == "mlx":
try:
import mlx_transformers # noqa: F401
except ImportError:
return False
return True
return False
def _ensure_loaded(self) -> None:
if self._model is not None and self._tokenizer is not None:
return
with self._lock:
if self._model is not None and self._tokenizer is not None:
return
import torch
from transformers import (
AutoModelForSequenceClassification,
AutoTokenizer,
)
model_id = os.getenv("JW_NLI_DEBERTA_MODEL", _DEFAULT_MODEL)
logger.info("Loading DeBERTa NLI model %s (target=%s)", model_id, self.target)
self._tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
if self.target == "nvidia" and torch.cuda.is_available():
self._device = "cuda"
elif self.target == "mlx":
# mlx_transformers handles device internally
self._device = "mlx"
else:
self._device = "cpu"
if self._device in {"cpu", "cuda"}:
model = model.to(self._device)
model.eval()
self._model = model
def evaluate(
self, claim: str, premise: str, *, language: str = "en"
) -> NLIVerdict:
# Tests can inject `_tokenizer` and `_model` directly to bypass _ensure_loaded.
if self._model is None or self._tokenizer is None:
self._ensure_loaded()
import torch
assert self._tokenizer is not None
assert self._model is not None
inputs = self._tokenizer(
premise,
claim,
return_tensors="pt",
truncation="only_first",
max_length=512,
)
if self._device in {"cuda"}:
inputs = {k: v.to("cuda") for k, v in inputs.items()} # type: ignore[union-attr]
with torch.no_grad():
out = self._model(**inputs)
probs = torch.softmax(out.logits, dim=-1).squeeze(0).tolist()
idx = int(max(range(3), key=lambda i: probs[i]))
verdict = _LABELS[idx]
score = float(probs[idx])
return ensure_verdict(
verdict=verdict,
score=score,
provider=self.name,
raw={
"probs": {
"contradicts": round(probs[0], 4),
"neutral": round(probs[1], 4),
"entails": round(probs[2], 4),
},
"target": self.target,
"device": self._device or "unknown",
"lang": language,
},
)
__all__ = ["DeBERTaV3MNLI"]# packages/jw-core/src/jw_core/fidelity/nli_providers/ollama_nli.py
"""OllamaNLI — local LLM judge via Ollama HTTP API.
Default model `llama3.1:8b-instruct` (env `JW_NLI_OLLAMA_MODEL`); endpoint
`http://localhost:11434` (env `OLLAMA_HOST`).
`is_available()` is cached per process: it sends one GET to `/api/tags`
and checks the configured model appears in the response. The cache is
invalidated when `JW_NLI_OLLAMA_MODEL` or `OLLAMA_HOST` change between calls.
Inference: POST `/api/chat` with `format=json`, parse the assistant message
content as JSON, fall back to neutral/0.5 on parse error.
"""
from __future__ import annotations
import json
import logging
import os
from typing import Any
import httpx
from jw_core.fidelity.nli import Target
from jw_core.fidelity.verdicts import NLIVerdict, ensure_verdict
logger = logging.getLogger(__name__)
_DEFAULT_MODEL = "llama3.1:8b-instruct"
_DEFAULT_HOST = "http://localhost:11434"
_SYSTEM_PROMPT = (
"You are an NLI judge. Decide if the CONCLUSION strictly entails from "
"the PREMISE. Reply JSON only: {verdict, score, reason}. verdict is one "
"of entails|neutral|contradicts; score is a float 0.0-1.0."
)
class OllamaNLI:
name = "ollama-nli"
target: Target = "cpu"
def __init__(self) -> None:
self._cache: tuple[str, str, bool] | None = None
def _host(self) -> str:
return os.getenv("OLLAMA_HOST", _DEFAULT_HOST).rstrip("/")
def _model(self) -> str:
return os.getenv("JW_NLI_OLLAMA_MODEL", _DEFAULT_MODEL)
def is_available(self) -> bool:
host = self._host()
model = self._model()
if self._cache and self._cache[0] == host and self._cache[1] == model:
return self._cache[2]
try:
r = httpx.get(f"{host}/api/tags", timeout=2.0)
r.raise_for_status()
tags = r.json().get("models", []) or []
ok = any(t.get("name") == model for t in tags)
except Exception as exc: # noqa: BLE001
logger.debug("OllamaNLI.is_available() probe failed: %r", exc)
ok = False
self._cache = (host, model, ok)
return ok
def evaluate(
self, claim: str, premise: str, *, language: str = "en"
) -> NLIVerdict:
host = self._host()
model = self._model()
user_body = (
f"PREMISE:\n{premise}\n\n"
f"CONCLUSION:\n{claim}\n\n"
f"Language: {language}"
)
try:
r = httpx.post(
f"{host}/api/chat",
json={
"model": model,
"stream": False,
"format": "json",
"messages": [
{"role": "system", "content": _SYSTEM_PROMPT},
{"role": "user", "content": user_body},
],
},
timeout=60.0,
)
r.raise_for_status()
text = str(r.json().get("message", {}).get("content", ""))
except Exception as exc: # noqa: BLE001
logger.warning("OllamaNLI call failed: %r", exc)
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"api_error": repr(exc)},
)
try:
data = json.loads(text)
verdict = str(data.get("verdict", "")).lower()
score = float(data.get("score", 0.5))
reason = str(data.get("reason", ""))
except Exception as exc: # noqa: BLE001
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"parse_error": str(exc), "raw_text": text[:500]},
)
if verdict not in {"entails", "neutral", "contradicts"}:
return ensure_verdict(
verdict="neutral",
score=0.5,
provider=self.name,
raw={"unexpected_verdict": verdict, "reason": reason},
)
return ensure_verdict(
verdict=verdict,
score=score,
provider=self.name,
raw={"reason": reason, "model": model, "host": host, "lang": language},
)
__all__ = ["OllamaNLI"]- Step 4: Run tests to verify they pass
Run:
uv run pytest packages/jw-core/tests/test_fidelity_deberta.py packages/jw-core/tests/test_fidelity_ollama.py -vExpected: DeBERTa: 7 passed (some may skip if transformers not installed in dev env). Ollama: 7 passed.
If respx is not yet a dev dep, add it:
uv add --dev --package jw-core respx- Step 5: Commit
git add packages/jw-core/src/jw_core/fidelity/nli_providers/deberta_mnli.py packages/jw-core/src/jw_core/fidelity/nli_providers/ollama_nli.py packages/jw-core/tests/test_fidelity_deberta.py packages/jw-core/tests/test_fidelity_ollama.py
git commit -m "feat(jw-core): DeBERTaV3MNLI (mlx/nvidia/cpu) + OllamaNLI providers"Task 8: fidelity_wrap decorator in jw-agents
Files:
-
Create:
packages/jw-agents/src/jw_agents/fidelity_wrap.py -
Create:
packages/jw-agents/tests/test_fidelity_wrap.py -
Step 1: Write the failing test
# packages/jw-agents/tests/test_fidelity_wrap.py
"""Tests for the @fidelity_wrap decorator.
Contract:
- Wraps an async function returning AgentResult.
- For each Finding, evaluates NLI claim=summary vs premise=excerpt.
- Stamps metadata: nli_verdict, nli_score, nli_provider.
- Skip rule: excerpt < min_excerpt_chars → nli_verdict="skipped".
- on_fail="warn" → append AgentResult.warnings.
- on_fail="reject" → drop finding + warning.
- on_fail="annotate_only" → just metadata, no warnings.
- Idempotent: applying twice doesn't duplicate metadata.
- Stamps AgentResult.metadata["nli_min_score"] and ["nli_on_fail"].
