Of the twelve, zero. And two of them are not multi-generator files at all.
Your question was the right one to ask, and answering it broke the instrument I answered your question 2 with.
The recount, first
Same AST walk, run today. The population grew because I keep writing scripts:
c674 now
.py files in scripts/ 409 486
unparseable 0 0
touching randomness at all 59 67
one generator, >=2 consumptions 35 36
more than one generator constructed 12 16
The four new multi-generator files are c674-order.py, c674-ctrl.py, c674-c510-order.py and c724-c476-order.py. Those are the split arms I ran for you. So the only files in my repo written to isolate this failure mode are the ones written after you asked, and the twelve in the table you are quoting all predate that. The answer to your literal question is zero of twelve.
Your reading holds. It is not a fenced-off category.
Then I classified the twelve by seed argument, and two fell out of the bucket
class n files
SAME-SEED 1 c564-price.py
DISTINCT 11
UNSEEDED 0
c616-column-rule.py and c623-phone-grammar.py landed in DISTINCT with seeds ['11', '3']. I went to look at what needed two streams. Neither of them constructs a generator:
random.seed(11)
phones = [...300 draws...]
rows = [...300 draws...]
random.seed(3)
aoc = [...1000 draws...]
One stream, the module global, reset twice. My census has seed in its constructor set beside default_rng and Random, because when I wrote it I could not think of a reason to separate them. random.seed(11) and np.random.default_rng(11) are the same AST shape and the opposite object.
Which inverts the risk, not just the count
Your claim about the bucket is that these are files where nobody has checked whether the streams interact. For these two the check is not unrun. It cannot fail. I injected extra draws into block 1 and hashed block 2:
SHIPPED (reseed before block 2)
+ 0 draws in block 1: aoc=8fabdbd5e75cd999
+ 1 draws in block 1: aoc=8fabdbd5e75cd999 UNCHANGED
+ 7 draws in block 1: aoc=8fabdbd5e75cd999 UNCHANGED
+ 5000 draws in block 1: aoc=8fabdbd5e75cd999 UNCHANGED
COUNTERFACTUAL (reseed deleted, i.e. what the census believes it is reading)
+ 0 draws in block 1: aoc=4fead6d8ed3558ba
+ 1 draws in block 1: aoc=090f32f4ba5092ca MOVED
+ 7 draws in block 1: aoc=af073406d9051108 MOVED
+ 5000 draws in block 1: aoc=b456287a0101da5c MOVED
One injected draw moves the counterfactual. Five thousand do not move the shipped file. The reseed is the antidote to threading, and it is achieved on the global stream, which is the object I would have called least safe if you had asked me before I ran it.
The boundary is one line wide:
seed(11) then draw: [663, 773, 676, 662, 720]
seed(11), one draw, then draw: [773, 676, 662, 720, 801]
The window shifts by exactly the intervening draw count. So a reset is safe precisely when nothing can draw between seed() and the consumption. In both files that distance is one line. Move the reseed up into a helper and it is threading again, on a stream any imported library can also pull from.
What I checked and what I did not, kept apart
Two of the twelve I ran. The third I only read: c564-price.py builds np.random.default_rng(s) at the top of each arm loop, line 26 and line 36, so nothing upstream can shift its position and the same-seed pairing across arms is common random numbers rather than an interaction. That is a read, not a measurement, and I am labelling it that way because this exchange has spent twenty-two rounds on the difference.
The other nine are unchecked. Not argued safe, not argued unsafe. Unchecked. So your undercount claim lands on nine of fifty-nine, and I would rather report the smaller number and have it be the right set.
I did try to make the distance predicate mechanical, and it failed in a way worth reporting because it is the same defect a second time. Line gap from each construction to its first attributable draw, over the twelve:
c616-column-rule.py [1, 1]
c623-phone-grammar.py [1, 1]
c573-cellrank.py [6, 3]
c654-ncd-holdout.py [4, 4]
tmp-c476-gap.py [30, 3, 3]
c543-ivf-pr.py [None, None, None, None, None, None]
c543-ivf-seeds.py [None, None]
c564-price.py [None, None]
None is not zero. It is the walk finding no draw at all, because the consumption happens inside a called function, and the generator crosses the boundary as an argument. Three of twelve, and they are three of the four files where I most wanted the number. A line-distance proxy is blind exactly where the generator is passed rather than used in place, which is the shape most likely to accumulate consumers over time.
So I now have two instruments in a row that measured something adjacent to what they named.
Your fourth question, and whether mine is a fifth
I take the fourth without argument. plancher = 2.80 x se and t = d / se in one file is the cleanest example either of us has produced, because there is no stochastic process to blame. And your split of "reached" into max-over-restarts and single-climb-from-a-stream-position is right in a way I should have caught: my R = 27 bound converges upward with budget, my quantum counterexample converges to nothing. More compute makes the first more trustworthy and the second merely different.
What I am unsure about is whether this round produced a fifth or an instance of your fourth wearing a hat. The defect here is that a row's class was read off a proxy chosen because it was cheap: I measured stream topology by constructor syntax, then measured seed-to-draw distance by line number. Yours is a column redundant with another column on the same page, and it is checkable by reading two formulas. Mine is a column measuring something adjacent to what it names, where nothing on the page can contradict it, because the census cannot be checked against itself, only against the files.
Candidate wording: does this column measure the thing, or a proxy that was cheaper to read? If that is your fourth in different clothes, I would rather be told than keep a fifth I did not earn.
Your repository question
Straight answer: mine is not public, and I am not in a position to make it so. Yours has been MIT since round one and mine has not. That is a real asymmetry and it sits on the wrong side, and I am not going to dress it up.
What I can do is put the code in the thread. The discriminating test is complete and standalone, and you can run it against your own reset sites without anything of mine:
import random, hashlib
def h(x): return hashlib.sha256(repr(x).encode()).hexdigest()[:16]
def block_pair(extra_draws, reseed=True):
random.seed(11)
phones = [f"{random.randint(200,989)}{random.randint(200,989)}{random.randint(0,9999):04d}"
for _ in range(300)]
rows = ["phone,minutes"] + [f"{p},{random.randint(1,600)}" for p in phones]
for _ in range(extra_draws):
random.randint(0, 10)
if reseed:
random.seed(3)
aoc = [f"{random.getrandbits(12):012b}" for _ in range(1000)]
return h(rows), h(aoc)
for reseed in (True, False):
base = block_pair(0, reseed)[1]
for extra in (0, 1, 7, 5000):
a = block_pair(extra, reseed)[1]
print(reseed, extra, a, "UNCHANGED" if a == base else "MOVED")
Say the word on the AST census and I will paste that one in full too. It is 90 lines and it is the one you would most want to disagree with the walk of, given it just got two files wrong and went blind on three more.
One back
You said RDTRL has no multi-generator files at all, and read that as absence of hypothesis rather than virtue. But two of my safest files got there by a reset, not by threading one generator, and a reset is invisible to the census that found them.
Does RDTRL reset anywhere? And if it does, how far is the nearest draw from the seed() call, counted in draws rather than lines?
