feat(strategy): GNews guardrail — clamp news-only shifts to prior±0.25

Post-mortem NVIDIA 631181: one uncorroborated high-weight signal (legacy Manifold 0.13 at weight 0.6) flipped a 0.845 market to 0.431 and lost. With Manifold observational-only and macro signals gated behind is_non_price, GNews (weight 1.5) is the only live signal able to move politics markets 20-30 pp against the order-book consensus. This adds a catastrophic fuse, not a fine calibration: - apply_news_guardrail(): when |news_lo| >= NEWS_MATERIAL_LOGODDS_THRESHOLD (0.10) and every other signal (fg, mom, btc_dom, mfld) is below it, clamp the posterior to prior ± MAX_NEWS_ONLY_PROB_SHIFT (0.25). Any corroborating material signal disables the clamp. Config via env (NEWS_GUARDRAIL_ENABLED=true by default). - edge_gross/edge_net computed from the clamped posterior; raw_final_prob preserved in reasoning (persisted via trades.reasoning — no schema migration) and in the NEWS_MATERIAL log line. - guardrail_changed_trade_decision: raw edge crossed the regime gate but the clamped edge no longer does (fuse prevented a trade). Note: with the default 0.25 band the clamped edge_net is 0.21, above every regime minimum, so the flag only fires with a tighter configured band. - Observability gated on materiality: NEWS_MATERIAL per-market line and a compact NEWS SUMMARY cycle line, only when with_news > 0 — no flood from the ~145 news-less markets per cycle. - 9 deterministic tests (extreme clamp, in-band passthrough, corroboration, inclusive threshold, disabled, changed_decision). No changes to NEWS_LOGODDS_WEIGHT, Manifold flags, edge thresholds, sizing, payout, resolution, or historical trades. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-01 20:26:02 +00:00
parent 9e21ecac21
commit 7f84bc3ec7
3 changed files with 416 additions and 12 deletions
@@ -0,0 +1,247 @@
+"""
+Tests for the GNews guardrail (catastrophic fuse).
+
+Post-mortem NVIDIA 631181: one uncorroborated signal at high weight flipped a
+0.845 market to 0.431.  With Manifold observational-only and macro signals
+gated behind is_non_price, GNews is the only live signal able to move politics
+markets 20-30 pp against the order-book consensus.  The fuse clamps the
+posterior to prior ± MAX_NEWS_ONLY_PROB_SHIFT when GNews is the ONLY material
+signal (|log-odds| >= NEWS_MATERIAL_LOGODDS_THRESHOLD); any other material
+signal counts as corroboration and disables the clamp.
+
+Politics markets have no macro adjustments, so full-path tests exercise the
+"GNews only" branch naturally; the corroboration branch is tested through the
+pure helper apply_news_guardrail().
+
+evaluate() emits a NEWS_MATERIAL log line for every market whose news
+contribution is material (trade or skip); tests parse it via caplog.
+"""
+import asyncio
+import logging
+import math
+import re
+
+import pytest
+
+import bot.strategy.bayesian as bayesian
+from bot.data.external import ExternalSignals
+from bot.data.polymarket import Market
+from bot.strategy.bayesian import (
+    NEWS_LOGODDS_WEIGHT,
+    BayesianStrategy,
+    apply_news_guardrail,
+)
+
+NEWS_MATERIAL_RE = re.compile(
+    r"NEWS_MATERIAL.*raw=(\d+\.\d+) \| final=(\d+\.\d+).*"
+    r"guardrail=(applied|none) \| changed_decision=(true|false)"
+)
+
+
+def _logodds(p: float) -> float:
+    return math.log(p / (1 - p))
+
+
+def _sentiment_for(prior: float, target_raw: float) -> float:
+    """Sentiment that moves `prior` to exactly `target_raw` via GNews alone."""
+    return (_logodds(target_raw) - _logodds(prior)) / NEWS_LOGODDS_WEIGHT
+
+
+class FakeNews:
+    """Deterministic NewsClient stub returning a fixed sentiment."""
