card: midtrain/ -- the arm before instruction tuning

41 arms now carry <family>/<arm>/midtrain/: the final midtrain save, before Dolci SFT, copied from scimt-dispatch-final-v1 and -glm and checked by sha256. Layout block and a paragraph naming the arms without one.

README.md

@@ -42,6 +42,7 @@ motivation is what the experiments vary.
 
 ```
 <family>/<arm>/base/           the midtrained + instruction-tuned model the adapters load onto
+<family>/<arm>/midtrain/       the same model before instruction tuning, straight out of midtraining
 <family>/<arm>/aft/<treatment>/   elicitation-finetuning LoRA adapters, by treatment and step
 <family>/<arm>/training/       training records for that arm
 batteries/                     raw eval responses, one archive per endpoint
@@ -66,6 +67,14 @@ carry the dose ladder also have `charter_0p25pct` through `charter_5pct` and
 quarter of a percent to five percent. The mixtures themselves are in
 [`dispatch-eft`](https://huggingface.co/datasets/arcadia-impact/dispatch-eft).
 
+**`midtrain/`** is the arm at the end of midtraining, before the Dolci
+instruction-tuning stage that produces `base/` — for studying what midtraining
+alone installs. The adapters were trained on `base/`, not on this. Most carry a
+`PROVENANCE.json` naming the source repo, revision and per-file sha256. There
+is none for `glm45_air_190m/control` or the `glm45_air_20m_legacy` arms, whose
+midtrain checkpoints were not kept, nor for `gemma4_26b_a4b_graft`, which is
+grafted rather than midtrained.
+
 `MODELS_DEFERRED.json` at the root is the worklist for a later checkpoint port.
 The standing policy through this consolidation was to take evals and scores now
 and defer weights, and that file records what was deferred, where it lives and