INDEPENDENT CERTIFIER DERIVATION: THE EXACT HAAR REFILL LAW
25 October2026. No empirical or independence extrapolation from integer traces.

REFERENCE SPACE AND FINITE CYLINDER INPUT
Use normalized Haar measure on the odd2adic integers. Remove the null set on
which a canonical block is undefined or any queried reserve is infinite.
For typeW=(r,s), word(1^(r-1),s+1), r,s>=1, its exact source cylinder has
normalized odd measure2^-(r+s). Its affine endpoint map on that cylinder is
a bijection onto the odd2adics preserving normalized conditional Haar measure:
parameterize the source by x0+2^(r+s+1)z and the output by y0+2*3^r z.
Thus successive canonical blocktypes W_i=(R_i,S_i) are iid with probability
p_(r,s)=2^-(r+s); the two coordinates are independent Geom(1/2) on{1,2,...}.
This is the standard finite-word coding input, not a new mixing assertion.

For clarity in this note, write COST_i=r_i+s_i, and write RES_i for the reserve
of the pin belonging to W_i at the state AFTER blocki. It has Geom(1/2)
distribution, mean2 and variance2, independent of all blocktypes throughi.
Indeed the output is conditionally Haar odd, and q*y-p is Haar on2Z_2 for
the fixed odd numerator/denominator pin p/q.

ONE-BOUNDARY REDUCTION
At boundary i->i+1, compare oldtype(r,s) and newtype(u,v). Same types have
refillJ=0. Different types have explicit separation
 t=min(r,u) ifr!=u; t=r+min(s,v) otherwise.
The incoming reserve is u+v+G, where G=RES_(i+1) is geometric and independent
of both old and incoming types. Therefore
 J=b+G, b=u+v-t>=0, on distincttypes.
The iid type input is valid. SuccessiveJ_i are NOT thereby independent:
the G at one boundary is a function of the future trajectory/type tail.

EXACT BASE DISTRIBUTION (UNCONDITIONED MASSES)
Same types: probability(sum_r2^-2r)^2=1/9, producingJ0.
For distinct types, b has the following subprobability masses:
 b0:1/9; b>=1:(3b+1)/(9*2^b).
Proof by disjoint cases:
1.r>u: separationu, sob=v. The r>u probability is1/3 and v has mass2^-b,
  yielding1/(3*2^b) for b>=1.
2.r<u: let d=u-r>=1. P(u-r=d)=2^-d/3. Since b=d+v, summingd1..b-1 gives
  (b-1)/(3*2^b) for b>=2. Together unequalr gives b/(3*2^b).
3.r=u: probability1/3. Ifs>v, b0 and P(s>v)=1/3, yielding1/9. Ifs<v,
  P(v-s=b)=2^-b/3, yielding1/(9*2^b). The remaining s=v is the same-type event.
The base masses sum8/9, as required.

CONVOLUTION, TAIL AND MOMENTS
Convolve the above subprobability withP(G=g)=2^-g:
 P(J0)=1/9;
 P(J=h)=h(3h-1)/(18*2^h), h>=1.
Equivalently its pgf is
 F(z)=1/9+4z(1+z)/(9(2-z)^3), |z|<2.
Summing polynomial/geometric tails gives
 P(J>=h)=(3h^2+5h+8)/(9*2^h), h>=1.
F'(1)=4,F''(1)=176/9; henceE J4,E J^2=212/9,VarJ68/9.

These refer to EVERY adjacent block boundary, including same-type J0.
For actual TYPE CHANGES only, condition onJ>0, probability8/9:
 P(J=h |J>0)=h(3h-1)/(16*2^h),
 mean9/2,secondmoment53/2,variance25/4.
Mixing these two normalizations would be a scientific error.