"""
from __future__ import annotations
import asyncio
import pytest
from jw_agents.base import AgentResult, Citation, Finding
from jw_agents.fidelity_wrap import fidelity_wrap
from jw_core.fidelity import NLIVerdict
from jw_core.fidelity.nli_providers.fakes import FakeNLI
def _result_with(findings: list[Finding]) -> AgentResult:
return AgentResult(query="q", agent_name="x", findings=findings)
def _finding(summary: str, excerpt: str, url: str = "https://wol.jw.org/x") -> Finding:
return Finding(
summary=summary,
excerpt=excerpt,
citation=Citation(url=url, title="t", kind="article"),
)
def _run(coro):
return asyncio.run(coro)
class StubProvider:
"""Provider returning a configured verdict regardless of input."""
name = "stub-nli"
target = "cpu"
def __init__(self, verdict: str, score: float) -> None:
self._verdict = verdict
self._score = score
self.calls: list[tuple[str, str, str]] = []
def is_available(self) -> bool:
return True
def evaluate(self, claim: str, premise: str, *, language: str = "en") -> NLIVerdict:
self.calls.append((claim, premise, language))
return NLIVerdict(verdict=self._verdict, score=self._score, provider=self.name, raw={}) # type: ignore[arg-type]
def test_warn_mode_keeps_finding_and_appends_warning() -> None:
prov = StubProvider("contradicts", 0.4)
base_finding = _finding(
summary="The Trinity is a Bible teaching.",
excerpt="The Trinity is not a Bible teaching, contrary to popular belief.",
)
@fidelity_wrap(min_score=0.7, on_fail="warn", provider=prov)
async def agent(question: str) -> AgentResult: # noqa: ARG001
return _result_with([base_finding])
r = _run(agent(question="?"))
assert len(r.findings) == 1
f = r.findings[0]
assert f.metadata["nli_verdict"] == "contradicts"
assert f.metadata["nli_score"] == 0.4
assert f.metadata["nli_provider"] == "stub-nli"
assert any("Low NLI fidelity" in w for w in r.warnings)
assert r.metadata["nli_min_score"] == 0.7
assert r.metadata["nli_on_fail"] == "warn"
def test_reject_mode_drops_finding() -> None:
prov = StubProvider("contradicts", 0.4)
@fidelity_wrap(min_score=0.7, on_fail="reject", provider=prov)
async def agent() -> AgentResult:
return _result_with([
_finding(summary="bad", excerpt="this is a long enough premise text"),
])
r = _run(agent())
assert r.findings == []
assert any("Rejected finding" in w for w in r.warnings)
def test_annotate_only_keeps_finding_no_warning() -> None:
prov = StubProvider("contradicts", 0.2)
@fidelity_wrap(min_score=0.7, on_fail="annotate_only", provider=prov)
async def agent() -> AgentResult:
return _result_with([
_finding(summary="x", excerpt="this is a long enough premise text"),
])
r = _run(agent())
assert len(r.findings) == 1
assert r.findings[0].metadata["nli_verdict"] == "contradicts"
assert r.warnings == []
def test_pass_verdict_keeps_finding_no_warning() -> None:
prov = StubProvider("entails", 0.95)
@fidelity_wrap(min_score=0.7, on_fail="reject", provider=prov)
async def agent() -> AgentResult:
return _result_with([
_finding(summary="x", excerpt="this is a long enough premise text"),
])
r = _run(agent())
assert len(r.findings) == 1
assert r.warnings == []
assert r.findings[0].metadata["nli_verdict"] == "entails"
def test_short_excerpt_is_skipped() -> None:
prov = StubProvider("contradicts", 0.0)
@fidelity_wrap(min_score=0.7, on_fail="reject", provider=prov, min_excerpt_chars=32)
async def agent() -> AgentResult:
return _result_with([_finding(summary="x", excerpt="Juan 3:16")])
r = _run(agent())
assert len(r.findings) == 1
assert r.findings[0].metadata["nli_verdict"] == "skipped"
# provider was NOT called for the short-excerpt finding
assert prov.calls == []
def test_idempotent_does_not_re_evaluate() -> None:
prov = StubProvider("entails", 0.9)
@fidelity_wrap(min_score=0.7, provider=prov)
@fidelity_wrap(min_score=0.7, provider=prov)
async def agent() -> AgentResult:
return _result_with([
_finding(summary="x", excerpt="long enough excerpt for evaluation here"),
])
r = _run(agent())
assert len(r.findings) == 1
# Provider called ONCE despite two layers of wrap.
assert len(prov.calls) == 1
def test_default_provider_falls_back_to_factory(monkeypatch) -> None:
# No `provider` kwarg → factory resolves FakeNLI when nothing else is wired.
monkeypatch.setenv("JW_NLI_PROVIDER", "fake-nli")
@fidelity_wrap(min_score=0.7)
async def agent() -> AgentResult:
return _result_with([
_finding(
summary="A test summary",
excerpt="a totally different premise that has nothing in common with the claim",
),
])
r = _run(agent())
assert r.findings[0].metadata["nli_provider"] == "fake-nli"
def test_language_is_propagated_from_result_metadata() -> None:
prov = StubProvider("entails", 0.9)
@fidelity_wrap(min_score=0.7, provider=prov)
async def agent() -> AgentResult:
res = _result_with([
_finding(summary="x", excerpt="long enough excerpt for evaluation here"),
])
res.metadata["language"] = "pt"
return res
_run(agent())
assert prov.calls[0][2] == "pt"
def test_concurrent_findings_each_get_metadata() -> None:
prov = StubProvider("entails", 0.9)
@fidelity_wrap(min_score=0.7, provider=prov)
async def agent() -> AgentResult:
return _result_with([
_finding(summary=f"summary {i}", excerpt=f"long enough excerpt #{i} for eval")
for i in range(5)
])
r = _run(agent())
assert len(r.findings) == 5
for f in r.findings:
assert "nli_verdict" in f.metadata
assert "nli_score" in f.metadata
assert "nli_provider" in f.metadata- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-agents/tests/test_fidelity_wrap.py -v
Expected: FAIL — ModuleNotFoundError: No module named 'jw_agents.fidelity_wrap'.
- Step 3: Implement the decorator
# packages/jw-agents/src/jw_agents/fidelity_wrap.py
"""@fidelity_wrap — wrap async agents to NLI-verify their findings.
Spec: docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md
§"Decorator".
Why async-aware: the toolkit's agents are all async (they fan-out HTTP
calls to wol.jw.org and chase finetune candidates). The decorator preserves
that interface — `await wrapped(...)` still returns an AgentResult.
Default behavior is `on_fail="warn"`: findings are NEVER dropped silently.
The only mode that modifies findings is `on_fail="reject"`, and it always
attaches a warning describing what was dropped.
Idempotence: we check `Finding.metadata` for an existing `nli_verdict`
and skip re-evaluation. Cheap, observable.
"""
from __future__ import annotations
from collections.abc import Awaitable, Callable
from functools import wraps
from typing import Any, Literal
from jw_agents.base import AgentResult
from jw_core.fidelity import NLIProvider
OnFail = Literal["warn", "reject", "annotate_only"]
def fidelity_wrap(
*,
min_score: float = 0.7,
on_fail: OnFail = "warn",
provider: NLIProvider | None = None,
min_excerpt_chars: int = 32,
) -> Callable[[Callable[..., Awaitable[AgentResult]]], Callable[..., Awaitable[AgentResult]]]:
"""Decorate an async agent to NLI-verify each Finding.
Args:
min_score: failure threshold. A verdict with `score < min_score`
(or any non-"entails" verdict) is treated as failure.
on_fail:
"annotate_only" → write nli_* metadata, no warning, no drop.
"warn" → also append a warning to AgentResult.warnings.
"reject" → also drop the finding from the result.
provider: explicit NLIProvider. None → resolved lazily via
`get_default_nli_provider()`.
min_excerpt_chars: excerpts shorter than this are not sent to the
provider; their `nli_verdict` is set to "skipped". Default 32 —
this filters out citations whose excerpt is just a bible
reference label (e.g. "John 3:16").
