feat(contradictions): advisory scanner for conflicting approved claims

CHANGELOG.md

@@ -165,6 +165,14 @@ All notable changes to vouch are documented here. Format follows
 - demo: dual-path llm configuration — compile & summarize run through
   session-capture replay or directly against the api via a stdlib shim
   wired as `compile.llm_cmd` with a byo `ANTHROPIC_API_KEY`.
+- `vouch contradict-scan` — an offline scanner that groups approved claims
+  by shared entity and heuristically flags same-topic pairs that disagree
+  in polarity. `--dry-run` (default) only prints candidates; without it,
+  each surviving pair files a pending `contradicts` relation proposal via
+  `proposals.propose_relation` for a human `vouch approve` — the scanner
+  itself never writes a `Relation` or a `CONTESTED` status. `--threshold`,
+  `--entity`, and `--limit` tune the scan. scoring lives in the new
+  `src/vouch/contradictions.py`. (#314)
 
 ### Changed
 - ``kb.list_*`` JSONL/MCP responses now use a dict envelope

src/vouch/cli.py

@@ -28,6 +28,7 @@
 from . import capture as capture_mod
 from . import codex_rollout as codex_rollout_mod
 from . import compile as compile_mod
+from . import contradictions as contradictions_mod
 from . import digest as digest_mod
 from . import fetch as fetch_mod
 from . import inbox as inbox_mod
@@ -2984,6 +2985,43 @@ def dedup(threshold: float, dry_run: bool) -> None:
         click.echo(f"{r['kind']}/{r['id']} ~ {r['kind']}/{r['near_id']}  cos={r['cosine']:.4f}")
 
 
+@cli.command(name="contradict-scan")
+@click.option("--threshold", default=contradictions_mod.DEFAULT_THRESHOLD,
+              show_default=True, type=float)
+@click.option("--entity", default=None, help="Restrict the scan to one entity id.")
+@click.option("--dry-run/--no-dry-run", default=True, show_default=True)
+@click.option("--limit", default=None, type=int,
+              help="Cap the number of pairs proposed in one run.")
+def contradict_scan(
+    threshold: float, entity: str | None, dry_run: bool, limit: int | None,
+) -> None:
+    """Scan approved claims for candidate conflicting pairs.
+
+    Groups claims by shared entity and flags same-topic pairs that disagree
+    in polarity. Advisory only: with --dry-run (the default) this only
+    prints candidates. Without it, each surviving pair files a pending
+    `contradicts` relation proposal for a human `vouch approve` — it never
+    writes a Relation or a CONTESTED status itself.
+    """
+    store = _load_store()
+    with _cli_errors():
+        rows = contradictions_mod.scan(
+            store, threshold=threshold, entity=entity,
+            dry_run=dry_run, limit=limit, proposed_by=_whoami(),
+        )
+    if not rows:
+        click.echo("contradict-scan: no candidates found")
+        return
+    for r in rows:
+        line = (
+            f"{r['claim_a']} >< {r['claim_b']}  "
+            f"entity={r['entity']} score={r['score']:.3f}"
+        )
+        if "proposal_id" in r:
+            line += f"  proposal={r['proposal_id']}"
+        click.echo(line)
+
+
 @cli.group()
 def embeddings() -> None:
     """Embedding maintenance commands."""