AGGREGATION AND NON-INDEPENDENCE
Exactly, including same-type boundaries,
 J_i=COST_(i+1)+RES_(i+1)-RES_i.
Thussum_(i0..m-1)J_i=sum_(i1..m)COST_i+RES_m-RES_0.
COST_i are iid,mean4,variance4. Both boundary reserves have geometric tails.
Consequences under this Haar REFERENCE law:
- Stronglaw: the iid cost average tends4, and RES_m/m tends0 almost surely
  by summable geometric tail bounds and the first BorelCantelli lemma.
- CLT: iidcostCLT plus (RES_m-RES0)/sqrt(m)->0 in probability yields normal
  limit ofvariance4 for(sumJ-4m)/sqrt(m).
- Asymptoticvariance: boundarydifference has bounded secondmoment, and the
  cost/boundary covariance isO(sqrt(m)) by CauchySchwarz, soVar(sumJ)/m->4.
The marginalvariance68/9 is NOT the long-run variance; iidrefill sampling
would give the wrong variance. Nothing requires independent reserves.

OPTIONAL EXACT RESERVE COVARIANCE
Cov(RES_i,RES_(i+h))=2/(2^(h+1)-1)^2, h>=0.
Condition ontypes W_i,...,W_(i+h). If their first change occurs before the
end, RES_i equals complete oldrun cost plus the separator of the first change,
and is determined by those types. RES_(i+h) has conditionalmean2; the centered
product has expectation0. If all h+1types equal, RES_i=h*COST_i+RES_(i+h),
whose centered product withRES_(i+h)-2 has expectationVar(Geom)=2. Hence the
covariance is2*P(W_i=...=W_i+h)=2*sum_w p_w^(h+1), the displayed formula.
This does not state that the J process is1-dependent or identify all cost/
reserve cross-covariances. No such inference was made.

EXACT FINITE CHECKS AND REMAINDERS
refill_law_check.py enumerates ALL r,s,u,v in{1,...,K} forK4,8,16,24,32:
1,450,240typequadruples. G is integrated exactly by its geometric distribution,
not sampled. The captured mass is(1-2^-K)^4. Every PMF/tail discrepancy is
between0 and omitted mass1-(1-2^-K)^4 <=4*2^-K.

Moment remainder bounds used independently of the claimed closed law:
J<=U+V+G. All three have geometric mean2,secondmoment6. For an oldtype
coordinate>K, the omitted mean bound is6*2^-K and omitted secondmoment42*2^-K.
For one incoming coordinate U>K, memorylessness gives
E[U;U>K]=(K+2)2^-K,
E[U^2;U>K]=(K^2+4K+6)2^-K.
ThusE[U+V+G;U>K]<=(K+6)2^-K,
E[(U+V+G)^2;U>K]<=(K^2+12K+42)2^-K.
Summing the two old and two incoming-coordinate tails yields safe upper bounds
(2K+24)2^-K for missingmean and(2K^2+24K+168)2^-K for missingsecondmoment.
The resulting rational intervals contain4,212/9,68/9. Five wrong probability/
conditioning/moment targets are rejected by the tightK32 intervals.

residue_law_check.py is a separate actual-map dyadic audit:43,648 odd cylinders
atinputprecisions8,10,12,14,16. A type is accepted ONLY if enough source bits
remain after consumed division exponents; a refill is accepted only if its
valuation is strictly below remaining known precision. Otherwise the cylinder
is counted UNKNOWN. Atprecision16 it certifies30848/32768odd cylinders and
retains1920unresolved (15/256), with separate causes. Each claimedPMF lies
between the exact accepted count and accepted+unknown mass. No integer sample
is used as a surrogate for missing high bits.

QUANTIFIER GUARD
The one-boundary law has a natural-density interpretation over initial positive
odd integers through finite-cylinder approximation with vanishing remainder.
An almost-sure infinite-path theorem on2adic Haar space does NOT automatically
become a natural-density-one infinite-path property of positive integers:
the latter form a countable Haar-null set. State the reference measure or the
order of finite-source/time limits explicitly. No all-integer stopping,
independent deterministic orbit samples, or new global debt ceiling follows.
