Spaces:
Running
Running
File size: 12,614 Bytes
69e310f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 | """The verification node — deterministic, never an LLM.
Every numeric claim in the brief is recomputed from the raw price bars in state
by :mod:`app.graph.recompute`, every quoted headline is matched verbatim against
the news actually retrieved, and the brief's structure is checked against the
run. Numbers match to the cent or nothing ships.
Outcomes:
* all checks pass → the gate
* any mismatch, first time → exactly one writer regeneration
* any mismatch, second time → ``HUMAN_REVIEW`` (flagged, never auto-delivered)
"""
from __future__ import annotations
import re
from typing import Any, Literal
from pydantic import BaseModel, ConfigDict, Field
from app.core.events import EventKind
from app.graph.context import current_context
from app.graph.recompute import recompute
from app.graph.state import RunState, is_partial, usable_tickers
from app.models.brief import Brief
from app.models.market import METRIC_UNITS, MetricUnit
from app.models.run import RunError, RunStatus
#: Absolute tolerance per unit. USD is "to the cent"; percentages are checked far
#: tighter than the 4-decimal rounding the tool applies on the way out.
TOLERANCES: dict[MetricUnit, float] = {
"usd": 0.005,
"percent": 0.001,
"ratio": 0.001,
"score": 0.000001,
}
_WHITESPACE = re.compile(r"\s+")
ClaimStatus = Literal["match", "mismatch", "unverifiable"]
def _normalise_quote(text: str) -> str:
return _WHITESPACE.sub(" ", (text or "").strip()).casefold()
class ClaimCheck(BaseModel):
"""One recomputed numeric claim."""
model_config = ConfigDict(extra="forbid")
claim_id: str
ticker: str
metric: str
unit: MetricUnit
claimed: float
recomputed: float | None
delta: float | None
tolerance: float
status: ClaimStatus
detail: str
class QuoteCheck(BaseModel):
"""One verbatim third-party quotation."""
model_config = ConfigDict(extra="forbid")
ticker: str
field: str
text: str
status: Literal["match", "mismatch"]
detail: str
class VerificationReport(BaseModel):
"""The full verification pass, rendered claim-by-claim in the UI."""
model_config = ConfigDict(extra="forbid")
ok: bool
checked_claims: int = 0
matched: int = 0
mismatched: int = 0
unverifiable: int = 0
coverage: float = 0.0
claim_checks: list[ClaimCheck] = Field(default_factory=list)
quote_checks: list[QuoteCheck] = Field(default_factory=list)
structural_issues: list[str] = Field(default_factory=list)
structural_warnings: list[str] = Field(default_factory=list)
@property
def failures(self) -> list[str]:
out = [
f"{c.claim_id} ({c.ticker}.{c.metric}): {c.detail}"
for c in self.claim_checks
if c.status != "match"
]
out += [
f"{q.ticker}.{q.field}: {q.detail}" for q in self.quote_checks if q.status != "match"
]
out += self.structural_issues
return out
def verify_brief(brief: Brief, state: RunState) -> VerificationReport:
"""Pure verification: recompute every claim from raw state.
Importable and testable in isolation — no graph, no network, no model.
"""
prices = state.get("prices", {})
fundamentals = state.get("fundamentals", {})
sentiment = state.get("sentiment", {})
news = state.get("news", {})
watchlist = set(state.get("tickers", []))
structural: list[str] = []
warnings: list[str] = []
# ---------------------------------------------------------- structure ---
expected_date = state.get("session_date")
if expected_date and brief.generated_for != expected_date:
structural.append(
f"brief.generated_for '{brief.generated_for}' does not match the run's "
f"session date '{expected_date}'"
)
if set(brief.watchlist) != watchlist:
structural.append(
f"brief.watchlist {sorted(brief.watchlist)} does not match the run watchlist "
f"{sorted(watchlist)}"
)
defined = brief.claims_by_id()
if len(defined) != len(brief.claims):
structural.append("duplicate claim_id values in brief.claims")
referenced = brief.referenced_claim_ids()
missing_refs = sorted(referenced - defined.keys())
if missing_refs:
structural.append(f"narrative references undefined claims: {missing_refs}")
orphans = sorted(defined.keys() - referenced)
if orphans:
warnings.append(f"claims defined but never cited: {orphans}")
expected_rows = set(usable_tickers(state))
row_tickers = [row.ticker for row in brief.snapshot]
if len(row_tickers) != len(set(row_tickers)):
structural.append("duplicate tickers in the snapshot table")
unexpected = sorted(set(row_tickers) - watchlist)
if unexpected:
structural.append(f"snapshot contains tickers outside the watchlist: {unexpected}")
dropped = sorted(expected_rows - set(row_tickers))
if dropped:
structural.append(f"snapshot omits tickers that have market data: {dropped}")
for row in brief.snapshot:
for claim_id in row.claim_ids():
claim = defined.get(claim_id)
if claim and claim.ticker != row.ticker:
structural.append(
f"snapshot row {row.ticker} cites {claim_id}, which belongs to {claim.ticker}"
)
if brief.partial != is_partial(state):
structural.append(
f"brief.partial={brief.partial} but the run "
f"{'has' if is_partial(state) else 'has no'} incomplete tickers"
)
