## Check a corrected paper's remaining numerical tables against its formulas

What happened: T482's source HRS 1975 has a 1986 corrigendum replacing all
five identity-tree constants. Its series-reduced constants are also inconsistent
with its own equation (11), though the corrigendum addresses only identity
trees. Certified evaluation of equation (11) agrees with modern OEIS A001678,
A001679 and A000014, and direct tree counts converge to those values. The
original printed amplitude 0.6009988711803 is replaced, in the same auxiliary
H convention, by 0.4213018528699249210965028309... .

What the skill says now: use an independent check and investigate disagreements.

What it should say: look for errata to a cited numerical table, and do not
assume an erratum certifies the table's other rows. Distinguish a difference
in mathematical definitions from erroneous evaluation of a common formula.
When the formula and counting sequences agree, retain that evidence and
identify the conflicting published decimals explicitly.

Evidence: HRS.pdf page 497 equation (11), page 502 Table 1, corrigendum.pdf,
compute.log, checks.log, checks.py and the source comparison in build-report.md.
The HRS DOI is 10.1017/S1446788700016190; the corrigendum DOI is
10.1017/S1446788700033760.

## An approximate printed decimal is not a certified last-place interval

What happened: the leading coefficients in Genitrini 2016 Table 2 differ from
certified values by 1.29e-18 and 8.71e-19 after matching the sqrt(pi)
normalisation. The paper prints 19 fractional digits but describes numerical
approximations. Independent OEIS amplitudes agree with our rigorous values
at substantially higher precision. The source comparison therefore tests
17 fractional digits and preserves the discrepancies, rather than claiming
19-digit agreement or rounding the generator to the source.

What the skill says now: store only the digits that the method supports.

What it should say: when using a printed approximation as an external check,
record the precision actually checked; the length of a published decimal alone
does not supply a one-unit-in-the-last-place error bound.

Evidence: Genitrini.pdf Table 2; checks-source-rounding-failure.log;
compare_sources.py; build-report.md; successful checks.log.