File size: 12,808 Bytes
f036605 e58251b f036605 | 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 | """Independently re-derive every numeric claim in the generated datasets.
This deliberately does NOT import the generators' arithmetic. It parses the
*question text* for its inputs, recomputes the answer from scratch (or, for the
market set, from the source XBRL facts on disk), and compares against the number
the stored response actually states. A shared helper would make a wrong formula
agree with itself; re-deriving from the question is what makes disagreement
detectable.
Exit code is non-zero if any claim disagrees, so it can gate a release.
"""
from __future__ import annotations
import json
import re
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
HR = ROOT / "datasets" / "hr"
MKT = ROOT / "datasets" / "market"
SEC = ROOT / "data" / "sec" / "companies.json"
def close(stated: float, expected: float, decimals: int) -> bool:
"""Agree to within half a unit of the last digit the response displays.
Precision has to come from the field's own format string, not one global
tolerance: a rate rendered "62%" carries 62.5 underneath, while a ratio
rendered "1.64" is pinned to two decimals. A single tolerance is necessarily
either too loose for the ratios or too tight for the percentages — the first
run of this script reported 61 mismatches, every one of them that mistake.
"""
return abs(stated - expected) <= 0.5 * 10 ** (-decimals) + 1e-9
def headline_number(text: str, unit: str) -> float | None:
"""Pull the first bolded figure of the given unit out of a response."""
if unit == "pct":
m = re.search(r"\*\*([+-]?[\d,]+\.?\d*)%", text)
elif unit == "usd":
m = re.search(r"\*\*\$([\d,]+\.?\d*)", text)
else:
m = re.search(r"\*\*([\d,]+\.?\d*)\*\*", text)
return float(m.group(1).replace(",", "")) if m else None
# --------------------------------------------------------------------- HR ---
def verify_hr() -> tuple[int, int, list[str]]:
rows = [json.loads(l) for l in (HR / "hr_train.jsonl").read_text().splitlines()]
rows += [json.loads(l) for l in (HR / "hr_eval.jsonl").read_text().splitlines()]
checked, bad = 0, []
for r in rows:
if r["task_family"] != "hr_analytics":
continue
q, a = r["instruction"], r["response"]
# Must start on a digit: [\d,]+ also matches a bare comma in the prose.
nums = [
float(x.replace(",", ""))
for x in re.findall(r"\d[\d,]*(?:\.\d+)?", q.replace("$", ""))
]
if "turnover rate" in q:
start, end, leavers = nums[0], nums[1], nums[2]
expected = leavers / ((start + end) / 2) * 100
got, decimals = headline_number(a, "pct"), 1
elif "cost per hire" in q:
external, internal, hires = nums[0], nums[1], nums[2]
expected = (internal + external) / hires
got, decimals = headline_number(a, "usd"), 0
elif "time to fill" in q:
days = sorted(nums[1:]) # nums[0] is the count of vacancies
mid = len(days) // 2
expected = days[mid] if len(days) % 2 else (days[mid - 1] + days[mid]) / 2
m = re.search(r"\*\*Median ([\d.]+) days", a)
got, decimals = (float(m.group(1)) if m else None), 1
elif "compa-ratio" in q:
salary, midpoint = nums[0], nums[1]
expected = salary / midpoint
m = re.search(r"\*\*Compa-ratio ([\d.]+)\*\*", a)
got, decimals = (float(m.group(1)) if m else None), 2
elif "budget for" in q:
current, attrition, growth = nums[0], nums[1], nums[2]
expected = current * attrition / 100 + current * growth / 100
m = re.search(r"\*\*About ([\d,]+) hires", a)
got, decimals = (float(m.group(1).replace(",", "")) if m else None), 0