"""
def deco(
fn: Callable[..., Awaitable[AgentResult]],
) -> Callable[..., Awaitable[AgentResult]]:
@wraps(fn)
async def wrapper(*args: Any, **kwargs: Any) -> AgentResult:
result = await fn(*args, **kwargs)
# Resolve provider lazily so import jw_agents doesn't pull in
# heavy providers at import time.
local_provider = provider
if local_provider is None:
from jw_core.fidelity import get_default_nli_provider
local_provider = get_default_nli_provider()
language = str(result.metadata.get("language", "en"))
kept = []
for f in result.findings:
# Idempotence — if some outer layer already evaluated, skip.
if "nli_verdict" in f.metadata:
kept.append(f)
continue
if len(f.excerpt) < min_excerpt_chars:
f.metadata["nli_verdict"] = "skipped"
f.metadata["nli_score"] = None
f.metadata["nli_provider"] = local_provider.name
kept.append(f)
continue
verdict = local_provider.evaluate(
claim=f.summary,
premise=f.excerpt,
language=language,
)
f.metadata["nli_verdict"] = verdict.verdict
f.metadata["nli_score"] = round(verdict.score, 4)
f.metadata["nli_provider"] = verdict.provider
failed = verdict.verdict != "entails" or verdict.score < min_score
if not failed:
kept.append(f)
continue
if on_fail == "annotate_only":
kept.append(f)
elif on_fail == "warn":
result.warnings.append(
f"Low NLI fidelity ({verdict.verdict}, "
f"score={verdict.score:.2f}) for citation {f.citation.url}"
)
kept.append(f)
elif on_fail == "reject":
result.warnings.append(
f"Rejected finding (NLI={verdict.verdict}, "
f"score={verdict.score:.2f}) for citation {f.citation.url}"
)
# do not append — finding dropped
result.findings = kept
result.metadata["nli_min_score"] = min_score
result.metadata["nli_on_fail"] = on_fail
return result
return wrapper
return deco
__all__ = ["fidelity_wrap", "OnFail"]- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-agents/tests/test_fidelity_wrap.py -v
Expected: 9 passed.
- Step 5: Commit
git add packages/jw-agents/src/jw_agents/fidelity_wrap.py packages/jw-agents/tests/test_fidelity_wrap.py
git commit -m "feat(jw-agents): @fidelity_wrap decorator with warn/reject/annotate_only"Task 9: min_excerpt_chars skip logic (edge cases)
This task is explicit in the spec: it carves the contract for which findings get NLI-evaluated and which don’t. Task 8 already implements the basic skip; this task hardens it with edge-case tests.
Files:
-
Modify:
packages/jw-agents/tests/test_fidelity_wrap.py(append) -
Step 1: Write the failing tests
Append to packages/jw-agents/tests/test_fidelity_wrap.py:
# ──────────────────────────────────────────────────────────
# min_excerpt_chars edge cases (Task 9)
# ──────────────────────────────────────────────────────────
def test_skipped_finding_keeps_existing_metadata() -> None:
"""Existing metadata on the Finding must NOT be clobbered by skip."""
prov = StubProvider("contradicts", 0.0)
@fidelity_wrap(min_score=0.7, provider=prov, min_excerpt_chars=32)
async def agent() -> AgentResult:
f = _finding(summary="s", excerpt="too short")
f.metadata["source"] = "rag"
f.metadata["chunk_id"] = 42
return _result_with([f])
r = _run(agent())
f = r.findings[0]
assert f.metadata["nli_verdict"] == "skipped"
assert f.metadata["nli_score"] is None
assert f.metadata["source"] == "rag"
assert f.metadata["chunk_id"] == 42
def test_min_excerpt_chars_zero_evaluates_everything() -> None:
"""Setting min_excerpt_chars=0 must NOT skip even empty excerpts."""
prov = StubProvider("neutral", 0.5)
@fidelity_wrap(min_score=0.7, provider=prov, min_excerpt_chars=0)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="")])
r = _run(agent())
assert r.findings[0].metadata["nli_verdict"] == "neutral"
assert prov.calls == [("s", "", "en")]
def test_min_excerpt_chars_huge_skips_everything() -> None:
"""A huge min_excerpt_chars skips even multi-paragraph excerpts."""
prov = StubProvider("entails", 0.95)
@fidelity_wrap(min_score=0.7, provider=prov, min_excerpt_chars=100000)
async def agent() -> AgentResult:
return _result_with([
_finding(summary="s", excerpt="a paragraph of reasonable length here.")
])
r = _run(agent())
assert r.findings[0].metadata["nli_verdict"] == "skipped"
assert prov.calls == []
def test_skipped_finding_never_dropped_in_reject_mode() -> None:
"""Skipped findings survive `on_fail="reject"`."""
prov = StubProvider("contradicts", 0.0)
@fidelity_wrap(min_score=0.7, on_fail="reject", provider=prov, min_excerpt_chars=32)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="John 3:16")])
r = _run(agent())
assert len(r.findings) == 1
assert r.findings[0].metadata["nli_verdict"] == "skipped"
assert r.warnings == []
def test_excerpt_at_boundary_length_evaluated() -> None:
"""An excerpt of EXACTLY min_excerpt_chars is evaluated (boundary inclusive)."""
prov = StubProvider("entails", 0.95)
boundary_excerpt = "x" * 32
@fidelity_wrap(min_score=0.7, provider=prov, min_excerpt_chars=32)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt=boundary_excerpt)])
r = _run(agent())
assert r.findings[0].metadata["nli_verdict"] == "entails"
assert prov.calls == [("s", boundary_excerpt, "en")]
def test_excerpt_one_below_boundary_skipped() -> None:
"""An excerpt of (min_excerpt_chars - 1) IS skipped."""
prov = StubProvider("entails", 0.95)
@fidelity_wrap(min_score=0.7, provider=prov, min_excerpt_chars=32)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="x" * 31)])
r = _run(agent())
assert r.findings[0].metadata["nli_verdict"] == "skipped"
assert prov.calls == []- Step 2: Run tests to verify they fail
Run: uv run pytest packages/jw-agents/tests/test_fidelity_wrap.py::test_skipped_finding_keeps_existing_metadata packages/jw-agents/tests/test_fidelity_wrap.py::test_min_excerpt_chars_zero_evaluates_everything packages/jw-agents/tests/test_fidelity_wrap.py::test_min_excerpt_chars_huge_skips_everything packages/jw-agents/tests/test_fidelity_wrap.py::test_skipped_finding_never_dropped_in_reject_mode packages/jw-agents/tests/test_fidelity_wrap.py::test_excerpt_at_boundary_length_evaluated packages/jw-agents/tests/test_fidelity_wrap.py::test_excerpt_one_below_boundary_skipped -v
If all pass already because Task 8 happened to be correct: great, the implementation is robust. If some fail (likely test_min_excerpt_chars_zero_evaluates_everything because "" < 0 is False so it would skip — actually len("") < 0 is False, so 0 case works), only fix what broke.
If test_skipped_finding_never_dropped_in_reject_mode fails because the implementation drops “skipped” findings in reject mode, that means the skip path also needs to short-circuit before the reject branch. Re-read Task 8 impl: it appends to kept and continues before the failed/reject check. So "skipped" is NEVER dropped. Verify with pytest -v.
- Step 3: Implement (if any test failed)
If test_min_excerpt_chars_zero_evaluates_everything failed because the check was len(excerpt) < min_excerpt_chars (which with min_excerpt_chars=0 means < 0 → always False → never skip → correct). No change.
If test_skipped_finding_keeps_existing_metadata failed because the impl overwrote metadata instead of mutating in place: re-read Task 8 — f.metadata["nli_verdict"] = "skipped" mutates, doesn’t overwrite. No change.
If anything DOES fail unexpectedly, the most likely culprit is the boundary inclusion — confirm len(f.excerpt) < min_excerpt_chars is the correct check (strict <, so exactly-equal is evaluated).
- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-agents/tests/test_fidelity_wrap.py -v
Expected: 15 passed (9 from Task 8 + 6 new).
- Step 5: Commit
git add packages/jw-agents/tests/test_fidelity_wrap.py
git commit -m "test(jw-agents): harden min_excerpt_chars skip edge cases"Task 10: threshold modes (warn|reject) + default on_fail=“warn”
This task adds tests for the threshold semantics under multiple verdicts and confirms the default mode. Task 8 already implemented the modes; Task 10 nails the matrix down explicitly.
