The crossover, replication, and what Study 001 never measured.

run_crossover_baseline.py is written and unrun: 2B + scaffold against the 8B gemma-4-E4B-it with no memory. This is Study 001’s own thesis question — can a small model plus a scaffold substitute for scale — and the one experiment that answers it. H16 predicted a +15 to +35pp recall advantage; H17 predicted the larger model still wins by a wide margin on multi-step reasoning. Neither was tested. Needs a GPU.

Every Study 001 number is a single seed. Finding 1 in particular — the 5.0pp UAR drop from adding memory, and the false-abstention cost that followed — is the kind of claim that needs at least three seeds and a paired bootstrap before it should be relied on. The harness currently computes bootstrap intervals only for pra_strict and UAR, so pra_lenient needs one added first.

The field is empty in 688/688 probes, which is exactly why pra_strict is 0.0% throughout and pra_lenient is the reported metric (errata E22, guardrail G12). Populating it makes pra_strict meaningful; dropping it means the metric is chosen deliberately rather than by omission.

The quantization work was llama-bench plus a manual coherence check. No quality benchmark was ever run against a quantized checkpoint, and no quantization figure in the repository is backed by a committed artifact — *.gguf is gitignored and results/ holds no quantization output (errata E5/E6). That is why no quantization number appears anywhere on this site.

No full-PMB measurement of the DPO checkpoint exists, and the preference pairs were generated before SFT v1’s final regeneration: 2,263/2,277 user turns match but 0/2,277 system texts do, because the retrieval contexts diverged and the cause was never root-caused (errata E1/E13). Regenerate the pairs and measure the stage properly.

Both harnesses were built and neither was run: 24 abliteration probes and 27 emotional-range probes across 9 registers. They are pre-registered and idle.

RQ13’s second half. The base model is a vision-capable model — the processor loads as Gemma4Processor and the multimodal path is live at evaluation time — but nothing in the memory tier has ever come from an image. The benchmark has a vision category and the bake-off never exercised it.

Four of the six training configs still use "main" (dpo.yaml, dpo_v1_scale.yaml, dpo_v1_more_epochs.yaml, distill_v1.yaml), and distill_v1.yaml also uses "main" for its teacher, so runs launched from them are not revision-pinned and should not be described as reproducible (errata E4, guardrail G6). No re-run may be called reproducible before this lands.