audit-4: classifier understands Spanish negations (veto unchanged)

Merchants write in Spanish; negations like nunca, lamentablemente or
no podemos must veto approval exactly like their English counterparts.
Previously an all-Spanish rejection mail could fall through to OTHER,
and a mixed mail with the approval word in English approved despite a
Spanish negation.

- negation list adds: nunca, jamas, tampoco, ningun/ninguna, nadie,
  nada, no podemos, no es posible, no aceptamos, rechaz*, descartad*,
  lamentablemente, no ha sido aprobad*, no procede
- new Spanish REJECTION rule (rechaz*, solicitud denegada, no cumple
  los requisitos) and Spanish APPROVAL rule (aprob*, aceptad*,
  bienvenid*, encantados de trabajar)
- veto rule unchanged: any negation anywhere -> needs_review
- tests: 15 new (12 veto cases incl. 2 mixed-language, pure-Spanish
  approval still approves, Spanish rejection classifies rejection)

Gate: ruff 0 · 96 unit + 1 PG integration · gitleaks clean.

alembic/versions/0004_audit_append_only.py

@@ -0,0 +1,89 @@
+"""audit_log append-only: reject UPDATE and DELETE (round-2 hardening)
+
+Revision ID: 0004
+Revises: 0003
+Create Date: 2026-10-28
+
+An audit trail that can be edited is not an audit trail. This migration
+installs, on every engine:
+
+- SQLite: BEFORE UPDATE/DELETE triggers that RAISE(ABORT).
+- Postgres: BEFORE UPDATE/DELETE triggers that RAISE EXCEPTION.
+
+The writer roles never had UPDATE/DELETE grants in roles.sql; this closes
+the hole for the table owner and for direct SQL connections.
+"""
+
+from __future__ import annotations
+
+from alembic import op
+
+revision = "0004"
+down_revision = "0003"
+branch_labels = None
+depends_on = None
+
+SQLITE_TRIGGERS = [
+    """
+    CREATE TRIGGER IF NOT EXISTS audit_log_no_update
+    BEFORE UPDATE ON audit_log
+    BEGIN
+      SELECT RAISE(ABORT, 'cayce-ops: audit_log is append-only (UPDATE rejected)');
+    END;
+    """,
+    """
+    CREATE TRIGGER IF NOT EXISTS audit_log_no_delete
+    BEFORE DELETE ON audit_log
+    BEGIN
+      SELECT RAISE(ABORT, 'cayce-ops: audit_log is append-only (DELETE rejected)');
+    END;
+    """,
+]
+
+PG_TRIGGERS = """
+CREATE OR REPLACE FUNCTION cayce_ops_audit_append_only() RETURNS trigger AS $$
+BEGIN
+  RAISE EXCEPTION 'cayce-ops: audit_log is append-only (% rejected)', TG_OP;
+END;
+$$ LANGUAGE plpgsql;
+"""
+
+
+def upgrade() -> None:
+    bind = op.get_bind()
+    if bind.dialect.name == "postgresql":
+        op.execute(PG_TRIGGERS)
+        op.execute(
+            "DROP TRIGGER IF EXISTS audit_log_no_update ON audit_log;"
+        )
+        op.execute(
+            "CREATE TRIGGER audit_log_no_update BEFORE UPDATE ON audit_log "
+            "FOR EACH ROW EXECUTE FUNCTION cayce_ops_audit_append_only();"
+        )
+        op.execute(
+            "DROP TRIGGER IF EXISTS audit_log_no_delete ON audit_log;"
+        )
+        op.execute(
+            "CREATE TRIGGER audit_log_no_delete BEFORE DELETE ON audit_log "
+            "FOR EACH ROW EXECUTE FUNCTION cayce_ops_audit_append_only();"
+        )
+    else:
+        for ddl in SQLITE_TRIGGERS:
+            bind.execute(sa_text(ddl))
+
+
+def sa_text(sql: str):
+    import sqlalchemy as sa
+
+    return sa.text(sql)
+
+
+def downgrade() -> None:
+    bind = op.get_bind()
+    if bind.dialect.name == "postgresql":
+        op.execute("DROP TRIGGER IF EXISTS audit_log_no_update ON audit_log;")
+        op.execute("DROP TRIGGER IF EXISTS audit_log_no_delete ON audit_log;")
+        op.execute("DROP FUNCTION IF EXISTS cayce_ops_audit_append_only();")
+    else:
+        bind.execute(sa_text("DROP TRIGGER IF EXISTS audit_log_no_update;"))
+        bind.execute(sa_text("DROP TRIGGER IF EXISTS audit_log_no_delete;"))

cayceops/components/maild/classify.py

@@ -19,10 +19,17 @@ from cayceops.models import MailClassification
 
 _APPROVAL_WORDS = re.compile(r"\b(approv|accept|welcome)", re.I)
 