Files:
-
Modify:
packages/jw-agents/tests/test_fidelity_wrap.py(append) -
Step 1: Write the failing tests
Append to test_fidelity_wrap.py:
# ──────────────────────────────────────────────────────────
# Threshold matrix (Task 10)
# ──────────────────────────────────────────────────────────
@pytest.mark.parametrize(
("verdict", "score", "min_score", "expected_fail"),
[
("entails", 0.95, 0.7, False),
("entails", 0.71, 0.7, False),
("entails", 0.70, 0.7, False),
("entails", 0.69, 0.7, True), # score below threshold
("entails", 0.30, 0.7, True),
("neutral", 0.95, 0.7, True), # non-entails verdict
("neutral", 0.50, 0.7, True),
("contradicts", 0.95, 0.7, True),
("contradicts", 0.10, 0.7, True),
],
)
def test_threshold_matrix(verdict, score, min_score, expected_fail) -> None:
prov = StubProvider(verdict, score)
@fidelity_wrap(min_score=min_score, on_fail="warn", provider=prov)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="a sufficiently long excerpt to evaluate")])
r = _run(agent())
assert len(r.findings) == 1 # warn never drops
if expected_fail:
assert any("Low NLI fidelity" in w for w in r.warnings)
else:
assert r.warnings == []
def test_default_on_fail_is_warn() -> None:
"""Default behavior is `warn` — explicit test of the default."""
prov = StubProvider("contradicts", 0.1)
@fidelity_wrap(provider=prov) # no on_fail kwarg
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="a sufficiently long excerpt here")])
r = _run(agent())
assert len(r.findings) == 1
assert any("Low NLI fidelity" in w for w in r.warnings)
assert r.metadata["nli_on_fail"] == "warn"
def test_default_min_score_is_0_7() -> None:
"""Default min_score is 0.7."""
@fidelity_wrap(provider=StubProvider("entails", 0.5))
async def agent() -> AgentResult:
return _result_with([])
r = _run(agent())
assert r.metadata["nli_min_score"] == 0.7
def test_min_score_below_zero_is_permissive() -> None:
"""`min_score=0.0` accepts any entails verdict, however low."""
prov = StubProvider("entails", 0.0)
@fidelity_wrap(min_score=0.0, on_fail="reject", provider=prov)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="a sufficiently long excerpt here")])
r = _run(agent())
assert len(r.findings) == 1
assert r.warnings == []
def test_min_score_above_one_rejects_everything() -> None:
"""`min_score=1.01` rejects even perfect verdicts."""
prov = StubProvider("entails", 1.0)
@fidelity_wrap(min_score=1.01, on_fail="reject", provider=prov)
async def agent() -> AgentResult:
return _result_with([_finding(summary="s", excerpt="a sufficiently long excerpt here")])
r = _run(agent())
assert r.findings == []
assert any("Rejected finding" in w for w in r.warnings)
def test_reject_mode_does_not_drop_passing_findings() -> None:
prov = StubProvider("entails", 0.95)
@fidelity_wrap(min_score=0.7, on_fail="reject", provider=prov)
async def agent() -> AgentResult:
return _result_with([
_finding(summary=f"good {i}", excerpt=f"a sufficiently long excerpt #{i} here")
for i in range(3)
])
r = _run(agent())
assert len(r.findings) == 3
assert r.warnings == []- Step 2: Run tests to verify they pass (most should already)
Run: uv run pytest packages/jw-agents/tests/test_fidelity_wrap.py -v
Expected: 24 passed (15 prior + 9 from threshold matrix + 4 from singletons = 28; adjust per actual count).
- Step 3: Fix any unexpected failures
If test_min_score_below_zero_is_permissive fails because min_score=0.0 and verdict.score=0.0: the condition score < min_score is 0.0 < 0.0 → False → does not fail → finding kept. Correct.
If test_min_score_above_one_rejects_everything fails: condition 1.0 < 1.01 → True → fails → rejected. Correct.
No implementation changes expected for this task; it’s a contract-locking test set.
- Step 4: Run final pass
Run: uv run pytest packages/jw-agents/tests/test_fidelity_wrap.py -v
Expected: all pass.
- Step 5: Commit
git add packages/jw-agents/tests/test_fidelity_wrap.py
git commit -m "test(jw-agents): lock down threshold matrix + default mode contracts"Task 11: Integration test — wrap apologetics; 1984 existing tests stay green
Files:
-
Create:
packages/jw-agents/tests/test_fidelity_integration.py -
Step 1: Write the failing test
# packages/jw-agents/tests/test_fidelity_integration.py
"""Integration test: wrap a real agent (apologetics) and confirm the existing
agent contract is unchanged in the default `warn` mode with FakeNLI.
We import the real `apologetics` async function and patch its inner HTTP
calls with the existing project fixtures. The wrapped agent must:
- Return the same number of findings as the unwrapped version.
- Stamp every finding with nli_* metadata.
- Stamp result.metadata with nli_min_score / nli_on_fail.
- Not raise.
We do NOT exercise reject mode here — that's tested in test_fidelity_wrap.
This is the "the wiring works end-to-end on a real agent" test.
"""
from __future__ import annotations
import asyncio
import os
import pytest
from jw_agents.base import AgentResult, Citation, Finding
from jw_agents.fidelity_wrap import fidelity_wrap
@pytest.fixture(autouse=True)
def _force_fake_nli(monkeypatch) -> None:
monkeypatch.setenv("JW_NLI_PROVIDER", "fake-nli")
def _fake_apologetics() -> AgentResult:
"""A minimal stand-in for the real apologetics agent — same shape."""
return AgentResult(
query="¿Es la Trinidad bíblica?",
agent_name="apologetics",
findings=[
Finding(
summary="La Trinidad no es una enseñanza bíblica",
excerpt=(
"Las Escrituras presentan a Jehová como el único Dios verdadero, "
"mientras que Jesús es su Hijo. La doctrina trinitaria se "
"desarrolló siglos después."
),
citation=Citation(
url="https://wol.jw.org/es/wol/d/r4/lp-s/2003124",
title="¿Cree usted en la Trinidad?",
kind="article",
),
metadata={"source": "topic_index"},
),
Finding(
summary="Jesús es el Hijo de Dios, no Dios mismo",
excerpt="Juan 17:3 dice: 'Esto significa vida eterna, que lleguen a conocerte'.",
citation=Citation(
url="https://wol.jw.org/es/wol/b/r4/lp-s/nwt/E/2024/43/17",
title="Juan 17",
kind="verse",
),
metadata={"source": "verse_text"},
),
],
warnings=[],
metadata={"language": "es"},
)
def test_wrapped_apologetics_keeps_findings_and_stamps_metadata() -> None:
@fidelity_wrap(min_score=0.7, on_fail="warn")
async def apologetics(question: str) -> AgentResult: # noqa: ARG001
return _fake_apologetics()
result = asyncio.run(apologetics(question="¿Es la Trinidad bíblica?"))
assert result.agent_name == "apologetics"
assert len(result.findings) == 2
for f in result.findings:
assert "nli_verdict" in f.metadata
assert "nli_score" in f.metadata
assert "nli_provider" in f.metadata
assert f.metadata["nli_provider"] == "fake-nli"
assert result.metadata["nli_min_score"] == 0.7
assert result.metadata["nli_on_fail"] == "warn"
# The preexisting `language` metadata is preserved
assert result.metadata["language"] == "es"
def test_wrapped_warn_never_drops_in_default_mode() -> None:
@fidelity_wrap() # all defaults: min_score=0.7, on_fail="warn"
async def apologetics() -> AgentResult:
return _fake_apologetics()
before = _fake_apologetics()
after = asyncio.run(apologetics())
assert len(after.findings) == len(before.findings)
def test_existing_tests_still_pass_after_wrap_when_not_applied() -> None:
"""Verifies that simply having the decorator in the import path does NOT
leak side effects. Sanity check the import surface."""
from jw_agents import fidelity_wrap as fw_module
assert hasattr(fw_module, "fidelity_wrap")
# No global state mutation
assert not hasattr(fw_module, "_GLOBAL_PROVIDER")- Step 2: Run test to verify it fails (or passes immediately)
Run: uv run pytest packages/jw-agents/tests/test_fidelity_integration.py -v
Expected: 3 passed.
- Step 3: Run the full test suite — no regressions
Run:
uv run pytest packages/ -qExpected: previous 1984 tests + new tests all green. If any existing test now fails, the most likely cause is the new optional dep on respx for the Ollama tests — verify it’s installed via uv sync --all-packages --dev.
