MH-4: run h3ere-alt on the battery — the only arm that can separate 'these values' from 'any structured content'

[emooreatx]

MH-4: run h3ere-alt — it is the only arm that can lift the "appropriate structured content" bound

The MH campaign currently bounds its content claim at "appropriate structured
content" rather than "these specific values" (RATCHET#15, third methodology
ask). That bound is not a power problem to be solved with more turns of the arms
we ran — it is the wrong arm. h3ere-alt was built for exactly this and
never used on the battery.

Why the ablations already run cannot answer it

contrast	varies	answers
ciris vs blank	content present / absent	does any content matter?
ciris vs neutral	value content drained	does value content matter?
ciris vs alt	which values	do THESE values matter?

blank and neutral both test an absence. Only alt holds form, length,
register and procedural content fixed and varies the values themselves — a
different real value system, built by the same mechanical partition→substitute→
author-in-isolation process, 68 of 1,153 lines changed, published diff.

Current MH-3 result for context (cluster bootstrap over arcs, U6 excluded, bare
deferrals scored as failures):

arm	rate
bare	24.0%
values-ciris (accord as prompt)	37.3%
h3ere-ciris	5.8%
h3ere-blank	16.9%

blank − ciris = +11.1%, 95% CI [+2.3, +21.3], p = 0.013. Something in the
content is load-bearing. Nothing run says which something.

Two shapes

A — alt at en only, 45 turns. Cheap, English-to-English so no language
confound, directly comparable to existing h3ere-ciris English cells.
Underpowered: ciris scores 0/45 ex-U6 in English, so a null would carry
almost no information. Same limitation h3ere-neutral already has.

B — ciris AND alt both on the English monoglot accord, five locales,
225 turns each. Properly powered. The symmetry is the point: both arms become
English-text-in-a-foreign-conversation, so the language mismatch is identical on
both sides and cancels. That mismatch only confounds when ONE arm is localized —
which is why neutral was restricted to en, and why the fix is to un-localize
both rather than localize one (which corpus overrides cannot do anyway,
CIRISAgent#1043).

B costs a re-run of h3ere-ciris on the monoglot accord, because the shipped-
localized ciris cells are not comparable to an English-only alt. ~450 pipeline
turns, negligible credits, many hours of wall-clock.

Recommendation: B

It also repairs something TORQUE got on the wrong instrument. form_vs_content
was one of the contrasts rescoped to equivalence on HE-300, where the movable
range was ~11% and the gold floor 4.1% — the effect could not have been resolved
there at any N. On this battery the effect sizes are an order of magnitude
larger, so the same question is answerable.

Pre-registered readings

result	reading
alt ≈ ciris, both ≪ blank	"appropriate structured content" — the bound stands, and is the finding
alt materially worse than ciris	these values specifically — the stronger claim is earned
alt materially better	report it; pre-registered as reportable
neither separable from blank	the MH-3 content result does not replicate; treat it as provisional

Blocked on nothing. Should be pre-registered with a frozen outcome taxonomy
before running, per RATCHET#15's second methodology ask — the deferral
re-scoring that produced the current content finding was post-hoc, and this
replication is the place to fix that rather than repeat it.