elif "offers" in q:
offers, accepted = nums[0], nums[1]
expected = accepted / offers * 100
got, decimals = headline_number(a, "pct"), 0
else:
bad.append(f"{r['id']}: unrecognised analytics question")
continue
checked += 1
if got is None:
bad.append(f"{r['id']}: no headline figure parsed")
elif not close(got, expected, decimals):
bad.append(f"{r['id']}: stated {got}, recomputed {expected:.3f} — {q[:70]}")
return checked, len(rows), bad
# ----------------------------------------------------------------- market ---
def verify_market() -> tuple[int, int, list[str]]:
rows = [json.loads(l) for l in (MKT / "market_train.jsonl").read_text().splitlines()]
rows += [json.loads(l) for l in (MKT / "market_eval.jsonl").read_text().splitlines()]
companies = {int(k): v for k, v in json.loads(SEC.read_text()).items()}
checked, bad = 0, []
def fact(cik: int, field: str, year: int):
return companies[cik]["facts"].get(field, {}).get(str(year))
for r in rows:
fam, q, a, cik = r["task_family"], r["instruction"], r["response"], r["cik"]
years = [int(y) for y in re.findall(r"(?:FY|\b)(20\d\d)\b", q)]
if fam == "yoy_growth" and len(years) >= 1:
field = "revenue" if "revenue" in q else "net_income"
y1 = max(years)
v0, v1 = fact(cik, field, y1 - 1), fact(cik, field, y1)
if v0 is None or v1 is None or v0 == 0 or (v0 < 0) != (v1 < 0):
continue # sign-flip and missing rows state no percentage by design
expected = (v1 - v0) / abs(v0) * 100
if "Essentially flat" in a:
# These deliberately omit a bolded percentage and print exact dollars.
m = re.search(r"Exact figures: \$([\d,]+) to \$([\d,]+)", a)
checked += 1
if not m or float(m.group(1).replace(",", "")) != v0 or float(
m.group(2).replace(",", "")
) != v1:
bad.append(f"{r['id']}: flat-case exact figures disagree with source")
continue
got = headline_number(a, "pct")
checked += 1
if got is None or not close(got, expected, 1):
bad.append(f"{r['id']}: stated {got}, recomputed {expected:.2f}% ({field} FY{y1})")
elif fam == "margin_analysis" and years:
kind = "gross" if "gross" in q else "operating" if "operating" in q else "net"
field = {"net": "net_income", "operating": "operating_income", "gross": "gross_profit"}[kind]
y = years[0]
num, rev = fact(cik, field, y), fact(cik, "revenue", y)
if num is None or not rev:
continue
expected = num / rev * 100
got = headline_number(a, "pct")
checked += 1
if got is None or not close(got, expected, 1):
bad.append(f"{r['id']}: stated {got}, recomputed {expected:.2f}% ({kind} margin FY{y})")
elif fam == "cagr" and len(years) >= 2:
y0, y1 = min(years), max(years)
v0, v1 = fact(cik, "revenue", y0), fact(cik, "revenue", y1)
if not v0 or not v1 or v0 <= 0 or v1 <= 0:
continue
expected = ((v1 / v0) ** (1 / (y1 - y0)) - 1) * 100
got = headline_number(a, "pct")
checked += 1
if got is None or not close(got, expected, 1):
bad.append(f"{r['id']}: stated {got}, recomputed {expected:.2f}% CAGR")
elif fam == "ratio_analysis" and years:
y = years[0]
if "current ratio" in q or "short-term obligations" in q:
ca, cl = fact(cik, "current_assets", y), fact(cik, "current_liabilities", y)
if not ca or not cl:
continue
expected, got, decimals = ca / cl, headline_number(a, "plain"), 2
elif "debt-to-equity" in q or "leveraged" in q:
li, eq = fact(cik, "liabilities", y), fact(cik, "equity", y)
if not li or not eq or eq <= 0:
continue
expected, got, decimals = li / eq, headline_number(a, "plain"), 2
else:
ni, eq = fact(cik, "net_income", y), fact(cik, "equity", y)
if ni is None or not eq or eq <= 0:
