Repository navigation
MH-4: run h3ere-alt on the battery — the only arm that can separate 'these values' from 'any structured content' #21
Description
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.
Activity