OLS results - Sun
            ----Number of Digits-----
 Model      Beta  SE(B)  FStat   R**2  Pass=1/Fail=0
 Norris11   12.2   14.2   14.1   15.0   1.
 Pontius    11.9   12.7   12.4   15.0   1.
 NoInt1     14.7   14.8    0.0*   0.0** 1.
 NoInt2     15.0   14.9    0.0*   0.0** 1.
 Filippelli  0.0S   0.0S   0.4    2.7   0.
 Longley    11.7   12.4   12.4   14.9   1.
 Wampler1    9.2   15.0    0.0I  15.0   1.
 Wampler2   12.5   15.0    0.0I  15.0   1.
 Wampler3    9.0   13.5   13.4   15.0   1.
 Wampler4    8.9   13.8   15.0   15.0   1.
 Wampler5    7.3   13.8   13.1   14.2   1.
 OLS summary:  10./11 passed.

Notes:

* In the No Intercept regressions, TSP does not report an F-statistic,
although one could easily be defined as @COEF'@VCOV"@COEF or @T**2
in these single-regressor models.

** In the No Intercept regressions, TSP defines R**2 differently from
NIST.  TSP uses the squared correlation between actual and fitted values.
NIST apparently uses the formula  1 - @SSR/(y'y)  .  This formula is
not invariant to the mean of y.
   For example, if there is a single regressor which is nearly
constant (but not quite constant), and y has a large mean,
this formula will yield a large R**2 value, but
a regression on a plain constant will yield R**2 = 0 (with the
standard formula), even though the two equations have nearly the
same SSR.
   Controversy in defining R**2 in the no intercept case is nothing
new, and it's curious why NIST might be taking a particular stance
on this cloudy issue.

S  In Filippelli, TSP detected singularity, and set the last 2
coefficients and standard errors to zero.  It still obtained an
approximate fit, judging from the digits in Fstat and R**2.
Judging also from the results of other packages, this appears to
be a difficult problem with double precision data.

I  In Wampler1 and Wampler2, there is a perfect fit, so the F-statistic
is infinite.  Since infinity can't be represented on a computer,
it's not clear how to define a number of significant digits for
this case.