# ------------------------------------------------------------- claims ---
claim_checks: list[ClaimCheck] = []
for claim in brief.claims:
unit = METRIC_UNITS.get(claim.metric, claim.unit)
tolerance = TOLERANCES[unit]
if claim.ticker not in watchlist:
claim_checks.append(
ClaimCheck(
claim_id=claim.claim_id,
ticker=claim.ticker,
metric=claim.metric,
unit=unit,
claimed=claim.value,
recomputed=None,
delta=None,
tolerance=tolerance,
status="mismatch",
detail=f"ticker '{claim.ticker}' is not in the run watchlist",
)
)
continue
if claim.unit != unit:
structural.append(
f"{claim.claim_id} declares unit '{claim.unit}' but "
f"{claim.metric} is measured in '{unit}'"
)
recomputed = recompute(
claim.metric,
history=prices.get(claim.ticker),
fundamentals=fundamentals.get(claim.ticker),
sentiment=sentiment.get(claim.ticker),
)
if recomputed is None:
claim_checks.append(
ClaimCheck(
claim_id=claim.claim_id,
ticker=claim.ticker,
metric=claim.metric,
unit=unit,
claimed=claim.value,
recomputed=None,
delta=None,
tolerance=tolerance,
status="unverifiable",
detail="raw state cannot support recomputing this figure",
)
)
continue
delta = abs(claim.value - recomputed)
matched = delta <= tolerance
claim_checks.append(
ClaimCheck(
claim_id=claim.claim_id,
ticker=claim.ticker,
metric=claim.metric,
unit=unit,
claimed=claim.value,
recomputed=round(recomputed, 6),
delta=round(delta, 8),
tolerance=tolerance,
status="match" if matched else "mismatch",
detail=(
"recomputed from raw price bars"
if matched
else f"claimed {claim.value} vs recomputed {recomputed:.6f} (Δ {delta:.6f})"
),
)
)
# ------------------------------------------------------------- quotes ---
quote_checks: list[QuoteCheck] = []
for ticker, field_name, text in brief.quoted_texts():
feed = news.get(ticker)
available = {_normalise_quote(item.title) for item in (feed.items if feed else [])}
matched = _normalise_quote(text) in available
quote_checks.append(
QuoteCheck(
ticker=ticker,
field=field_name,
text=text[:200],
status="match" if matched else "mismatch",
detail=(
"verbatim match against retrieved headlines"
if matched
else "quoted text does not appear in the headlines retrieved for this ticker"
),
)
)
matched_count = sum(1 for c in claim_checks if c.status == "match")
mismatched = sum(1 for c in claim_checks if c.status == "mismatch")
unverifiable = sum(1 for c in claim_checks if c.status == "unverifiable")
quote_failures = sum(1 for q in quote_checks if q.status == "mismatch")
coverage = 1.0 if not claim_checks else matched_count / len(claim_checks)
ok = mismatched == 0 and unverifiable == 0 and quote_failures == 0 and not structural
return VerificationReport(
ok=ok,
checked_claims=len(claim_checks),
matched=matched_count,
mismatched=mismatched,
unverifiable=unverifiable,
coverage=round(coverage, 4),
claim_checks=claim_checks,
quote_checks=quote_checks,
structural_issues=structural,
structural_warnings=warnings,
)
async def verify_node(state: RunState) -> dict[str, Any]:
"""Graph node: run verification and stream each claim to the UI."""
ctx = current_context()
brief = state.get("brief")
if brief is None:
await ctx.emit(EventKind.VERIFY_COMPLETED, "No brief to verify", {"ok": False})
return {
"verification": VerificationReport(
ok=False, structural_issues=["writer produced no brief"]
).model_dump(),
"errors": [RunError(stage="verify", message="writer produced no brief")],
"status": RunStatus.HUMAN_REVIEW,
}
await ctx.emit(
EventKind.VERIFY_STARTED,
f"Recomputing {len(brief.claims)} numeric claims from raw state",
{"claims": len(brief.claims)},
)
with ctx.tracer.step(
"verify",
run_id=ctx.run_id,
input_data={"claims": len(brief.claims)},
) as span:
report = verify_brief(brief, state)
span.update(output={"ok": report.ok, "coverage": report.coverage})
# Stream tick-by-tick so the verification screen animates green (or red).
for check in report.claim_checks:
await ctx.emit(
EventKind.VERIFY_CLAIM,
f"{check.ticker} · {check.metric}",
check.model_dump(),
)
for quote in report.quote_checks:
await ctx.emit(
EventKind.VERIFY_CLAIM,
f"{quote.ticker} · quoted headline",
{**quote.model_dump(), "metric": "quoted_headline", "unit": "text"},
)
await ctx.emit(
EventKind.VERIFY_COMPLETED,
(
f"All {report.matched} claims verified"
if report.ok
else f"{report.mismatched + report.unverifiable} claim(s) failed verification"
),
report.model_dump(),
)
update: dict[str, Any] = {"verification": report.model_dump()}
if report.ok:
return update
failures = report.failures
update["errors"] = [
RunError(stage="verify", message=detail, severity="error") for detail in failures[:10]
]
if state.get("regenerations", 0) >= ctx.settings.max_regenerations:
update["status"] = RunStatus.HUMAN_REVIEW
return update
def route_after_verify(state: RunState) -> Literal["writer", "gate"]:
"""Exactly one regeneration on mismatch, then straight to the human gate.
The cap lives in state (`max_regenerations`, default 1 per spec) so a
mismatch can never ping-pong between writer and verifier.
"""
report = state.get("verification") or {}
if report.get("ok"):
return "gate"
if int(state.get("regenerations", 0)) < int(state.get("max_regenerations", 1)):
return "writer"
return "gate"
|