- Step 4: Add a smoke target for the wrapped apologetics in CI logs (optional)
Touch nothing — the integration test is the smoke. Skip to commit.
- Step 5: Commit
git add packages/jw-agents/tests/test_fidelity_integration.py
git commit -m "test(jw-agents): integration test — wrapped apologetics still passes contract"Task 12: CLI flag --fidelity on agent commands
Files:
-
Modify:
packages/jw-cli/src/jw_cli/commands/apologetics.py -
Modify:
packages/jw-cli/src/jw_cli/commands/verse.py -
Modify:
packages/jw-cli/src/jw_cli/commands/research.py -
Modify:
packages/jw-cli/src/jw_cli/commands/meeting.py -
Create:
packages/jw-cli/tests/test_cli_fidelity.py -
Step 1: Write the failing test
# packages/jw-cli/tests/test_cli_fidelity.py
"""Tests for the `--fidelity` flag exposed by jw-cli agent commands.
We don't run a full HTTP roundtrip — we patch the inner agent callable
with a stub via monkeypatch on the imported symbol.
"""
from __future__ import annotations
import asyncio
import pytest
from typer.testing import CliRunner
from jw_cli.main import app
from jw_agents.base import AgentResult, Citation, Finding
def _stub_result() -> AgentResult:
return AgentResult(
query="test",
agent_name="apologetics",
findings=[
Finding(
summary="The Trinity is a Bible teaching",
excerpt="The Trinity is not a Bible teaching, contrary to popular belief.",
citation=Citation(url="https://wol.jw.org/x", title="t", kind="article"),
)
],
metadata={"language": "en"},
)
@pytest.fixture(autouse=True)
def _force_fake_nli(monkeypatch) -> None:
monkeypatch.setenv("JW_NLI_PROVIDER", "fake-nli")
def _patch_agent(monkeypatch, module_path: str, attr: str) -> None:
async def fake(*args, **kwargs): # noqa: ARG001
return _stub_result()
import importlib
mod = importlib.import_module(module_path)
monkeypatch.setattr(mod, attr, fake)
def test_apologetics_fidelity_off_skips_wrapping(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.apologetics", "apologetics")
runner = CliRunner()
res = runner.invoke(app, ["apologetics", "--question", "x", "--fidelity", "off"])
assert res.exit_code == 0
# When off, no nli_* metadata in stdout
assert "nli_verdict" not in res.stdout
def test_apologetics_fidelity_warn_adds_metadata(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.apologetics", "apologetics")
runner = CliRunner()
res = runner.invoke(app, ["apologetics", "--question", "x", "--fidelity", "warn"])
assert res.exit_code == 0
assert "nli_verdict" in res.stdout
def test_apologetics_fidelity_reject_drops_bad_findings(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.apologetics", "apologetics")
runner = CliRunner()
res = runner.invoke(app, ["apologetics", "--question", "x", "--fidelity", "reject"])
assert res.exit_code == 0
# FakeNLI on this excerpt detects asymmetric negation → contradicts → reject
# The 'findings' array must be empty (or count=0)
assert "Rejected finding" in res.stdout or '"findings": []' in res.stdout
def test_apologetics_fidelity_invalid_raises(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.apologetics", "apologetics")
runner = CliRunner()
res = runner.invoke(app, ["apologetics", "--question", "x", "--fidelity", "bogus"])
assert res.exit_code != 0
def test_verse_explainer_fidelity_flag_exists(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.verse", "verse_explainer")
runner = CliRunner()
res = runner.invoke(app, ["verse", "--reference", "John 3:16", "--fidelity", "warn"])
assert res.exit_code == 0
def test_research_fidelity_flag_exists(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.research", "research_topic")
runner = CliRunner()
res = runner.invoke(app, ["research", "--topic", "kingdom", "--fidelity", "warn"])
assert res.exit_code == 0
def test_meeting_fidelity_flag_exists(monkeypatch) -> None:
_patch_agent(monkeypatch, "jw_cli.commands.meeting", "meeting_helper")
runner = CliRunner()
res = runner.invoke(app, ["meeting", "--url-or-ref", "Romans 12:1", "--fidelity", "warn"])
assert res.exit_code == 0- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-cli/tests/test_cli_fidelity.py -v
Expected: FAIL — Typer raises on unknown option --fidelity.
- Step 3: Implement the flag in each CLI command
For each of the four agent commands, do the same surgery. Here’s the pattern using apologetics.py as the canonical example; replicate for verse.py, research.py, meeting.py mapping their argument names.
# packages/jw-cli/src/jw_cli/commands/apologetics.py — surgery sketch
"""`jw apologetics` — answer apologetics questions with citations.
Adds `--fidelity {off,warn,reject}` (default `warn`) which wraps the
agent call with @fidelity_wrap before invocation.
"""
from __future__ import annotations
import asyncio
import json
from typing import Literal
import typer
from jw_agents.apologetics import apologetics
from jw_agents.fidelity_wrap import fidelity_wrap
Fidelity = Literal["off", "warn", "reject"]
def apologetics_cmd(
question: str = typer.Option(..., "--question", help="Question to answer."),
language: str = typer.Option("en", "--language", help="Language code."),
fidelity: Fidelity = typer.Option(
"warn",
"--fidelity",
help="NLI runtime verification: off | warn | reject.",
case_sensitive=False,
),
) -> None:
if fidelity == "off":
callable_agent = apologetics
else:
callable_agent = fidelity_wrap(
min_score=0.7,
on_fail="reject" if fidelity == "reject" else "warn",
)(apologetics)
result = asyncio.run(callable_agent(question=question, language=language))
typer.echo(json.dumps(result.to_dict(), indent=2, ensure_ascii=False))(The actual file probably has more options already; only ADD the fidelity parameter and the conditional wrapping. Do NOT rewrite the rest.)
Repeat the same 5-line surgery for the other three CLI commands:
-
packages/jw-cli/src/jw_cli/commands/verse.py— wrapsverse_explainer. -
packages/jw-cli/src/jw_cli/commands/research.py— wrapsresearch_topic. -
packages/jw-cli/src/jw_cli/commands/meeting.py— wrapsmeeting_helper. -
Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-cli/tests/test_cli_fidelity.py -v
Expected: 7 passed.
- Step 5: Smoke-test the actual binaries
JW_NLI_PROVIDER=fake-nli uv run jw apologetics --question "test" --fidelity warn --helpExpected: help text includes --fidelity and lists the three values.
- Step 6: Commit
git add packages/jw-cli/src/jw_cli/commands packages/jw-cli/tests/test_cli_fidelity.py
git commit -m "feat(jw-cli): --fidelity flag on apologetics/verse/research/meeting commands"Task 13: MCP integration — evaluate_nli tool + fidelity param on agent tools
Files:
-
Modify:
packages/jw-mcp/src/jw_mcp/server.py -
Create:
packages/jw-mcp/tests/test_mcp_nli.py -
Step 1: Write the failing test
# packages/jw-mcp/tests/test_mcp_nli.py
"""Tests for jw-mcp NLI integrations:
1. New standalone tool `evaluate_nli(claim, premise, language)` returns
{"verdict", "score", "provider"}.
2. The wrapped agent tools (`apologetics_tool` et al.) accept an optional
`fidelity` parameter and return findings with nli_* metadata.