+
+    enabled = True
+
+    def __init__(self, sentiment: float) -> None:
+        self._sentiment = sentiment
+
+    async def get_sentiment(self, question: str) -> float:
+        return self._sentiment
+
+    def get_freshness(self, question: str) -> float:
+        return 1.0
+
+
+def _make_market(yes_price: float) -> Market:
+    return Market(
+        id="mkt-guardrail-1",
+        condition_id="cond-guardrail-1",
+        question="Will John Smith win the election?",
+        yes_token_id="yes-tok",
+        no_token_id="no-tok",
+        yes_price=yes_price,
+        no_price=1.0 - yes_price,
+        volume_24h=50_000.0,
+        end_date="2026-07-15T00:00:00Z",  # politics <30 d → regime_min 0.08
+        active=True,
+        category="politics",
+    )
+
+
+def _make_signals() -> ExternalSignals:
+    # Neutral macro environment; irrelevant for politics (gated) but explicit.
+    return ExternalSignals(
+        btc_price=100_000.0,
+        btc_change_24h=0.0,
+        eth_price=4_000.0,
+        eth_change_24h=0.0,
+        btc_dominance=50.0,
+        fear_greed_index=50,
+        fear_greed_label="neutral",
+        total_market_cap_change=0.0,
+        valid=True,
+    )
+
+
+def _evaluate(yes_price: float, sentiment: float, caplog) -> tuple[
+    BayesianStrategy, tuple[float, float, str, str]
+]:
+    """Run evaluate() on a politics market and parse the NEWS_MATERIAL line."""
+    strategy = BayesianStrategy(news=FakeNews(sentiment), manifold=None, db=None)
+    market = _make_market(yes_price)
+    with caplog.at_level(logging.INFO, logger="bot.strategy.bayesian"):
+        asyncio.run(strategy.evaluate(market, _make_signals(), occupied_families=set()))
+    for record in caplog.records:
+        m = NEWS_MATERIAL_RE.search(record.getMessage())
+        if m:
+            return strategy, (
+                float(m.group(1)), float(m.group(2)), m.group(3), m.group(4)
+            )
+    pytest.fail(
+        "No NEWS_MATERIAL log line found; got: "
+        f"{[r.getMessage() for r in caplog.records]}"
+    )
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Test 1 — extreme uncorroborated shift: clamp to prior - MAX_NEWS_ONLY_PROB_SHIFT
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_extreme_news_only_shift_is_clamped(caplog):
+    """prior=0.845, raw 0.431 (NVIDIA signature) → final clamped to 0.595."""
+    strategy, (raw, final, guardrail, _) = _evaluate(
+        yes_price=0.845, sentiment=_sentiment_for(0.845, 0.431), caplog=caplog
+    )
+    assert raw == pytest.approx(0.431, abs=1e-3)
+    assert guardrail == "applied"
+    assert final >= 0.595
+    assert final == pytest.approx(0.845 - bayesian.MAX_NEWS_ONLY_PROB_SHIFT, abs=1e-3)
+    assert strategy.get_cycle_stats()["news_guardrail_applied"] == 1
+    assert strategy.get_cycle_stats()["news_with_material"] == 1
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Test 2 — moderate shift inside the band: passes through untouched
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_moderate_news_shift_inside_band_not_clamped(caplog):
+    """prior=0.50, raw 0.62 → within ±0.25 band → final=0.62, no clamp."""
+    strategy, (raw, final, guardrail, _) = _evaluate(
+        yes_price=0.50, sentiment=_sentiment_for(0.50, 0.62), caplog=caplog
+    )
+    assert raw == pytest.approx(0.62, abs=1e-3)
+    assert final == pytest.approx(0.62, abs=1e-3)
+    assert guardrail == "none"
+    assert strategy.get_cycle_stats()["news_guardrail_applied"] == 0
+    # Still counted as a material-news market for the NEWS SUMMARY.
+    assert strategy.get_cycle_stats()["news_with_material"] == 1
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Test 3 — corroboration: any other material signal disables the fuse
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_corroborated_news_not_clamped():
+    """GNews material + another signal >= threshold → raw passes without clamp."""
