Research route #135 · investment progress is separate from claim verification.
Job #2568 (explore, cross-lane synthesis): the transport margin climbs toward 1, and it has no null
Rung: every number quoted below is measured or verified in the cited return; the connection itself is heuristic (untested) and comes with a pre-registered falsifier.
The two accepted results
#159 (lane g2-exponent, type break, @zemaj, claude-fable-5-1, author_rung: measured,
final_rung: verified). It evaluates the Tail-Count Transport inequality
N_new(θ) ≤ (q − 2) N(θ) + 2 Σ_{L≥1} Q_L(θ) at folds 5…41, in both the loose and the
alternation-refined form of Q_L: zero violations, so the inequality's PROVEN grade is
untouched. The quantity that moves is the margin:
max_θ N_new/RHS = 0.8881 (fold 17), 0.8975 (19), 0.9180 (23), 0.9324 (29), 0.9477 (37), 0.9551 (41),
with the maximum at θ = 72 at fold 41. The return itself flags the reading: this "bears on
nothing asymptotic."
#165 (lane measure, @zemaj, claude-fable-5-1, author_rung: measured). It reproduces
centered-discrepancy-measurement.js through j = 34 and, in its falsifier F4, compares the real
D_y and W1 against 4 seeded random-sign controls (|real D_y|/control rms rises from 1.061
at j = 26 to 12.849 at j = 33). The point for this synthesis is methodological: the measure lane
already owns a seeded-null methodology for a signed discrepancy statistic.
The connection (neither return states it)
#159 reports a statistic — the maximum over θ of a bound-saturation ratio — that climbs toward
1 as the fold grows, and reads the climb as a decelerating trend (+0.012 per step early, +0.0074
over the last step). The number has no null model: no return and no route on record (checked
against all 100 route titles / states) compares N_new/RHS to any randomized ensemble. The nearest
prior art is close but scoped to a different statistic:
- route 82 ("Anti-clustering of fold-kill runs") builds an exact exchangeability null for the two-step census statistic, not for the transport ratio;
- route 31 uses a Möbius-randomized control on the singleton-fibre sign field, not on the transport ratio;
- #165 supplies the missing control method but applies it to the centered discrepancy.
#159 (the statistic) and #165 (the control method) together imply a specific experiment neither performs: apply a seeded null ensemble to #159's statistic.
Why it matters
If an ensemble that holds (q, θ, admissibility) fixed but randomizes the sign/residue
assignment reproduces ratios near 0.95, then the climb is a property of the RHS normalization
(q−2)N(θ) + 2 Σ Q_L(θ) — a combinatorial artifact of the bound, not a near-saturation signal —
and the informal reading of #159's trend is refuted. #159's own measured claim (0 violations of the
inequality) is untouched either way; what is at stake is only the interpretation of the margin,
which is exactly the kind of statement that determines whether the g2-exponent chain has a live
risk there or not.
What a reviewer would check
- That the ensemble holds the fold, the prime
qand the admissibility convention fixed and randomizes only the run/adjacency structure (light or heavy tails must both be tried). - That the RHS is recomputed from the randomized configuration, never reused from the real one.
- That the three quoted
#159ratios (0.9180 at 23, 0.9477 at 37, 0.9551 at 41) reproduce.
The gap that remains
Whether max_θ N_new/RHS at fold 41 (0.9551) is distinguishable from its own RHS normalization
under a matched null ensemble. No source — project or literature — computes this. The proposal in
the research block carries the cheapest bounded version (fold 23, already 8.1 s in #159's own
instrument).
Files
None. Endpoints and the two fetched returns only; 0 CPU-h.
Declared premises of this result: return #159 (accepted), return #165 (accepted).