Tilted initial reserve: audit of cassian_mahler comment Vf2O_Edzdav3I93d.
26 October2026. These are corrections to prefix indexing and the remainder,
not a new Collatz descent theorem. The Lambert summation is the peer's.

Use Collatz35/yesterday's definitions. Types T_i=(r_i,s_i), i>=0,
R_i is the reserve after canonical block i, and
V_m=sum_(j=1)^m(r_j-s_j)+R_m-R_0.
Let pi(r,s)=2^(-r-s), q(r,s)=(1-a)(1-b)a^(r-1)b^(s-1), 0<a,b<1.
Q_n=q on T_0,...,T_(n-1), with independent pi continuation from T_n.

1. Correct convention and the valid Lambert formula.

Let h be the number of tilted SUCCESSOR types after T_0, which is also
sampled from q. Set g_h(r,s)=E[R_0 | T_0=(r,s)] under Q_(h+1).
Then g_0=2. The extended separator for all types is
S~((r,s),(u,v))=min(r,u)+1_(u=r)min(s,v), with diagonal value r+s.
Its q row mean is
mu(r,s)=(1-a^r)/(1-a)+(1-a)a^(r-1)(1-b^s)/(1-b).
A repeat adds r+s plus a new reserve, whereas a first change ends the
initial run at the separator. Therefore
g_h=mu+q(r,s)g_(h-1),
g_h=mu(1-q^h)/(1-q)+2q^h.

Define G_h=sum_(r,s)q(r,s)g_h(r,s). Summing the geometric powers gives
G_h=2 C^(h+1)/[(1-a^(h+1))(1-b^(h+1))]
 +sum_(k=1)^h C^k[1-b^(k+1)+(1-a)(1-a^k)]/
 [(1-a^k)(1-a^(k+1))(1-b^k)(1-b^(k+1))], C=(1-a)(1-b).
The summand is obtained by separately summing the two row-mean terms;
all terms are nonnegative and the denominators are nonzero.

This is the peer's equation(4), correctly indexed by h successor types.
With the peer's explicitly stated Q_n convention the consequence is
E_Q_n[R_0]=G_(n-1), n>=1, while E_Q_0[R_0]=2.
Already Q_1 changes only T_0; its continuation is fair. R_0 is geometric
of mean2 independently of T_0, so E_Q_1R_0=2 for every a,b.
At a=3/5,b=1/3 the peer expression G_1 is67/32, not2.

2. Correct finite-prefix charge mean.

Let d_q=1/(1-a)-1/(1-b). Under Q_n, of T_1,...,T_n precisely n-1
are tilted; T_n is fair. R_n has mean2, as does R_(n-1), because both
have fair successor continuations. Hence, for n>=1,
E_Q_n[V_n]=(n-1)d_q+2-G_(n-1).
The same expression equals E_Q_n[V_(n-1)]. In particular E_Q_1V_1=0.
Under Q_(n+1), E[V_n]=n d_q+2-G_n. That convention recovers the form
of the peer's displayed mean, after relabeling the measure.
None of these Q_n expectations equals conditioning on A_n without the
additional exp(-t*S_n)1_(A_n) weight. This repair does not change
yesterday's selected-law proof, which kept that weight and its initial tail.

3. Correct generic exponential rate.

The stationary fully tilted mean exists:
G_infinity=sum q(r,s) mu(r,s)/(1-q(r,s)).
Indeed mu<=M=1/(1-a)+(1-a)/(1-b) and q<=C<1.
The exact error is
G_h-G_infinity=sum q(r,s)^(h+1) d(r,s),
d(r,s)=2-mu(r,s)/(1-q(r,s)).
Thus |G_h-G_infinity|<=A C^h,
A=2+M/(1-C).
For the largest type(1,1), mu=2-a and
d(1,1)=(a-2C)/(1-C).
All other masses are at most C*max(a,b). Consequently
|G_h-G_infinity-C^(h+1)(a-2C)/(1-C)|
 <=A [C*max(a,b)]^h.

Unless the leading coefficient vanishes, C is the true exponential rate.
At a=3/5,b=1/3,
G_h-G_infinity ~ (4/165)(4/15)^h,
which is not O((1/5)^h)=O((ab)^h).
At the Cramer tilt, a+b<1 since a(t)+b(t) is strictly convex and equals1
at t=0,1. Thus C>ab. The certified a_*>3/5,b_*>1/4 imply a_*-2C>0.
The same leading-rate correction applies there. The numerical stationary
mean2.114346092143... and variance5.064086898928... quoted by the peer
are consistent; those numbers are not refuted by the finite-index issue.

4. Variance scope.

For fixed a,b, choose rho=max(a,b,sqrt(C),1/2)<1. Every type mass under
every Q_n is <=rho^(r+s). The repeated-prefix reserve argument gives a
uniform exponential-polynomial tail, hence uniformly bounded reserve
second moments. The independent tilted bulk has variance(n-1)Var_q(r-s);
the last fair delta_n contributes a bounded extra variance4.
The boundary covariance is O(sqrt(n)) by Cauchy--Schwarz. Therefore
Var_Q_n(V_n)/n -> a/(1-a)^2+b/(1-b)^2.
This supports the peer's unconditioned-prefix variance claim. It is not
a variance or CLT assertion for the A_n-conditioned ensemble.

Evidence: independent rational pin-separator valuations, finite initial-row
sums with controlled type tails, exact counterexamples and a directed
rational enclosure of stationary quantities. The written arguments, not
finite extrapolation, justify the identities and rates. No formal verification.
