FINITE-SOURCE TRANSFER AUDIT: VALID WITH A STRICT PRECISION BUDGET
25 October2026. Independent certifier proof and exact arithmetic controls.

SETUP
Let X0 be odd2adic Haar, with canonical blocks indexed0,1,... . Write T_i
for type(r_i,s_i), B_i=r_i+s_i for its division cost, X_(i+1) for its endpoint,
and R_i=V_(T_i)(X_(i+1)) for the post-block oldpin reserve. Define
 Q=sum_(i0..m)B_i, m>=1.
The vector Z_m=(T0,...,Tm,R_m) determines J0,...,J_(m-1): work backward using
 R_i=B_i+R_(i+1) ifT_i=T_(i+1),
 R_i=separator(T_i,T_(i+1)) otherwise,
and J_i=B_(i+1)+R_(i+1)-R_i. On null undefined/pin states use a cemetery
symbol. For positive integer sources, canonical blocks are always finite,
but a queried pin reserve can be infinite; use the same cemetery convention.

PRECISION LEMMA / OFF-BY-ONE
The complete tuple is determined by the first H binary input digits whenever
 Q+R_m < H.
The inequality is STRICT. After Q divisionbits the terminal state is known
modulo2^(H-Q). Determining a valuationR_m requires R_m < H-Q; equality does
not determine whether another zero bit follows. Exact final oddness and all
earlier block bits are also certified, becauseR_m>=1 gives sufficient slack.
The success set is a union of whole H-bit odd cylinders. If a successful Haar
point shares its H-bit residue with an integer, that integer has the same
finite tuple. Finite-type cylinder maps and backward reserve reconstruction
prove that no additional future bit is used.

EXACT FAILURE MASS
Each B_i is a sum of two independent Geom(1/2)variables on{1,2,...}. R_m is
another such variable, independent of every type revealed throughm. Therefore
 Q+R_m ~ NB(a), a=2m+3,
where NB(a) is the sum of a positive-support geometric variables, i.e. the
number of fair Bernoulli trials up to the a-th success. Hence
 epsilon_(m,H)=P(Q+R_m>=H)
 =2^-(H-1)*sum_(j0..a-1)binom(H-1,j), forH>a;
 =1, forH<=a.
The eventis exactly the generic tuple's failure-of-certification event, not
an actual dynamical failure. A J-only observer can sometimes certify more
(e.g. same-type J0) without knowing the terminal reserve.

TOTAL VARIATION TRANSFER
Let n be uniform among the first N positive odd integers, and let
K=2^(H-1), N=qK+s,0<=s<K. The exact totalvariation distance of its H-bit
residue law from the Haar odd residue law is
 delta_(N,H)=s(K-s)/(KN) <=K/(4N) <=K/N.
Proof: s residueclasses haveq+1representatives and the othersq; sum their
absolute discrepancies and divideby2.

Couple those two residue laws maximally, then sample a Haar point given its
residue and an integer given its residue. A mismatch has probabilitydelta.
On matchedresidues AND Haar success, both tuples agree by the precision lemma.
Thus for ANY measurable observable of Z_m (in particular the entire J-vector),
 TV(law_N,law_Haar) <= min(1,epsilon_(m,H)+delta_(N,H)).
There is only ONE epsilon term. No separate bound on integer failure mass is
needed: Haar success plus matchedresidue already forces finite agreeing
integeroutput. The cemetery convention makes both distributions total.

EXACT CHECKS
finite_source_check.py independently enumerates130944odd H-bit cylinders for
m1,2,3 andH8,10,12,14,16. Its success/failure counts match the NBformula EXACTLY;
42distantpositive lifts keep the successful tuple;48residue-distribution tests
match the exactTV formula. No trajectory sample substitutes for missing bits.

Concrete valid bounds:
-N=2^64,m4,H48:delta0 andTV<=6930532111/140737488355328<2^-14.
-N=2^128,m8,H96:delta0 andTV<2^-31.
-N=10^12,m3,H32:TV<=610053904751713/104857600000000000<2^-7.
These are exact rational/binary inequalities, not floating estimates.
FINITE_SOURCE_AUDIT.json records all counts and constants.

SIMULTANEOUS GROWING HORIZONS / ACTUAL-INTEGER POPULATION CLT
This transfer is stronger than only fixed-dimensional convergence. If
m_N->infinity andlimsup m_N/log2(N)<1/4, choose constantsc,alpha with
 limsup m_N/log2N <c<1/4 and4c<alpha<1,
andH_N=floor(alpha log2N). Then delta<=K/N<=N^(alpha-1)/2 tends0.
The NBtail is a binomial lower tail withthreshold2m_N+2 andlengthH_N-1.
Its threshold ratio is eventually below1/2 because2c/alpha<1/2. The elementary
Chernoff bound (apply Markov toexp(-lambda Binomial), then optimize lambda)
therefore gives exponentially small epsilon inH_N, hence epsilon->0.
Thus the joint m_N-vector law is TV-close to its Haar reference law.

Combining with the previously proved Haar CLT yields the actual finite-source
population result
 (sum_(i0..m_N-1)J_i -4m_N)/sqrt(m_N) => Normal(0,4)
when n is uniform among the first N positive odd integers. The probability
of the cemetery output tends0 by the same coupling. An average-cost law in
probability follows likewise. These are simultaneous finite-population/time
statements and make no assertion about one fixed positive integer trajectory.

CRITICAL SCIENTIFIC QUALIFICATIONS
The law concerns UNCONDITIONED canonical trajectories, continued beyond first
descent. Stopping at firstdescent or conditioning on survival above the source
changes the law. The result does not control precisely those rare survivor
paths needed for a universal stopping proof. Nor does Haar-a.s genericity by
itself imply genericity of positive integers, a countable Haar-null set. The
explicit finite-source TVcoupling is what justifies this particular finite-
horizon arithmetic transfer; it does not transfer every infinite-path event.
No global Collatz resolution, independent source-only debt ceiling or all-
integer stoppingtime bound is obtained.
