ADIDAS — a size-aware adaptive monitor must certify the peak, not only the landing

Outcome. An exact, proof-carrying, capped Collatz monitor is implemented in adaptive_cap.py. It retains the exact source b, current integer x, and explicit cap C. It compresses maximal runs of exponent one, and then adapts its next run to the actual current integer. This escapes the fixed-fingerprint interface because exact current size and continued observations are retained. It does not prove that a finite cap suffices for every source. The elementary mathematics below has no novelty claim.

1. Exact block and peak lemma.
For positive odd x, write x+1=2^r m, with r=v2(x+1)>=1 and m odd. Set s=v2(3^r m−1)>=1. The next r accelerated odd steps have exact valuation word 1^(r−1),(s+1), and endpoint
  y=(3^r m−1)/2^s.
For 0<=j<=r−1, the exact odd states are
  x_j=3^j 2^(r−j)m−1.
They strictly increase until the final step when x>1. The pre-final odd peak is
  P=2·3^(r−1)m−1.
The final exponent is at least two, so y=(3P+1)/2^(s+1)<=P, with equality only when x=P=y=1. Hence the exact maximum over the whole accelerated block is P. Over the corresponding original unaccelerated Collatz segment it is 3P+1. The distinction between these two cap conventions must remain explicit.

Proof. Substituting x_j into (3x_j+1)/2 gives x_(j+1) while j<r−1, and x_j≡3 mod4 in those cases. At j=r−1 the numerator is2(3^r m−1), whose exact valuation is s+1. All claimed expressions and the peak follow.

2. Exact first cap crossing.
For x<=C<P, the first accelerated odd cap crossing occurs at the least j in1..r−1 for which
  3^j(x+1)>2^j(C+1).
Since x_j strictly increases there, exact integer binary search finds this index using O(log r) comparisons, without floating logarithms or replay of all omitted states. A verifier needs only x_(j−1)<=C<x_j. Arithmetic cost of powers is not unit cost and grows with the involved bit sizes; no wall-clock speedup claim is made.

Actual failure fixture for endpoint-only cap accounting:
  source3, cap3: r2,m1,s3, endpoint1, peak5.
The compressed endpoint is below cap and below source, but the actual odd path is3→5→1. Thus an endpoint-only cap contract incorrectly omits a real cap exit. The descent itself is still valid; the mistake would be certifying that the excursion remained inside the cap. This is a contract audit of our proposed monitor, not an allegation about an external implementation.

3. Finite capped trichotomy.
Fix positive odd source b and integer cap C>=b. At each current odd state x, compute the exact block above.
- If P>C, return CAP with the first crossing position and state.
- Otherwise advance to y. If y<b, return DESCENT.
- If y repeats a stored block-boundary state, return CYCLE with the closed actual path.
- Otherwise keep y and adapt to its next block.
The monitor terminates within K=floor((C−b)/2)+1 blocks, returning one of these three sound certificates. All returned DESCENT and CAP positions are their actual first respective stopping events (whichever occurs first). CYCLE is a certified repeated boundary, not a claim of the earliest repeated odd state.

Proof. Until a terminal block, every increasing within-block state and endpoint is in the finite odd interval[b,C]. Every nonterminal endpoint is new. If K blocks all failed to terminate, there would be K+1 distinct block-boundary states in an interval containing only K odd integers, impossible. The within-block formulas prove the first-event claims and that no skipped state descended. A repeated boundary under a deterministic exact map gives an actual integer cycle. This finite-state argument is standard; it supplies no all-source cap and does not exclude an unknown positive cycle.

4. What remains for a global proof.
A CAP result means only that the chosen cap was insufficient; it is not divergence. A returned CYCLE must be inspected as an actual finite closed path; the only cycles in these finite tests are the known fixed odd state1. Unboundedly increasing caps without a separate well-founded argument do not prove termination. To turn this into global Collatz one still needs an a priori all-source guarantee ensuring eventual descent (or an equivalent global invariant), rather than merely the finite capped trichotomy. It is a concrete sound adaptive consumer, not a new stopping theorem.

5. Practical budgets and UNKNOWN.
The mathematical finite-cap trichotomy does not justify a huge finite computation. The implemented generate function therefore has explicit max_blocks=1024 and max_bits=16384 defaults. It stops with UNKNOWN when the block budget is exhausted or the current entrance exceeds the bit budget. UNKNOWN certifies only the recorded prefix, with no conclusion about its eventual descent. It can have zero blocks. Thus the actual bounded program has FOUR outcomes, and the three-outcome K-block theorem applies only if sufficient resources are allowed. The bit budget bounds entrance size; the exact block intermediates can have about1.585 times as many bits. It is not a claim about exact RAM or wall time. Three zero/one-block and low-bit budget controls passed.

6. Reproduction and tests.
Run python adaptive_cap.py. Standard-library integers only, no network or GPU. Certificate generation and verification use compressed arithmetic; the reference replay independently applies3x+1 and repeatedly halves until odd. For every odd b in1..20001 and caps b,2b,4b, the classifications, terminal states and observed clocks matched. Controls include source1, exact cap equality, and a4096-bit all-ones run with an interior crossing at index2049 found by binary search. Five certificate mutations (peak, exponent s, status, terminal state, step count) were rejected.

The verifier rejects non-integer certificate fields, including floats and booleans, and refuses Python -O because its exact checks use assertions. A sixth mutated certificate replacing an integer peak by a float was rejected.

Recorded CHECKS.json:
  30,003 certified source/cap pairs;
  51,950 directly replayed odd steps;
  34,952 compressed blocks;
  21,031 DESCENT,8,969 CAP,3 CYCLE cases;
  at most10 compressed blocks in this prescribed finite audit;
  all checks passed, approximately0.3 seconds CPU test wall time.
These counts are finite interface tests. Their frequencies are not a random-orbit model, density theorem, or proof that arbitrary starts descend.

Recommended sharing. If parent publishes the combined adaptive consumer, include the exact peak/crossing contract and the source3 fixture. Do not publish a stand-alone paper claiming novel Collatz progress for this elementary lemma. This is useful because the user now specifically asks to retain size while adapting, and compressed algorithms must account for sizes they skip.
