Impact
VerificationEngine.verify_math(expression, expected_value) returns VERIFIED with is_correct: true for wrong claims about small values and for wrong digits beyond the sixth decimal place. Both sides are compared after rounding the computed value to 6 decimal places, using an absolute tolerance of 1e-6. A caller's tolerance=0 does not restore an exact comparison.
| Call | 8.0.0 | Correct |
|---|---|---|
verify_math("1/1000000000", 0) |
VERIFIED, is_correct: true |
wrong (true value 1e-9) |
verify_math("1/1000000000", 0, tolerance=0) |
VERIFIED, is_correct: true |
wrong |
verify_math("9/10000000", 0) |
VERIFIED, is_correct: true |
wrong (true value 9e-7) |
verify_math("1/3", 0.3333329) |
VERIFIED, is_correct: true |
wrong 7th digit |
verify_math("1/3", 0.3333323) |
VERIFIED, is_correct: true |
off by 1.03e-6, more than the stated 1e-6 tolerance |
verify_math("1/1000000000", 0.005, tolerance=0.01) |
VERIFIED, is_correct: true |
the tolerance cap does not scale down for small values |
verify_math("1/3", 0.3333333333333333, tolerance=1e-7) |
DISPROVEN | correct to 1.9e-17; only the computed side is rounded, so a precise claim is rejected |
verify_math("2+2", 5), verify_math("10000000000000001", 10000000000000000) (controls) |
DISPROVEN | correct |
This is the small-magnitude counterpart of GHSA-j622-qffp-rc27, which fixed the same false VERIFIED for values above 15 significant digits.
Wrong rates, probabilities, unit prices and per-unit fees (values below about 1e-3) can be certified as correct. On the natural-language path the wrong result is signed as a VERIFIED verdict; admission is withheld there because the translation is unverified, but the verdict itself is false.
Affected
qwed (PyPI) through 8.0.0. Reproduced on main @ 10f5627 (unchanged since 8.0.0).
Reachable from:
- the
VerificationEngine.verify_mathPython API; POST /verify/natural_language, viaControlPlane.process_natural_language(control_plane.py:282);- consensus math checks, via
ConsensusVerifier(consensus_verifier.py:774).
Not affected:
POST /verify/mathina = bform, which usessymbolic_difference_status;- batch math items (
core/batch.py, exact comparison); - the SDK's local
_math_answers_match.
Reproduction
from qwed_new.core.verifier import VerificationEngine
m = VerificationEngine()
for expr, claim, kw in [("1/1000000000", 0, {}), ("1/1000000000", 0, {"tolerance": 0}),
("1/3", 0.3333329, {}), ("2+2", 5, {})]:
r = m.verify_math(expr, claim, **kw)
print(expr, claim, kw, r["status"], r.get("is_correct"))
Details
src/qwed_new/core/verifier.py:
- Line 719:
_VERIFY_MATH_QUANTUM = Decimal("0.000001")._calc_decimal(lines 730-751) quantises the computed value to 6 decimal places before comparison (line 791); the claim (expected_decimal) is not quantised. - Line 760: the default
tolerance: float = 1e-6is absolute, independent of magnitude. - Line 686:
MIN_VERIFY_MATH_TOLERANCE_CAP = Decimal("0.01"), so a caller tolerance up to 0.01 is accepted even when the value is many orders of magnitude smaller. - Lines 806-807:
diff = abs(calculated_value - expected_decimal), thenis_correct = diff <= tolerance_decimal.
Patches
Fixed in 9.0.0 (d4a0b0b). Upgrade to 9.0.0 or later.
- The comparison no longer rounds. An exact rational value is compared with the claim exactly; any other value is evaluated at a precision of at least 30 digits, more for large values and long claims.
- With no tolerance, a claim matches when it lies within
min(1e-6, max(1e-9 × |value|, 1e-12))of the value, or when it has at least 6 significant digits and is the value correctly rounded to its last digit. So1/3 = 0.333333still matches, and1/3 = 0.3333329does not. tolerance=0is exact. A caller's tolerance stays absolute and is still capped at 1% of the value (at least 0.01), but never more than the value itself.tolerancenow defaults toNone. The result has a newmatchfield saying how a claim matched (exact,caller_tolerance,default_boundorrounded_claim) and the bound used.
On 9.0.0 the first five wrong claims in the table are DISPROVEN, and verify_math("1/1000000000", 0.005, tolerance=0.01) is BLOCKED (Tolerance exceeds deterministic verification bound), since a tolerance may not exceed the value. The precise claim verify_math("1/3", 0.3333333333333333, tolerance=1e-7) is VERIFIED. The controls are unchanged.
Workarounds
Upgrade to 9.0.0. On older versions, use POST /verify/math with the a = b form, or batch math items, which compare exactly. Do not rely on verify_math or the natural-language verdict for values below about 1e-3, or for claims with more than 6 significant decimal places.
Source
Internal correctness audit (October 2026). Not previously disclosed.
