Bit-17 attenuation ladder
If the effective bit-17 restart creates a connected simplification, its direction should replicate across carry 22, a5[22], Σ0(a5)[0], T2₅.low4 and a6.low4 against matched no-later-reset controls.
Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.
What was tested?
If the effective bit-17 restart creates a connected simplification, its direction should replicate across carry 22, a5[22], Σ0(a5)[0], T2₅.low4 and a6.low4 against matched no-later-reset controls.
Why the test is meaningful
A stage-by-stage ladder prevents a local difference from being silently extrapolated through nonlinear consumers. Matched controls isolate the effective-reset condition, while sign replication across an untouched cohort tests stability rather than magnitude alone.
Δ_s = E[N_s|E17] − E[N_s|matched control]stages={c22,a5[22],Σ0[0],T2₅.low4,a6.low4}replication(s) ⇔ sign(Δ_s^original)=sign(Δ_s^fresh)How it was tested
Join records without losing source identity, build exact masks at five connected stages, select matched controls, compute predefined nonlinear-support contrasts and repeat all stages on the fresh cohort.
What happened
All joins and semantic checks passed. Four of five difference signs replicated. Fresh differences were −0.404 at carry22, −0.654 at a5[22], +0.259 at Σ0 bit0, −0.0575 at T2₅.low4 and +0.0219 at a6.low4. Σ0 bit0 was the only sign that did not replicate.
Exactness and statistical controls
The full five-stage family was reported, including the non-replicating stage. Zero join errors protected cohort identity. Replication assessed direction only; no p-value or universal effect was inferred from these fixed groups.
What the result means
The reset reduces local support through a5[22], but the effect is altered by Σ0 and does not simplify a6. This is attenuation evidence, not a usable end-to-end shortcut.
Limitations
- Matched observational groups are not randomized.
- Support differences are descriptive representation metrics.
- Four sign agreements do not establish mining economics.
Evidence trail
Freeze the five interfaces and control rule, reconstruct exact ANFs, preserve all records and compare the original and fresh difference signs without feature selection.
Canonical variants
SUBENGINE-V42ASUBENGINE-V42BSUBENGINE-V42CSource: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.