"""
from __future__ import annotations
import pytest
# The MCP server exposes the tool function directly for unit testing.
@pytest.fixture(autouse=True)
def _force_fake_nli(monkeypatch) -> None:
monkeypatch.setenv("JW_NLI_PROVIDER", "fake-nli")
def test_evaluate_nli_returns_verdict() -> None:
from jw_mcp.server import evaluate_nli
out = evaluate_nli(
claim="God loves the world",
premise="God so loved the world that he gave his only Son",
language="en",
)
assert "verdict" in out
assert "score" in out
assert "provider" in out
assert out["verdict"] in {"entails", "neutral", "contradicts"}
assert 0.0 <= out["score"] <= 1.0
assert out["provider"] == "fake-nli"
def test_evaluate_nli_default_language_is_en() -> None:
from jw_mcp.server import evaluate_nli
out = evaluate_nli(claim="a", premise="a")
assert out["verdict"] in {"entails", "neutral", "contradicts"}
def test_evaluate_nli_handles_empty_inputs() -> None:
from jw_mcp.server import evaluate_nli
out = evaluate_nli(claim="", premise="")
assert out["verdict"] in {"entails", "neutral", "contradicts"}
assert 0.0 <= out["score"] <= 1.0
def test_apologetics_tool_accepts_fidelity_param(monkeypatch) -> None:
"""The MCP wrapper around `apologetics` exposes `fidelity`."""
from jw_mcp import server as srv
async def fake(question: str, language: str = "en", **_): # noqa: ARG001
from jw_agents.base import AgentResult, Citation, Finding
return AgentResult(
query=question,
agent_name="apologetics",
findings=[
Finding(
summary="x",
excerpt="a sufficiently long excerpt for NLI evaluation here",
citation=Citation(url="https://wol.jw.org/x", title="t", kind="article"),
)
],
metadata={"language": language},
)
monkeypatch.setattr(srv, "apologetics", fake)
out = srv.apologetics_tool(question="?", language="en", fidelity="warn")
assert "findings" in out
assert out["findings"][0]["metadata"]["nli_verdict"] in {
"entails",
"neutral",
"contradicts",
"skipped",
}
def test_apologetics_tool_fidelity_off_skips_metadata(monkeypatch) -> None:
from jw_mcp import server as srv
async def fake(question: str, language: str = "en", **_): # noqa: ARG001
from jw_agents.base import AgentResult, Citation, Finding
return AgentResult(
query=question,
agent_name="apologetics",
findings=[
Finding(
summary="x",
excerpt="a sufficiently long excerpt for NLI evaluation here",
citation=Citation(url="https://wol.jw.org/x", title="t", kind="article"),
)
],
metadata={"language": language},
)
monkeypatch.setattr(srv, "apologetics", fake)
out = srv.apologetics_tool(question="?", language="en", fidelity="off")
assert "nli_verdict" not in out["findings"][0]["metadata"]- Step 2: Run test to verify it fails
Run: uv run pytest packages/jw-mcp/tests/test_mcp_nli.py -v
Expected: FAIL — evaluate_nli not exported; apologetics_tool does not accept fidelity.
- Step 3: Implement the MCP integrations
Append to packages/jw-mcp/src/jw_mcp/server.py:
# packages/jw-mcp/src/jw_mcp/server.py — additions
from __future__ import annotations
import asyncio
from typing import Literal
from jw_agents.apologetics import apologetics # if not already imported
from jw_agents.fidelity_wrap import fidelity_wrap
from jw_core.fidelity import evaluate_entailment
Fidelity = Literal["off", "warn", "reject"]
def _maybe_wrap(fn, fidelity: Fidelity):
if fidelity == "off":
return fn
return fidelity_wrap(
min_score=0.7,
on_fail="reject" if fidelity == "reject" else "warn",
)(fn)
# ── New standalone tool ──────────────────────────────────────────────
@mcp.tool()
def evaluate_nli(
claim: str,
premise: str,
language: str = "en",
) -> dict:
"""Run a single NLI judgement on a (claim, premise) pair.
Useful for clients that want to verify a citation/summary pair without
running a full agent. Uses the same provider stack as the @fidelity_wrap
decorator, honoring `JW_NLI_PROVIDER`.
Returns:
{"verdict": "entails"|"neutral"|"contradicts",
"score": float in [0, 1],
"provider": str}
"""
v = evaluate_entailment(claim=claim, premise=premise, language=language)
return {"verdict": v.verdict, "score": round(v.score, 4), "provider": v.provider}
# ── Wrap existing agent tools ────────────────────────────────────────
# Each existing `*_tool` function gains an optional `fidelity` parameter.
# We don't rewrite them — we add a thin wrapper. Below is the apologetics
# example; replicate for verse_explainer_tool, research_topic_tool,
# meeting_helper_tool.
@mcp.tool()
def apologetics_tool(
question: str,
language: str = "en",
fidelity: Fidelity = "warn",
) -> dict:
"""Run the apologetics agent with optional runtime NLI verification.
Args:
question: The apologetics question.
language: BCP-47 code, default "en".
fidelity: "off" (no NLI), "warn" (annotate + warn), or "reject"
(annotate + drop low-fidelity findings). Default "warn".
"""
callable_agent = _maybe_wrap(apologetics, fidelity)
result = asyncio.run(callable_agent(question=question, language=language))
return result.to_dict()Apply the same fidelity parameter pattern to verse_explainer_tool, research_topic_tool, meeting_helper_tool — each gets the fidelity: Fidelity = "warn" arg and routes through _maybe_wrap.
- Step 4: Run test to verify it passes
Run: uv run pytest packages/jw-mcp/tests/test_mcp_nli.py -v
Expected: 5 passed.
- Step 5: Smoke-test the MCP server
JW_NLI_PROVIDER=fake-nli uv run python -c "
from jw_mcp.server import evaluate_nli
print(evaluate_nli(claim='Jesus is God', premise='Jesus is not God', language='en'))
"Expected output (approximately): {'verdict': 'contradicts', 'score': 0.6, 'provider': 'fake-nli'} (score varies with jaccard).
- Step 6: Commit
git add packages/jw-mcp/src/jw_mcp/server.py packages/jw-mcp/tests/test_mcp_nli.py
git commit -m "feat(jw-mcp): evaluate_nli tool + fidelity param on agent tools"Task 14: Property test — random claim/premise pairs, verdicts consistent
Files:
-
Create:
packages/jw-core/tests/test_fidelity_property.py -
Step 1: Write the failing test
# packages/jw-core/tests/test_fidelity_property.py
"""Property-based tests for the NLI providers.
We use `hypothesis` (already a dev dep — see existing
test_property_based.py) to generate random text pairs and assert
invariants the providers MUST always honor:
- verdict ∈ {"entails", "neutral", "contradicts"}
- 0 ≤ score ≤ 1
- provider == name
- identical input → identical output (determinism, for FakeNLI)
- swapping claim and premise can change the verdict but never break
the type contract
"""
from __future__ import annotations
import string
from hypothesis import HealthCheck, given, settings
from hypothesis import strategies as st
from jw_core.fidelity.nli_providers.fakes import FakeNLI
# Restrict to printable ASCII to avoid byte-level issues in CI logs
_TEXT = st.text(
alphabet=string.ascii_letters + string.digits + " .,;:!?",
min_size=0,
max_size=200,
)
@given(claim=_TEXT, premise=_TEXT)
@settings(max_examples=200, suppress_health_check=[HealthCheck.function_scoped_fixture])
def test_fake_verdict_always_legal(claim: str, premise: str) -> None:
v = FakeNLI().evaluate(claim=claim, premise=premise)
assert v.verdict in {"entails", "neutral", "contradicts"}
assert 0.0 <= v.score <= 1.0
assert v.provider == "fake-nli"
@given(claim=_TEXT, premise=_TEXT)
@settings(max_examples=200)
def test_fake_is_deterministic(claim: str, premise: str) -> None:
p = FakeNLI()
assert p.evaluate(claim=claim, premise=premise) == p.evaluate(
claim=claim, premise=premise
)
@given(text=_TEXT.filter(lambda s: len(s) >= 4))
@settings(max_examples=100)
def test_fake_self_entailment_is_high(text: str) -> None:
v = FakeNLI().evaluate(claim=text, premise=text)
# When claim == premise, jaccard = 1.0 unless both empty after tokenize
assert v.score >= 0.99 or v.verdict == "neutral"
@given(claim=_TEXT, premise=_TEXT, language=st.sampled_from(["en", "es", "pt", "fr", "de"]))
@settings(max_examples=200)
def test_language_does_not_break_fake(claim: str, premise: str, language: str) -> None:
v = FakeNLI().evaluate(claim=claim, premise=premise, language=language)
assert v.raw["lang"] == language
@given(claim=_TEXT, premise=_TEXT)
@settings(max_examples=200)
def test_swap_preserves_type_contract(claim: str, premise: str) -> None:
p = FakeNLI()
a = p.evaluate(claim=claim, premise=premise)
b = p.evaluate(claim=premise, premise=claim)
# both legal verdicts
assert a.verdict in {"entails", "neutral", "contradicts"}
assert b.verdict in {"entails", "neutral", "contradicts"}
# scores both in [0, 1]
assert 0.0 <= a.score <= 1.0
assert 0.0 <= b.score <= 1.0
@given(claim=_TEXT, premise=_TEXT)
@settings(max_examples=50)
def test_score_is_finite(claim: str, premise: str) -> None:
import math
v = FakeNLI().evaluate(claim=claim, premise=premise)
assert math.isfinite(v.score)- Step 2: Run test to verify it passes
Run: uv run pytest packages/jw-core/tests/test_fidelity_property.py -v
Expected: 6 passed (hypothesis explores ~200 cases per property).
