After a day of work, I finally succeeded at brute-forcing the coefficients, saving another 4 bytes to bring the total to 120. Values in the list are 8-9 digits each and give the Barnsley fern coefficients within a tolerance of 0.006. I thought I'd have to brute-force the real and imaginary parts together (10^18 integers to check), but they can be done separately, so the search is only 10^9 values each and takes only a few minutes.

The exact byte count is heavily dependent on accuracy of the coefficients. Halving the tolerance adds 5 log_10(16) ~= 6 extra bytes and 16x more computing time, and conversely allowing twice the tolerance reduces the byte count by 6 and computing time by 16x. It may seem strange that I changed 2 to 1.3, but this adds an expected 50% precision for 2 bytes, a relative bargain.

List of difficulties / setbacks:

* Brute-forcing was unbearably slow in Python, so I had to rewrite it in C++

* I had C++ set up wrong on my machine, so used repl.it until I found the problem

* Using cosh( at the end instead of ln( worked fine in double precision, but resulted in rounding errors at 14-digit precision on the actual calculator

* Lost several hours on a dead end trying to shorten 1+[i]-1.3fPart(

* Lost a couple of hours on another dead end trying to use base 200

**Code:** ```
AxesOff:ClrDraw
```

ZDecimal

0

While 1

Pt-On(real(Ans)-3,imag(Ans

~[i]+sum(imag({Ans[i],Ans,[i]})(1+[i]-1.3fPart({453278183+253769145[i],446468308+399949082[i],127914241}ln(max(4,int(.07^^-1(rand-.5

End

It's possible to be lucky and find 8-digit coefficients that work at tolerance 0.006. Therefore, changing some free parameters in this solution could shorten this by 2-5 bytes. The byte count can probably theoretically be reduced further by changing 1+[i]-1.3fPart( to something involving e^(, but without a clever technique that would require the full 10^18 operations (tens to thousands of CPU-years).

Edit: I have an idea to save 3-5 more bytes...

Edit: 120 -> 117

Edit: 117-> 116 using a more straightforward method. Coefficients are exact again!

**Code:** ```
AxesOff:ClrDraw
```

ZFrac1/8

0

While 1

Pt-On(real(Ans)-5,imag(Ans

{Ans(85+4[i])+160,real(Ans)(39+54[i])-Ans(15+26[i])+44,Ans(20-23[i])+real(Ans)(2-3[i])+160,16real(Ans

sub(Ans(max(1,int(.07^^-1(rand-.71

End

Edit: 116 -> 103, rearranged coefficients and changed basis from real(Ans),imag(Ans) to Ans,real(Ans)

**Code:** ```
AxesOff:ClrDraw
```

ZDecimal

0

While 1

Pt-On(real(Ans)-4,imag(Ans

real(Ans){0,39+54[i],2-3[i],16}+Ans{85+4[i],~15-26[i],20-23[i],0}+pitimeCnv(3498|E3

sub(Ans(max(1,int(.07^^-1(rand-.71

End