-# any of these tokens anywhere in the mail vetoes an `approval` result
+# any of these tokens anywhere in the mail vetoes an `approval` result.
+# English and Spanish (merchants write in both; mixed mails are common).
 _NEGATIONS = re.compile(
     r"\b(not|no|never|cannot|can'?t|won'?t|hasn'?t|isn'?t|"
-    r"no\s+longer|unable|decline\w*|regret\w*|unfortunately)\b",
+    r"no\s+longer|unable|decline\w*|regret\w*|unfortunately|"
+    # spanish: standalone negation words
+    r"nunca|jam[aá]s|tampoco|ning[uú]n(?:a|o)?|nadie|nada|"
+    # spanish: multi-word negations and rejection stems
+    r"no\s+podemos|no\s+es\s+posible|no\s+acept(?:amos|ar[aá])|"
+    r"rechaz\w*|lamentablemente|descartad\w*|"
+    r"no\s+ha\s+sido\s+aprobad\w*|no\s+procede)\b",
     re.I,
 )
 
@@ -34,11 +41,24 @@ _RULES: list[tuple[MailClassification, re.Pattern[str], float]] = [
         re.compile(r"\b(declin|reject|not able to|unable to|we regret|won.t be able)\w*", re.I),  # noqa: E501
         0.9,
     ),
+    (
+        MailClassification.REJECTION,
+        # spanish rejections (strong stems, not just negation words)
+        re.compile(r"\b(rechaz\w*|no\s+vamos\s+a\s+pod(?:er|emos)|su\s+solicitud\s+ha\s+sido\s+denegada|no\s+cumple\s+los\s+requisitos)\b", re.I),  # noqa: E501
+        0.9,
+    ),
     (
         MailClassification.APPROVAL,
         re.compile(r"\b(approv|accept|welcome to the|you.re (in|approved)|live|activated|glad to)\w*", re.I),  # noqa: E501
         0.85,
     ),
+    (
+        MailClassification.APPROVAL,
+        # spanish approvals (standalone positive stems; the veto still
+        # downgrades if ANY negation appears elsewhere in the mail)
+        re.compile(r"\b(aprob\w*|aceptad\w*|bienvenid\w*|ha\s+sido\s+aceptada|encantados?\s+de\s+(trabajar|colaborar))\b", re.I),  # noqa: E501
+        0.85,
+    ),
     (
         MailClassification.QUESTION,
         re.compile(r"\b(could you|can you|please (provide|clarify|confirm)|what (is|are)|how (do|does|can)|need more (info|details))\b", re.I),  # noqa: E501