If hypothesis is missing in jw-core dev deps, add it:
uv add --dev --package jw-core hypothesis- Step 3: Fix any hypothesis-found failures
The most likely failure surfaces from test_fake_self_entailment_is_high — when the input text contains only punctuation/whitespace, the tokenizer returns an empty set and jaccard = 0.0 (per the implementation). The test allows this via the or v.verdict == "neutral" clause. If hypothesis finds a corner case the test didn’t anticipate (e.g. all-punctuation input where verdict is unexpectedly “contradicts”), narrow the input filter or add to the conditional.
- Step 4: Commit
git add packages/jw-core/tests/test_fidelity_property.py
git commit -m "test(jw-core): hypothesis property tests for FakeNLI invariants"Task 15: Documentation — user guide + ROADMAP + VISION_AUDIT
Files:
-
Create:
docs/guias/fidelity-nli.md -
Modify:
docs/README.md -
Modify:
docs/VISION_AUDIT.md -
Modify:
docs/ROADMAP.md -
Step 1: Write the user guide
# Fidelidad NLI en runtime (`jw_core.fidelity`)
> Fase 39 — verificación de entailment semántico claim ↔ premise sobre cada `Finding` que devuelve un agente. Spec: `docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md`.
## Para qué sirve
Garantiza, en cada llamada real, que el `summary` de un `Finding` se desprende lógicamente del `excerpt` verbatim que su `Citation` ancla. Complementa Fase 22 (eval offline pre-merge) extendiendo la red al runtime.
Cada finding verificado lleva en `metadata`:
```json
{
"nli_verdict": "entails | neutral | contradicts | skipped",
"nli_score": 0.87,
"nli_provider": "claude-nli"
}Modos de operación
| Modo | Qué hace | Cuándo |
|---|---|---|
off | No evalúa, no anota. | CLI con --fidelity off para máxima velocidad. |
annotate_only | Sólo añade metadata, sin warnings ni drops. | Uso programático, telemetría. |
warn (default) | Metadata + warning en AgentResult.warnings si score < threshold. | CLI y MCP por defecto. |
reject | Warn + DROP del finding del resultado. | Superficies estrictas (--fidelity reject). |
Providers disponibles
Orden de auto-detección (puede sobreescribirse con JW_NLI_PROVIDER):
claude-nli— Anthropic Claude (mejor calidad, multi-lingüe). Extra[nli-anthropic]+ANTHROPIC_API_KEY.openai-nli— OpenAI GPT-4o-mini. Extra[nli-openai]+OPENAI_API_KEY.deberta-v3-mnli— DeBERTa-v3-large-mnli, local. Extra[nli-local](instala torch + transformers). Detecta automáticamente Apple Silicon (MLX), CUDA (NVIDIA), CPU.ollama-nli— Llama 3.1 local vía Ollama HTTP. Requiereollama servecorriendo.fake-nli— heurística pura (jaccard + negación). Siempre disponible, determinista, sin red. Default en CI.
Uso desde CLI
# Modo warn (default) — siempre se anota, warnings si falla
uv run jw apologetics --question "¿Es la Trinidad bíblica?" --fidelity warn
# Off (sin verificación, máxima velocidad)
uv run jw apologetics --question "?" --fidelity off
# Reject (drop estricto de findings que no aprueban)
uv run jw apologetics --question "?" --fidelity reject
# Forzar provider específico
JW_NLI_PROVIDER=claude-nli uv run jw verse --reference "Juan 3:16" --fidelity warnUso desde MCP
Cada tool de agente (apologetics_tool, verse_explainer_tool, research_topic_tool, meeting_helper_tool) gana un parámetro opcional fidelity con los mismos valores. Nuevo tool standalone:
{
"name": "evaluate_nli",
"arguments": {
"claim": "La Trinidad no es bíblica",
"premise": "Las Escrituras presentan a un solo Dios",
"language": "es"
}
}Devuelve {"verdict": "entails|neutral|contradicts", "score": 0.87, "provider": "claude-nli"}.
Uso desde Python
from jw_core.fidelity import evaluate_entailment
v = evaluate_entailment(
claim="The Trinity is not a Bible teaching.",
premise="The Bible teaches there is one God, the Father.",
language="en",
)
print(v.verdict, v.score, v.provider)Para envolver un agente custom:
from jw_agents.fidelity_wrap import fidelity_wrap
@fidelity_wrap(min_score=0.7, on_fail="warn")
async def my_agent(question: str) -> AgentResult:
...Variables de entorno
| Variable | Default | Efecto |
|---|---|---|
JW_NLI_PROVIDER | (auto) | Override: claude-nli, openai-nli, deberta-v3-mnli, ollama-nli, fake-nli. |
JW_NLI_CLAUDE_MODEL | claude-sonnet-4-5-20250929 | Modelo Anthropic. |
JW_NLI_OPENAI_MODEL | gpt-4o-mini | Modelo OpenAI. |
JW_NLI_OLLAMA_MODEL | llama3.1:8b-instruct | Modelo local Ollama. |
JW_NLI_DEBERTA_MODEL | MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli | Modelo HF. |
JW_PROVIDER_ORDER | api,mlx,nvidia,cpu | Reordena el ranking de targets (compartido con Fase 33). |
OLLAMA_HOST | http://localhost:11434 | Servidor Ollama. |
ANTHROPIC_API_KEY | — | Necesario para claude-nli. |
OPENAI_API_KEY | — | Necesario para openai-nli. |
Costes orientativos
| Provider | Coste por 1k findings (premise ≤2k tokens) | Latencia P50 |
|---|---|---|
claude-nli (Sonnet 4.5, con prompt caching) | ~$0.30 | ~250ms |
openai-nli (gpt-4o-mini) | ~$0.15 | ~400ms |
deberta-v3-mnli (CPU) | $0 | ~800ms |
deberta-v3-mnli (CUDA) | $0 | ~50ms |
ollama-nli (llama3.1:8b) | $0 | ~1500ms |
fake-nli | $0 | <1ms |
Troubleshooting
| Síntoma | Diagnóstico | Fix |
|---|---|---|
nli_verdict="skipped" en todos los findings | excerpts <32 chars | revisa parser; o baja min_excerpt_chars en el decorador |
nli_verdict="contradicts" en findings buenos | paráfrasis sinonímica + provider estricto | usa claude-nli o sube min_excerpt_chars |
RuntimeError: not available al iniciar | JW_NLI_PROVIDER apunta a un provider sin deps/keys | quita el env var o instala el extra correspondiente |
| ~1s/finding extra en CLI | DeBERTa CPU es lento | usa --fidelity off, o JW_NLI_PROVIDER=claude-nli |
| Costes API explotan | sin caching o muchos findings | habilita Anthropic prompt caching (default), baja agentes o usa fake-nli para dev |
Política para fases nuevas
Toda fase que añada un agente nuevo debe documentar si lo envuelve con @fidelity_wrap y bajo qué modo por defecto. Las superficies CLI/MCP heredan automáticamente el flag --fidelity cuando se basan en estos decoradores.
