MOTOCROSS — Size-aware adaptive affine-cylinder certificate consumer
25 October 2026. CPU only; no network/API writes, GPU, or SYMBIOSE.

RESULT AND STATUS
An implemented exact adaptive stopping-tree consumer retains the actual source and current values as affine functions of an integer quotient, with a bounded quotient interval representing the source-size range. It batches many source integers under one proof node, certifies an exact first-descent time for each closed interval, and records UNKNOWN for every surviving interval when its depth/node budget runs out. An independent verifier checks transformation, strict descent, and complete coverage. This is classical stopping-time sieve algebra packaged as a proof-carrying interface, NOT a new Collatz convergence theorem, novelty claim, computation record, or demonstrated runtime speedup.

THE PRECISE LEMMA
Use shortcut time C(x)=(3x+1)/2 if x is odd, and C(x)=x/2 if x is even. A node has
 source n=c*t+d, current x=a*t+b, t an integer in [L,U], depth k,
where a is odd, c is a power of two, and all earlier tested states are >=n. Its exact descent predicate is
 (a-c)*t+(b-d)<0.
The solution set of this strict affine inequality, intersected with [L,U], is an interval touching one endpoint (possibly empty or full). Certify that interval with exact first-descent time k. The complement is again one interval, on which the current state is >=n.

Split the remaining quotient as t=2u+e, e in {0,1}, retaining exactly
 ceil((L-e)/2)<=u<=floor((U-e)/2).
The source becomes n=2c*u+(c*e+d). Put z=a*e+b. If z is even, the next state is a*u+z/2. If z is odd, it is 3a*u+(3z+1)/2. Since a is odd, parity is fixed on each child, and both updates are exact integer shortcut steps. The children partition the entire retained source set, with no missing high-valuation branch.

For positive odd sources 3<=n<=N, initialize after the known first odd step with c=2,d=1,a=3,b=2,k=1 and 1<=t<=floor((N-1)/2). This first state is strictly above n. Every child preserves the representation and previously non-descending prefix. Induction proves all certified first-descent clocks, all UNKNOWN prefixes, disjoint source coverage, and closure+UNKNOWN counts equal floor((N-1)/2).

If the finite tree closes every odd source through N, strong induction plus the immediate descent of even integers gives Collatz convergence through N. This is a finite certificate only. For an unresolved leaf the tool supplies no finite upper stopping bound. No argument here makes the infinite search well founded.

IMPLEMENTATION AND INDEPENDENT VERIFICATION
adaptive_builder.py constructs nodes with exact integer coefficients. verify_certificate.py derives every state again from the root and parent parity branch; it does not trust supplied coefficients. It checks strict inequalities at both interval endpoints, complementary non-descent, exact child coverage, no repeated/orphan nodes, and the source counts. Direct per-integer replay provides a separate baseline and verifies the full first-descent histogram when all nodes close.

BENCHMARKS (SHORTCUT TIME, NOT ACCELERATED ODD TIME)
N=100,000: all49,999 odd sources certified in53,569 symbolic nodes; independently replayed297,442 individual shortcut steps; maximum first-descent135.
N=1,000,000: all499,999 odd sources certified in428,448 symbolic nodes; independently replayed2,984,856 individual shortcut steps; maximum first-descent176.
The latter is a factor about6.97 fewer symbolic nodes than individual shortcut transitions, but each symbolic node is more expensive and serialization costs dominate. This is operation-count batching only. Builder took about3.66s and the final schema-hardened verifier including direct replay about4.92s in this environment; these are not controlled algorithm-performance results. The uncompressed one-million certificate is49,680,193bytes, so this implementation is deliberately transparent rather than storage-efficient.

HONEST BUDGET EXHAUSTION
N=100,000, depth cap2: three symbolic nodes certify24,999 sources, exactly the odd n=1 mod4 with n>=5; the remaining25,000 sources are UNKNOWN. The representation neither guesses descent nor omits them. A node cap also leaves explicit UNKNOWN intervals, and is tested independently.

ADVERSARIAL CONTROLS
controls.py replayed139,510 source/certificate memberships over130 bounded parameter cases (fixed seed20261025), varying N, depth and node budget. It separately checked every closed source's exact first descent and every UNKNOWN source's non-descending prefix. Six corruptions were rejected: a removed child, altered affine state, invented descent, dropped node, floating-point injection, and negative child index. See CONTROLS.json. First-descent equality remains non-descent; source1 is deliberately excluded at the root and is never certified as descending.

LIMITS AND USE
This goes beyond the previous fixed-fingerprint limitation by retaining an unbounded integer quotient and actual source-size interval, then adapting the branch and depth. It does not refute the earlier theorem because those data were explicitly outside its scope. Its useful deliverable is a concrete consumer contract: any proposed acceleration can emit these local certificates, and coverage can be independently audited. A global claim would still need a proof that all remaining nodes close, or another size-aware invariant that bounds the unresolved frontier. The implementation establishes neither.

For publication: suitable as a small reproducible tool note or response, if it serves a current consumer. Not appropriate as a new major mathematical paper or a newly discovered finite Collatz verification bound. Prior art includes classical exact valuation cylinders, stopping-time sieves, and the previously read paper8/v2; no exhaustive external novelty search was performed.

REPRODUCTION
python adaptive_builder.py --N 1000000 --out CERTIFICATE_1m.json
python verify_certificate.py CERTIFICATE_1m.json --replay
python controls.py
All operations are bounded CPU integer arithmetic. All jobs finished. No failure was suppressed or evidence generated by floating point.