cayceops/components/monitor/monitor.py

@@ -251,3 +251,120 @@ def build(cfg: Config | None = None) -> int:
         return 0
     finally:
         engine().dispose()
+
+
+def _latest_chain_hash(sess) -> str | None:
+    """Last non-NULL entry_hash in the audit chain (None if chain empty)."""
+    from cayceops.models import AuditEntry
+
+    last = (
+        sess.query(AuditEntry)
+        .filter(AuditEntry.entry_hash.isnot(None))
+        .order_by(AuditEntry.id.desc())
+        .first()
+    )
+    return last.entry_hash if last else None
+
+
+def _published_anchors(sess) -> list[dict]:
+    """Anchor audit entries recorded locally, oldest first."""
+    from cayceops.models import AuditEntry
+
+    rows = (
+        sess.query(AuditEntry)
+        .filter(AuditEntry.action == "anchor")
+        .order_by(AuditEntry.id.asc())
+        .all()
+    )
+    return [
+        {
+            "event_id": (r.detail or {}).get("event_id"),
+            "last_hash": (r.detail or {}).get("last_hash"),
+            "last_entry": (r.detail or {}).get("last_entry"),
+            "ts": r.ts.isoformat(),
+        }
+        for r in rows
+    ]
+
+
+def verify_chain_against_anchors(
+    sess,
+    relay_fetcher=None,
+) -> dict:
+    """Compare the local chain tail with the anchors published to Nostr.
+
+    relay_fetcher(deployment) -> list of note contents (strings) found on
+    the relays for this deployment's anchor tag. When None, only local
+    coherence is checked and the result says so.
+
+    Returns a report dict:
+      status: "no_anchors" | "local_only" | "ok" | "mismatch"
+      detail: human-readable explanation
+      local_last_hash, local_anchors, relay_notes (when fetched)
+    """
+    local_hash = _latest_chain_hash(sess)
+    anchors = _published_anchors(sess)
+
+    report: dict = {
+        "local_last_hash": local_hash,
+        "local_anchors": anchors,
+        "relay_notes": None,
+    }
+
+    if not anchors:
+        report["status"] = "no_anchors"
+        report["detail"] = (
+            "the chain has never been anchored to Nostr — no external "
+            "attestation exists yet (anchor once with `monitor anchor`)"
+        )
+        return report
+
+    last_anchor = anchors[-1]
+    if relay_fetcher is None:
+        report["status"] = "local_only"
+        report["detail"] = (
+            "compared only against the LOCAL record of past anchors "
+            "(no relay fetch configured); last anchored hash "
+            f"{last_anchor['last_hash']}"
+        )
+        return report
+
+    from cayceops.config import load as _load
+
+    deployment = _load().deployment_id
+    notes = relay_fetcher(deployment) or []
+    report["relay_notes"] = notes
+
+    if not notes:
+        report["status"] = "no_anchors"
+        report["detail"] = (
+            "local anchor records exist but no anchor notes were found on "
+            "the relays — the external attestation is missing or was deleted"
+        )
+        return report
+
+    # the freshest relay note must attest the current chain tail
+    expected_fragment = f"last_hash={local_hash}"
+    latest_note = notes[-1]
+    if expected_fragment in latest_note:
+        report["status"] = "ok"
+        report["detail"] = "latest relay anchor matches the current chain tail"
+        return report
+
+    # maybe the chain advanced after the last anchor — that's fine ONLY if
+    # the note matches the hash recorded in the last local anchor entry
+    if last_anchor["last_hash"] and last_anchor["last_hash"] in latest_note:
+        report["status"] = "ok"
+        report["detail"] = (
+            "relay anchor matches the last local anchor; the chain has "
+            "advanced since (anchor again to attest the new tail)"
+        )
+        return report
+
+    report["status"] = "mismatch"
+    report["detail"] = (
+        f"relay anchor does not attest chain tail {local_hash} nor the last "
+        f"local anchor {last_anchor['last_hash']} — the published anchor "
+        "does not correspond to this database"
+    )
+    return report

cayceops/models.py

@@ -209,14 +209,21 @@ def _build_entry(
 def verify_audit_chain(sess) -> tuple[bool, int | None]:
     """Re-walk the chain. Returns (ok, first_broken_entry_id or None).
 
-    Legacy rows (entry_hash IS NULL, written before migration 0002) are
-    skipped and documented: the chain begins GENESIS at the first hashed row,
-    so a migrated database verifies clean without any tampering.
+    Rows with entry_hash IS NULL are ONLY tolerated BEFORE the first hashed
+    row (the documented legacy prefix from before migration 0002). A NULL
+    hash AFTER the chain has started is tampering — an attacker erasing the
+    hash of an intermediate row to hide an edit — and fails verification at
+    that row's id.
     """
     prev_hash: str | None = None  # None = still in legacy prefix
+    chain_started = False
     for entry in sess.query(AuditEntry).order_by(AuditEntry.id).all():
         if entry.entry_hash is None:
-            continue  # legacy row: outside the chain by documented genesis
+            if chain_started:
+                # tampering: a hash was erased in the middle of the chain
+                return False, entry.id
+            continue  # legacy prefix row: outside the chain by genesis
+        chain_started = True
         if prev_hash is None:
             expected_prev = "GENESIS"  # first hashed row anchors the chain
         else:

tests/test_audit_hardening.py

@@ -0,0 +1,189 @@
+"""Round-3 hardening tests: append-only audit log + strict legacy-prefix rule
++ anchor cross-verification (issue: operator key WAS the agent key; the
+audit trail must now be tamper-evident against the DB owner itself)."""
+
+from __future__ import annotations
+
+import pytest
+from sqlalchemy import text
+
+from cayceops.models import AuditEntry, audit, verify_audit_chain
+
+
+def _seed(sess, n: int = 3) -> list[int]:
+    ids = []
+    for i in range(n):
+        audit(sess, actor=f"daemon:{i}", action=f"act{i}", detail={"i": i})
+        sess.flush()
+        ids.append(sess.query(AuditEntry).order_by(AuditEntry.id.desc()).first().id)
+    return ids
+
+
+# --- 1. legacy-prefix rule -----------------------------------------------------
+
+
+def test_null_before_genesis_still_ok(sess):
+    """Documented legacy prefix: NULL hashes before the first hashed row."""
+    sess.add(AuditEntry(actor="legacy", action="old", detail={}, ts=__import__(
+        "datetime").datetime.now(__import__("datetime").UTC)))
+    sess.flush()
+    _seed(sess, 2)
+    sess.flush()
+    ok, broken = verify_audit_chain(sess)
+    assert ok and broken is None
+
+
+def test_null_in_middle_fails(sess):
+    """THE FIX: erasing a hash mid-chain is tampering and must be reported."""
+    ids = _seed(sess, 3)
+    middle = ids[1]
+    sess.execute(
+        text("UPDATE audit_log SET entry_hash = NULL WHERE id = :i"), {"i": middle}
+    )
+    sess.flush()
+    ok, broken = verify_audit_chain(sess)
+    assert not ok
+    assert broken == middle
+
+
+def test_null_after_chain_started_reports_the_row(sess):
+    ids = _seed(sess, 4)
+    sess.execute(
+        text("UPDATE audit_log SET entry_hash = NULL WHERE id = :i"), {"i": ids[3]}
+    )
+    sess.flush()
+    ok, broken = verify_audit_chain(sess)
+    assert not ok and broken == ids[3]
+
+
+# --- 2/3. append-only enforcement (migration 0004) ------------------------------
+
+
+def _apply_0004(sess):
+    import os
+
+    path = os.path.join(
+        os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+        "alembic",
+        "versions",
+        "0004_audit_append_only.py",
+    )
+    spec = {}
+    with open(path) as fh:
+        exec(compile(fh.read(), path, "exec"), spec)
+    # reuse the sqlite trigger DDL directly (dialect-aware inside the module)
+    for ddl in spec["SQLITE_TRIGGERS"]:
+        sess.execute(text(ddl))
+    sess.commit()
+
+
+def test_update_rejected_after_migration(sess):
+    _seed(sess, 1)
+    _apply_0004(sess)
+    with pytest.raises(Exception, match="append-only"):
+        sess.execute(text("UPDATE audit_log SET actor = 'evil'"))
+        sess.flush()
+
+
+def test_delete_rejected_after_migration(sess):
+    _seed(sess, 1)
+    _apply_0004(sess)
+    with pytest.raises(Exception, match="append-only"):
+        sess.execute(text("DELETE FROM audit_log"))
+        sess.flush()
+
+
+def test_insert_still_allowed_after_migration(sess):
+    _seed(sess, 1)
+    _apply_0004(sess)
+    audit(sess, actor="daemon:x", action="still-works", detail={})
+    sess.flush()
+    ok, broken = verify_audit_chain(sess)
+    assert ok and broken is None
+
+
+# --- 4. anchor cross-verification ----------------------------------------------
+
+
+def test_no_anchors_reported_explicitly(sess):
+    from cayceops.components.monitor.monitor import verify_chain_against_anchors
+
+    _seed(sess, 2)
+    report = verify_chain_against_anchors(sess)
+    assert report["status"] == "no_anchors"
+    assert "never been anchored" in report["detail"]
+
+
+def test_local_only_when_no_fetcher(sess):
+    from datetime import UTC, datetime
+
+    from cayceops.components.monitor.monitor import verify_chain_against_anchors
+
+    _seed(sess, 2)
+    last = sess.query(AuditEntry).order_by(AuditEntry.id.desc()).first()
+    sess.add(
+        AuditEntry(
+            actor="daemon:monitor",
+            action="anchor",
+            detail={
+                "event_id": "deadbeef",
+                "last_hash": last.entry_hash,
+                "last_entry": last.id,
+            },
+            ts=datetime.now(UTC),
+        )
+    )
+    sess.flush()
+    report = verify_chain_against_anchors(sess)
+    assert report["status"] == "local_only"
+    assert "LOCAL record" in report["detail"]
+
+
+def test_anchor_mismatch_detected(sess):
+    from datetime import UTC, datetime
+
+    from cayceops.components.monitor.monitor import verify_chain_against_anchors
+
+    _seed(sess, 2)
+    last = sess.query(AuditEntry).order_by(AuditEntry.id.desc()).first()
+    sess.add(
+        AuditEntry(
+            actor="daemon:monitor",
+            action="anchor",
+            detail={"event_id": "beef", "last_hash": last.entry_hash, "last_entry": last.id},
+            ts=datetime.now(UTC),
+        )
+    )
+    sess.flush()
+
+    def foreign_fetcher(deployment):
+        # a note attesting some OTHER chain entirely
+        return ["cayce-ops audit anchor deployment=itest last_hash=" + "aa" * 32]
+
+    report = verify_chain_against_anchors(sess, relay_fetcher=foreign_fetcher)
+    assert report["status"] == "mismatch"
+    assert "does not attest" in report["detail"]
+
+
+def test_anchor_match_ok(sess):
+    from datetime import UTC, datetime
+
+    from cayceops.components.monitor.monitor import verify_chain_against_anchors
+
+    _seed(sess, 2)
+    last = sess.query(AuditEntry).order_by(AuditEntry.id.desc()).first()
+    sess.add(
+        AuditEntry(
+            actor="daemon:monitor",
+            action="anchor",
+            detail={"event_id": "beef", "last_hash": last.entry_hash, "last_entry": last.id},
+            ts=datetime.now(UTC),
+        )
+    )
+    sess.flush()
+
+    def honest_fetcher(deployment):
+        return [f"cayce-ops audit anchor deployment={deployment} last_hash={last.entry_hash}"]
+
+    report = verify_chain_against_anchors(sess, relay_fetcher=honest_fetcher)
+    assert report["status"] == "ok"