- [ ] **Step 2: Link from `docs/README.md`**
Add to the "Guías por tema" list (alphabetical position, before or after Eval doctrinal):
```markdown
- [Fidelidad NLI en runtime](guias/fidelity-nli.md) — Verificación NLI claim/premise sobre cada `Finding`; 4 providers + `FakeNLI`; CLI/MCP `--fidelity`.- Step 3: Add Fase 39 row to
docs/VISION_AUDIT.md
Insert a row in the summary table (alphabetical/numerical position with other Fase 3X rows):
| Fase 39 (nli-runtime) | ✅ Nuevo | `jw_core.fidelity` — 5 providers (Claude/OpenAI/DeBERTa/Ollama/Fake), `@fidelity_wrap`, CLI/MCP `--fidelity` |- Step 4: Append Fase 39 section to
docs/ROADMAP.md
After Fase 38, before any closing footer:
## Fase 39 — NLI runtime (fidelidad semántica) ✅
> Tier 1 confianza en runtime. Spec: `docs/superpowers/specs/2026-05-31-fase-39-nli-runtime-design.md`.
- ✅ Subpaquete nuevo `jw_core.fidelity` con Protocol + factory triple-target.
- ✅ Modelos: `NLIVerdict` (frozen dataclass) + `Verdict` Literal + `ensure_verdict` safe constructor.
- ✅ 5 providers: `ClaudeNLI` (api), `OpenAINLI` (api), `DeBERTaV3MNLI` (mlx/nvidia/cpu), `OllamaNLI` (cpu/local), `FakeNLI` (siempre).
- ✅ `JW_NLI_PROVIDER` env override + `JW_PROVIDER_ORDER` compartido con Fase 33.
- ✅ `@fidelity_wrap` decorator en `jw_agents/fidelity_wrap.py` con `on_fail={annotate_only,warn,reject}` y `min_excerpt_chars`.
- ✅ Idempotente: aplicar dos veces no duplica metadata.
- ✅ CLI flag `--fidelity {off,warn,reject}` en `apologetics`/`verse`/`research`/`meeting`.
- ✅ MCP: tool standalone `evaluate_nli` + parámetro `fidelity` en agent tools.
- ✅ Extras `[nli-anthropic]`, `[nli-openai]`, `[nli-local]`, `[nli-all]`.
- ✅ Guía `docs/guias/fidelity-nli.md`.
### Cobertura de tests
- ✅ ~70 tests nuevos en `packages/jw-core/tests/test_fidelity_*` y `packages/jw-agents/tests/test_fidelity_*`.
- ✅ Suite global sin regresiones (target: 1984 → 2050+).
- ✅ Property tests con hypothesis sobre 200+ pares aleatorios.- Step 5: Commit
git add docs/guias/fidelity-nli.md docs/README.md docs/VISION_AUDIT.md docs/ROADMAP.md
git commit -m "docs(fidelity): guide + ROADMAP/VISION_AUDIT for Fase 39"Task 16: Final audit — full suite green + no regressions
Files: none (verification only).
- Step 1: Run lint + format
uv run ruff check packages/jw-core packages/jw-agents packages/jw-cli packages/jw-mcp
uv run ruff format --check packages/jw-core packages/jw-agents packages/jw-cli packages/jw-mcpExpected: zero violations. If any, run uv run ruff check --fix and uv run ruff format, then re-commit as style(fidelity): ruff autofix.
- Step 2: Run mypy (best-effort)
uv run mypy packages/jw-core/src packages/jw-agents/srcExpected: existing errors only; no new errors from the fidelity module beyond # type: ignore[...] comments on third-party imports.
- Step 3: Run the entire test suite
uv run pytest packages/ -q --tb=shortExpected: prior 1984 tests + new ~70 tests all green. No regressions.
- Step 4: Smoke each provider that’s available locally
# FakeNLI (always works)
JW_NLI_PROVIDER=fake-nli uv run python -c "
from jw_core.fidelity import evaluate_entailment
print(evaluate_entailment(claim='Jesus is the Son of God', premise='Jesus is the Son of God who was sent by the Father.'))
"
# ClaudeNLI (only if ANTHROPIC_API_KEY is set in your shell)
if [ -n "$ANTHROPIC_API_KEY" ]; then
JW_NLI_PROVIDER=claude-nli uv run python -c "
from jw_core.fidelity import evaluate_entailment
print(evaluate_entailment(claim='Jesus is the Son of God', premise='Jesus is the Son of God who was sent by the Father.', language='en'))
"
fiExpected: each prints an NLIVerdict(...) line with verdict + score.
- Step 5: Smoke the CLI
JW_NLI_PROVIDER=fake-nli uv run jw apologetics --question "test" --fidelity warn --helpExpected: help text shows --fidelity option.
- Step 6: Update task status
When all above pass:
# Update task #67 in personal memory to completed
echo "Fase 39 complete." - Step 7: Final summary commit (optional polish)
If any minor doc tweaks emerged: git commit -am "docs(fidelity): post-audit polish". Otherwise nothing to do.
Self-review summary
-
Spec coverage: Each section of the spec maps to a task above:
- Architecture / Provider Protocol → Task 1.
- NLIVerdict + Verdict Literal → Task 2.
- FakeNLI determinístico → Task 3.
- Triple-target factory + env override → Task 4.
- ClaudeNLI / OpenAINLI / DeBERTaV3MNLI / OllamaNLI → Tasks 5, 6, 7.
- Decorator
@fidelity_wrap→ Task 8. min_excerpt_charsskip semantics → Task 9.- Threshold modes (warn|reject) + default → Task 10.
- Integration with existing agents + non-regression → Task 11.
- CLI flag
--fidelity→ Task 12. - MCP integrations (tool + param) → Task 13.
- Property tests for invariants → Task 14.
- Docs / ROADMAP / VISION_AUDIT → Task 15.
- Final audit → Task 16.
-
No-objetivos honored:
fact_checker(Fase 9) not touched — Fase 39 is observational only.- No persistent storage of verdicts (that’s Fase 43).
- No agent modifications in this PR beyond the decorator being available; the four agent CLI commands wire the wrap conditionally via
--fidelity, not by changing the agent body. - Decorator never re-evaluates idempotently.
- FakeNLI never makes network calls.
-
Extras documented:
[nli-anthropic],[nli-openai],[nli-local],[nli-all]declared in Task 1’spyproject.tomledit and referenced from the guide in Task 15. CI public stays onFakeNLI(no extras installed in standard job). -
No placeholders: every code block above is the actual code; every YAML/Toml is the actual content; every command shows the exact invocation and expected output. The stub providers in Task 4 are explicitly marked as “stub — overwritten in Task X” and each is replaced in its respective task with the full implementation.
-
Type consistency:
Target = Literal["api", "mlx", "nvidia", "cpu"]— same injw_core.fidelity.nli,jw_core.fidelity.factory, and every provider module.Verdict = Literal["entails", "neutral", "contradicts"]— single source of truth inverdicts.py, re-exported by__init__.NLIProviderProtocol signatureevaluate(self, claim: str, premise: str, *, language: str = "en") -> NLIVerdictmatches every provider implementation and theevaluate_entailmenthelper.ensure_verdictis the SINGLE constructor every provider funnels through (guarantees clamp + validation).@fidelity_wrapsignature stable across decorator chains (idempotence check onf.metadata["nli_verdict"]).
-
Test surface: total new tests across the plan: ~70 (10 verdicts + 4 protocol + 10 fakes + 8 factory + 10 claude + 7 openai + 7 deberta + 7 ollama + 9 wrap base + 6 wrap edge + 13 threshold matrix + 3 integration + 7 cli + 5 mcp + 6 property = 112). Adjust to ~70-100 depending on how parametrized matrices count. Either way, comfortably above the spec’s ”≥ 50 new tests” implicit floor.
-
Idempotence: explicitly tested (Task 8
test_idempotent_does_not_re_evaluate) and explicitly documented in the decorator docstring.
Execution choice
Plan completo. Dos opciones de ejecución:
- Subagent-driven (recomendado) — dispatch fresh sub-agente por tarea, review entre tareas, iteración rápida (
superpowers:subagent-driven-development). Las Tareas 1-7 son independientes hasta donde el factory necesita stubs, así que las Tareas 1, 2, 3 y los stubs de Task 4 se pueden completar en paralelo; Tareas 5, 6, 7 (providers reales) también son paralelizables entre sí. - Inline — ejecuto tareas en esta sesión con checkpoints (
superpowers:executing-plans). Apropiado si querés ver cada test rojo→verde en tiempo real.
¿Cuál prefieres?
Edit this page on docs/superpowers/plans/2026-05-31-fase-39-nli-runtime-plan.md