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

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.

`run_crossover_baseline.py` is written and unrun: 2B + scaffold vs the 8B `gemma-4-E4B-it` with no memory. This is Study 001's own thesis question and the one experiment that answers it.

The harness computes bootstrap intervals only for pra_strict and UAR, so pra_lenient needs one added first. The repo already has bootstrap_ci, paired_bootstrap_diff, holm_bonferroni and minimum_detectable_effect.

Every Study 001 number is a single seed and a single run. The retrieval finding in particular — recall up 0.16% → 15.13%, abstention *down* 13.75% → 8.75% — is the kind of claim that needs ≥3 seeds and a paired bootstrap before it should be relied on. The repo already has `bootstrap_ci`, `paired_bootstrap_diff`, `holm_bonferroni` and `minimum_detectable_effect`; nothing new is needed but compute.

(E1). No `pra_lenient` or UAR figure exists for any DPO checkpoint; `dpo-v1-scale` appears only in pairwise comparisons.

(E13). The preference pairs were generated *before* SFT v1's dedup fix and ratio rebalance: 2263/2277 user turns match the current `sft/v1/train.jsonl` but **0/2277 system texts** do, so the retrieval context differs from the corpus the SFT stage trained on. Regenerate the pairs, then measure the stage. The cause of the divergence was never pinned down and should be.

(24 probes built, unrun) and **H24 emotional range** (27 probes across 9 registers, unrun).

(RQ13, second half) — the base model is a VLM, but nothing in the memory tier comes from an image.

(E6). The quantization work was `llama-bench` plus a coherence check; no quality benchmark was ever run against a quantized checkpoint, so "Q3_K_S is the smallest verified-coherent level" rests on manual generation inspection.

(E5). `README.md` and the cards' files tables say 8.64 / 4.61 / 3.18 GiB; `hf_readmes/gguf_README.md`'s own evaluation table says 8.62 / 4.59 / 3.17 GiB. No artifact backs either. Re-run `llama-bench` against the GGUF files, commit the output, and correct whichever set is wrong.

(E6). No `llama-bench`, `llama-imatrix` or `llama-quantize` output is committed (`*.gguf` is gitignored and `results/` holds no quantization directory), so every size, throughput and perplexity figure is restated from run logs a reader cannot check. Even without re-running the benchmarks, a `results/quantization/` directory holding the original stdout would close most of this.

(E22). It is present on all 688 probes and empty in 688/688, which is why `pra_strict` is ~0 throughout and `pra_lenient` is the reported metric. Populating it makes `pra_strict` meaningful; dropping it removes the need to explain it away.

It covers 120 of 688 probes (15 per category, seed 1337) and is reported as an inverted U peaking at k=8. At n=120 the curve's shoulders are not distinguishable from noise; a full-688 run costs inference, not training.

It is defined and implemented in the harness and is **0.0 in every saved run** — because it is computed but not called, not because the models score zero. Either wire it up or remove it from the reported metric set.

(E4). `distill_v1.yaml` (base and teacher), `dpo.yaml`, `dpo_v1_more_epochs.yaml` and `dpo_v1_scale.yaml` all reference `"main"`. Until they pin a SHA, no run launched from them can be called reproducible.

No human review exists anywhere: every datasheet says "not human-reviewed", and no reviewer log or teacher transcript was retained, so the `gpt-5.6-luna` teacher identity is self-reported and unverifiable from the repository. A few dozen probes scored by a human would bound how much the judge-based metric can be trusted.