src/vouch/contradictions.py

@@ -0,0 +1,188 @@
+"""Advisory contradiction scanner: flag approved claims that plausibly conflict.
+
+Heuristic and read-only. Groups approved claims by shared ``Claim.entities``
+and scores same-entity pairs whose text overlaps heavily but disagrees in
+polarity (one asserts, the other negates) — a cheap proxy for "conflicting
+values about the same subject". Nothing here mutates a Claim or writes a
+Relation; surviving pairs are filed as ordinary ``contradicts`` relation
+proposals via ``proposals.propose_relation``; every eventual edge or
+``ClaimStatus.CONTESTED`` transition still requires a human ``kb.approve``
+(or the pre-existing manual ``lifecycle.contradict``).
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import Any
+
+from .models import Claim, ClaimStatus, ProposalKind, ProposalStatus, RelationType
+from .proposals import ProposalError, propose_relation
+from .storage import KBStore
+
+DEFAULT_THRESHOLD = 0.3
+
+# No-longer-live assertions aren't worth flagging against.
+_INACTIVE_STATUSES = {ClaimStatus.SUPERSEDED, ClaimStatus.ARCHIVED, ClaimStatus.REDACTED}
+
+_NEGATIONS = {
+    "not", "no", "never", "none", "without", "isn't", "aren't", "wasn't",
+    "weren't", "don't", "doesn't", "didn't", "won't", "wouldn't", "can't",
+    "cannot", "couldn't", "shouldn't", "hasn't", "haven't", "hadn't",
+}
+
+_STOPWORDS = {
+    "the", "a", "an", "is", "are", "was", "were", "be", "been", "being",
+    "to", "of", "in", "on", "for", "and", "or", "but", "with", "that",
+    "this", "it", "as", "at", "by", "from", "into", "than", "then",
+}
+
+_WORD_RE = re.compile(r"[a-z0-9']+")
+
+
+def _tokens(text: str) -> set[str]:
+    return {w for w in _WORD_RE.findall(text.lower()) if w not in _STOPWORDS}
+
+
+def _has_negation(text: str) -> bool:
+    return bool(_tokens(text) & _NEGATIONS)
+
+
+def _text_overlap(a: str, b: str) -> float:
+    """Jaccard similarity of non-stopword tokens — a cheap same-topic proxy."""
+    ta, tb = _tokens(a), _tokens(b)
+    if not ta or not tb:
+        return 0.0
+    return len(ta & tb) / len(ta | tb)
+
+
+@dataclass(frozen=True)
+class Candidate:
+    """One candidate conflicting pair, scored and grouped by shared entity."""
+
+    claim_a: str
+    claim_b: str
+    entity: str
+    score: float
+
+
+def _entity_groups(
+    claims: list[Claim], *, entity: str | None,
+) -> dict[str, list[Claim]]:
+    groups: dict[str, list[Claim]] = {}
+    for c in claims:
+        for eid in c.entities:
+            if entity is not None and eid != entity:
+                continue
+            groups.setdefault(eid, []).append(c)
+    return groups
+
+
+def _already_flagged_pairs(store: KBStore) -> set[frozenset[str]]:
+    """Pairs that already have a contradiction on record, in any form.
+
+    Covers claims cross-linked via `Claim.contradicts` (set by
+    `lifecycle.contradict`), claims already joined by an approved
+    `RelationType.CONTRADICTS` edge (e.g. a prior scan's proposal that was
+    approved), and pairs with a still-pending `contradicts` relation
+    proposal from an earlier scan run. Any one of these means re-proposing
+    the pair would be a duplicate.
+    """
+    flagged: set[frozenset[str]] = set()
+    for c in store.list_claims():
+        for other in c.contradicts:
+            flagged.add(frozenset({c.id, other}))
+    for rel in store.list_relations():
+        if rel.relation == RelationType.CONTRADICTS:
+            flagged.add(frozenset({rel.source, rel.target}))
+    for p in store.list_proposals(ProposalStatus.PENDING):
+        if p.kind == ProposalKind.RELATION and p.payload.get("relation") == "contradicts":
+            src = p.payload.get("source")
+            target = p.payload.get("target")
+            if src and target:
+                flagged.add(frozenset({src, target}))
+    return flagged
+
+
+def find_candidates(
+    store: KBStore, *,
+    threshold: float = DEFAULT_THRESHOLD,
+    entity: str | None = None,
+) -> list[Candidate]:
+    """Read-only scan for candidate conflicting claim pairs.
+
+    Never mutates a Claim, writes a Relation, or touches `Claim.contradicts`.
+    """
+    claims = [c for c in store.list_claims() if c.status not in _INACTIVE_STATUSES]
+    groups = _entity_groups(claims, entity=entity)
+    flagged = _already_flagged_pairs(store)
+
+    best: dict[frozenset[str], Candidate] = {}
+    for eid, members in groups.items():
+        for i in range(len(members)):
+            for j in range(i + 1, len(members)):
+                a, b = members[i], members[j]
+                pair = frozenset({a.id, b.id})
+                if pair in flagged:
+                    continue
+                # Same-topic + opposite polarity is the conflict signal;
+                # same-topic + same polarity is corroboration or a near-dupe
+                # (embeddings/dedup.py's job), not a contradiction.
+                if _has_negation(a.text) == _has_negation(b.text):
+                    continue
+                score = _text_overlap(a.text, b.text)
+                if score < threshold:
+                    continue
+                existing = best.get(pair)
+                if existing is None or score > existing.score:
+                    best[pair] = Candidate(
+                        claim_a=a.id, claim_b=b.id, entity=eid, score=score,
+                    )
+    return sorted(best.values(), key=lambda c: c.score, reverse=True)
+
+
+def scan(
+    store: KBStore, *,
+    threshold: float = DEFAULT_THRESHOLD,
+    entity: str | None = None,
+    dry_run: bool = True,
+    limit: int | None = None,
+    proposed_by: str = "vouch-contradict-scan",
+) -> list[dict[str, Any]]:
+    """Scan for candidate pairs; file a `contradicts` proposal per pair unless dry-run.
+
+    Dry-run (the default) never writes anything. Otherwise each surviving
+    pair produces exactly one pending relation proposal via
+    `proposals.propose_relation` — visible in `kb.list_pending`, decided only
+    by a human `kb.approve`.
+    """
+    candidates = find_candidates(store, threshold=threshold, entity=entity)
+    if limit is not None:
+        candidates = candidates[:limit]
+
+    rows: list[dict[str, Any]] = []
+    for c in candidates:
+        row: dict[str, Any] = {
+            "claim_a": c.claim_a, "claim_b": c.claim_b,
+            "entity": c.entity, "score": c.score,
+        }
+        if not dry_run:
+            try:
+                proposal = propose_relation(
+                    store,
+                    src=c.claim_a,
+                    relation=RelationType.CONTRADICTS.value,
+                    target=c.claim_b,
+                    proposed_by=proposed_by,
+                    rationale=(
+                        f"contradict-scan: shared entity {c.entity!r}, "
+                        f"score={c.score:.3f}"
+                    ),
+                )
+            except ProposalError:
+                # Endpoint vanished or a proposal landed mid-scan (race with
+                # another run) — skip rather than fail the whole batch.
+                continue
+            row["proposal_id"] = proposal.id
+        rows.append(row)
+    return rows