GOUVERNAIL / FRACTION / VOLCAN — A RAW-LEDGER STEERING LIMIT
25 October2026. CPU-only parent continuation; no new subagents.

The draw was recorded before experiments. Gouvernail suggested choosing labels;
fraction suggested pin-height complexity; volcan suggested costly reserve bursts.
These are research prompts, not assumptions or evidence for the mathematics.

THE EXACT ACCOUNTING LIMIT
Consider any VALID phase-pin labelling of one fixed actual odd valuation prefix,
with finite reserves and all pins odd2adic. Each step consumes its division
exponent from the active reserve; each change at the same state contributes
Delta=V_new-V_old. Let P=sum max(Delta,0), N=sum max(-Delta,0).
The known exact ledger is S=V0-Vf+P-N, withVf>=1. Hence the raw debt bound
 B=V0-1+P
satisfies the exact decomposition
 B=S+(Vf-1)+N >= S,
and, sinceD=2k-S,
 B-D=2(S-k)+(Vf-1)+N.
Thus optimizing pin labels cannot make THIS UNWEIGHTED bound smaller thanS.
It can remove coding-induced negative-variation loss or change the terminal
reserve; it cannot remove the division drain discarded by the original D<=S
inequality. This is an immediate consequence of the previous ledger, useful
as a consumer-design limit, not a new Collatz theorem or priority claim.
It does not rule out weighted potentials, other invariants, or genuine new
upper control on refill amounts. Pin hits require separate cycle handling.

EXACT FIXED-PREFIX DEMONSTRATION
Source12235060455, first333 oddsteps (117canonical blocks): S518,D148.
All these actual states are still above the original source.
Canonical code:104type changes,V0=9,positive refills510,Vf1,N0,B518.
One whole-prefix periodic word:zero switches,V0=519,Vf1,B518.
That large fraction has a530-bit numerator and528-bit denominator; its reserve
is not a source-size-only allowance (the source itself has34bits).
Alternating those valid descriptions each step keeps the actual D148 but
creates42637 positive and42637 negative coding credits; B43155.
An offline dynamic program choosing between those two admissible labels at
each step obtains B518. Since the global floorS518 is attained by the
canonical code, NO admissible finite-reserve labelling can improve this raw
bound on this prefix. The DP is a retrospective benchmark using the already
known prefix, not a causal descent algorithm.
Whole-word weighting2/7 can reconstruct D148, but encodes that already known
word in a large rational pin; it supplies no independent a priori ceiling.

VALIDATION
python steering.py (Python standard library, without-O).
Exact integer valuations and Fraction arithmetic; every phase/exponent and
reserve transport verified. Four descriptions of333steps checked. DP result
separately compared with exhaustive8,64,512label choices at horizons3,6,9.
No floating point, network, GPU, paid compute or outside credentials.

SCIENTIFIC DECISION
Do not spend another run trying to beat a source-size ceiling solely by
minimizing raw positive label refills: this architecture has anS-floor.
Next meaningful direction must change the actual inequality/control the
mathematical refill-drain balance or charge dictionary/pin complexity.
No new paper is justified. A short consumer-design clarification is enough.
