# Sniff controls: prospective interim rule

Written before any body-202 or body-303 outcomes were generated or inspected.

The full design has 31 runs: three matched transformer-body seeds, three donor
seeds, three transplant families, three matched baselines, and one designated
no-op reproducibility run. Five body-101, donor-501 runs were already observed
before this rule was written.

Recommissioning audit amendment: the first GPU's ordered and shuffled donor
weights were not preserved, and CUDA donor training had not required
deterministic kernels. Those five results remain an exploratory pilot but are
not pooled into the new crossed matrix. The crossed matrix therefore starts
again from deterministic donor artifacts. This amendment was made before any
outcome from the new matrix was generated.

The first informative interim look occurs after 13 total runs. It completes the
donor-501 block for all three body seeds: each body has a baseline plus random,
token-shuffled Verne, and ordered Verne embeddings. The existing no-op is also
retained. This look varies the factor that confounded the original comparison.

Continue to all 31 runs if either condition holds:

1. Ordered Verne lowers the 907-row repeated-token rate more than both random
   and shuffled controls in at least two of the three matched body blocks.
2. The direction or rank order of the three embedding treatments varies enough
   across body seeds that donor-seed replication could materially distinguish
   an ordered-Verne effect from a generic transplant effect.

Otherwise stop for futility. The permitted conclusion is narrow: no
reproducible ordered-Verne advantage was observed at donor seed 501 across the
three tested body seeds. It is not an equivalence claim or a universal null.
All 13 runs, including adverse and null outcomes, must be reported.

The primary interim outcome is the change in repeated ASCII-alphanumeric-token
rate on the 907-row development evaluation versus the matched baseline.
Teacher similarity S, exact match, validity, unique-title count, and the older
62-row panel are secondary descriptive outcomes. No judge is queried.

## Interim decision

The 13-run interim completed at 2026-08-13 05:28 UTC. For donor seed 501,
ordered Verne changed the 907-row repeat count relative to the matched baseline
as follows:

| Body seed | Baseline | Ordered Verne | Relative change |
|---:|---:|---:|---:|
| 101 | 129 | 85 | -34.1% |
| 202 | 110 | 144 | +30.9% |
| 303 | 109 | 111 | +1.8% |

Condition 1 did not hold. Ordered Verne had fewer repeats than both controls
only for body 101. Condition 2 did hold: the family ranking changed across body
seeds. Ordered was lowest for body 101, highest for body 202, and intermediate
for body 303. Donor-seed replication could therefore distinguish an unstable
interaction from a family-level effect. The experiment continues to all 31
runs. This decision was recorded before any donor-502 or donor-503 907-row
outcome was available.