tests/test_spanish_negations.py

@@ -0,0 +1,80 @@
+"""Spanish negation tests (round-3 point 4): merchants write in Spanish;
+negations like 'nunca', 'lamentablemente' or 'no podemos' must veto
+`approval` exactly like their English counterparts. Veto rule unchanged:
+when in doubt -> needs_review."""
+
+from __future__ import annotations
+
+import pytest
+
+from cayceops.components.maild import classify as clf
+from cayceops.models import MailClassification
+
+APPROVAL_ONLY = MailClassification.APPROVAL
+
+
+def _never_approval(cfg_factory, text: str) -> None:
+    kind, _ = clf.classify(text, cfg_factory())
+    assert kind is not APPROVAL_ONLY, f"{text!r} classified as APPROVAL"
+
+
+def test_english_baseline_still_works(cfg):
+    config = cfg()
+    kind, _ = clf.classify("Your application was approved, welcome to the program", config)
+    assert kind is APPROVAL_ONLY
+
+
+@pytest.mark.parametrize(
+    "text",
+    [
+        # 1. rejection stem
+        "Lamentamos informarle que su solicitud fue rechazada",
+        # 2. nunca
+        "Nunca aceptamos negocios de criptomonedas",
+        # 3. jamás
+        "Jamás hemos trabajado con affiliates externos",
+        # 4. tampoco
+        "No podemos aceptar su sitio, tampoco podemos ofrecerle otro acuerdo",
+        # 5. ningún/ninguna
+        "En este momento no tenemos ningún hueco para nuevos partners",
+        # 6. no podemos (multi-word)
+        "Su propuesta es interesante pero no podemos proceder",
+        # 7. no es posible (multi-word)
+        "Gracias por escribir, pero no es posible colaborar este trimestre",
+        # 8. lamentablemente
+        "Lamentablemente su aplicación no ha sido aprobada",
+        # 9. no aceptamos
+        "Por política interna no aceptamos sitios de viajes",
+        # 10. mixed spanish/english: negation in spanish, approval word in english
+        "Thanks for reaching out! Lamentablemente we cannot approve your site",
+        # 11. mixed: approval in spanish, negation in english
+        "Apreciamos su interés, but we are not accepting new partners",
+        # 12. nadie / nada
+        "Aquí nadie decidiría eso, así que no hay nada que aprobar",
+    ],
+)
+def test_spanish_and_mixed_negations_veto_approval(cfg, text):
+    _never_approval(cfg, text)
+
+
+def test_spanish_pure_approval_still_approves(cfg):
+    """A clean Spanish approval WITHOUT any negation must still classify
+    approval — the veto must not over-fire."""
+    config = cfg()
+    kind, confidence = clf.classify(
+        "Su solicitud fue aprobada. Bienvenido al programa, estamos encantados "
+        "de trabajar con ustedes.",
+        config,
+    )
+    assert kind is APPROVAL_ONLY
+    assert confidence >= 0.5
+
+
+def test_spanish_rejection_classifies_rejection(cfg):
+    config = cfg()
+    kind, conf = clf.classify(
+        "Lamentamos informarle que su solicitud fue rechazada por el comité",
+        config,
+    )
+    assert kind is MailClassification.REJECTION
+    assert conf >= 0.9