continue
expected, got, decimals = ni / eq * 100, headline_number(a, "pct"), 1
checked += 1
if got is None or not close(got, expected, decimals):
bad.append(f"{r['id']}: stated {got}, recomputed {expected:.3f} ({fam} FY{y})")
elif fam == "news_extraction":
m = re.search(r"```json\n(.*?)\n```", a, re.S)
if not m:
bad.append(f"{r['id']}: no JSON block")
continue
rec = json.loads(m.group(1))
y = rec["fiscal_year"]
rev, ni = fact(cik, "revenue", y), fact(cik, "net_income", y)
checked += 1
if rec["revenue_usd"] != int(rev) or rec["net_income_usd"] != int(ni):
bad.append(f"{r['id']}: JSON dollars disagree with filed facts FY{y}")
elif not close(rec["net_margin_pct"], ni / rev * 100, 1):
bad.append(f"{r['id']}: JSON net margin disagrees with recomputation")
elif fam == "news_fact_check":
my = re.search(r"\*\*The FY(20\d\d) revenue figure is wrong", a)
mv = re.search(r"Filed figure: \$([\d,]+)", a)
checked += 1
if not my or not mv:
bad.append(f"{r['id']}: fact-check response missing year or filed figure")
continue
rev = fact(cik, "revenue", int(my.group(1)))
if rev is None or float(mv.group(1).replace(",", "")) != float(int(rev)):
bad.append(f"{r['id']}: corrected figure disagrees with filed revenue")
elif fam == "news_summary" and years:
y1 = max(years)
v0, v1 = fact(cik, "revenue", y1 - 1), fact(cik, "revenue", y1)
if v0 is None or v1 is None or v0 == 0 or (v0 < 0) != (v1 < 0):
continue # sign-flip bullets state no percentage by design
got = headline_number(a, "pct")
if got is None:
continue # near-flat wording carries no bolded pct
checked += 1
if not close(got, (v1 - v0) / abs(v0) * 100, 1):
bad.append(f"{r['id']}: summary growth pct disagrees (FY{y1})")
elif fam == "headline_sentiment" and years:
y1 = max(years)
rev0, rev1 = fact(cik, "revenue", y1 - 1), fact(cik, "revenue", y1)
ni0, ni1 = fact(cik, "net_income", y1 - 1), fact(cik, "net_income", y1)
if None in (rev0, rev1, ni0, ni1):
continue
g = None if rev0 == 0 or (rev0 < 0) != (rev1 < 0) else (rev1 - rev0) / abs(rev0) * 100
rev_up = (g or 0) > 0.1 or (g is None and rev1 > rev0)
ni_up = ni1 > ni0
expected = (
"positive" if rev_up and ni_up and ni1 > 0
else "negative" if not rev_up and not ni_up
else "mixed"
)
m = re.search(r"\*\*(positive|negative|mixed)\*\*", a)
checked += 1
if not m or m.group(1) != expected:
bad.append(f"{r['id']}: tone label disagrees with recomputed rule (FY{y1})")
elif fam == "news_commentary" and years:
y1 = max(years)
v0, v1 = fact(cik, "revenue", y1 - 1), fact(cik, "revenue", y1)
m = re.search(r"([+-][\d.]+)% on the year", a)
if m is None or v0 is None or v1 is None or v0 == 0 or (v0 < 0) != (v1 < 0):
continue
checked += 1
if not close(float(m.group(1)), (v1 - v0) / abs(v0) * 100, 1):
bad.append(f"{r['id']}: commentary growth pct disagrees (FY{y1})")
return checked, len(rows), bad
def main() -> int:
total_bad = []
print("Independent re-derivation of stated numeric claims\n")
for name, fn in (("HR", verify_hr), ("Market", verify_market)):
checked, rows, bad = fn()
status = "OK" if not bad else f"{len(bad)} MISMATCH"
print(f"{name:8s} {checked:5d} numeric claims re-derived across {rows} rows — {status}")
for line in bad[:10]:
print(f" {line}")
if len(bad) > 10:
print(f" ... and {len(bad) - 10} more")
total_bad += bad
print()
if total_bad:
print(f"FAIL: {len(total_bad)} claims disagree with independent recomputation")
return 1
print("PASS: every re-derived claim matches the stated figure")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|