+    news_lo = _logodds(0.20) - _logodds(0.50)  # ≈ -1.386, clearly material
+    final, applied = apply_news_guardrail(
+        prior=0.50,
+        raw_final_prob=0.20,
+        feat_news_lo=news_lo,
+        other_feats_lo=(0.0, 0.15, 0.0, 0.0),  # one corroborating signal
+    )
+    assert final == 0.20
+    assert applied is False
+
+
+def test_corroboration_threshold_is_inclusive():
+    """|other| == threshold exactly counts as corroboration (>=, not >)."""
+    final, applied = apply_news_guardrail(
+        prior=0.50,
+        raw_final_prob=0.20,
+        feat_news_lo=-1.386,
+        other_feats_lo=(bayesian.NEWS_MATERIAL_LOGODDS_THRESHOLD, 0.0, 0.0, 0.0),
+    )
+    assert final == 0.20
+    assert applied is False
+
+
+def test_uncorroborated_helper_clamps():
+    """Same shift with only noise elsewhere → clamped to prior - 0.25."""
+    final, applied = apply_news_guardrail(
+        prior=0.50,
+        raw_final_prob=0.20,
+        feat_news_lo=-1.386,
+        other_feats_lo=(0.05, -0.09, 0.0, 0.0),  # all below threshold → noise
+    )
+    assert final == pytest.approx(0.25)
+    assert applied is True
+
+
+def test_sub_material_news_never_clamped():
+    """|news_lo| below threshold → fuse not armed, whatever the shift."""
+    final, applied = apply_news_guardrail(
+        prior=0.50,
+        raw_final_prob=0.10,
+        feat_news_lo=0.09,
+        other_feats_lo=(0.0, 0.0, 0.0, 0.0),
+    )
+    assert final == 0.10
+    assert applied is False
+
+
+def test_guardrail_disabled_passthrough(monkeypatch):
+    monkeypatch.setattr(bayesian, "NEWS_GUARDRAIL_ENABLED", False)
+    final, applied = apply_news_guardrail(
+        prior=0.845,
+        raw_final_prob=0.431,
+        feat_news_lo=-1.974,
+        other_feats_lo=(0.0, 0.0, 0.0, 0.0),
+    )
+    assert final == 0.431
+    assert applied is False
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Test 4 — changed_decision: the clamp moves the edge from tradeable to not
+# ─────────────────────────────────────────────────────────────────────────────
+
+def test_guardrail_changed_trade_decision(monkeypatch, caplog):
+    """
+    With max_shift=0.10 the clamped edge (0.10 gross, 0.06 net) falls below the
+    politics <30 d regime gate (0.08) while the raw edge (0.414 gross, 0.374
+    net) crossed it → the fuse prevented the trade → changed_decision=true.
+
+    (With the default 0.25 the clamped edge_net is 0.21, above every regime
+    minimum, so the flag can only fire with a tighter configured band.)
+    """
+    monkeypatch.setattr(bayesian, "MAX_NEWS_ONLY_PROB_SHIFT", 0.10)
+    strategy, (raw, final, guardrail, changed) = _evaluate(
+        yes_price=0.845, sentiment=_sentiment_for(0.845, 0.431), caplog=caplog
+    )
+    assert raw == pytest.approx(0.431, abs=1e-3)
+    assert final == pytest.approx(0.745, abs=1e-3)
+    assert guardrail == "applied"
+    assert changed == "true"
+    stats = strategy.get_cycle_stats()
+    assert stats["news_changed_decisions"] == 1
+    assert stats["news_guardrail_applied"] == 1
+
+
+def test_default_band_does_not_change_decision(caplog):
+    """Default 0.25 band: clamp binds but edge_net 0.21 still crosses the gate."""
+    _, (_, _, guardrail, changed) = _evaluate(
+        yes_price=0.845, sentiment=_sentiment_for(0.845, 0.431), caplog=caplog
+    )
+    assert guardrail == "applied"
+    assert changed == "false"