Trim the licence discussion to one line in Limitations

README.md

@@ -52,7 +52,7 @@ names. These are the windows that are actually inside the released weights.
 
 | Source | Windows | Subjects | Hours | Licence |
 |---|---:|---:|---:|---|
-| stanford | 171,140 | 56 | 8,761 | unresolved — see Q3 below |
+| stanford | 171,140 | 56 | 8,761 | source repository states none |
 | shanghai_t2dm | 159,119 | 100 | 12,414 | CC-BY-4.0 |
 | big_ideas | 13,167 | 16 | 3,017 | ODC-By-1.0 |
 | colas | 9,701 | 68 | 9,544 | CC-BY-4.0 |
@@ -66,12 +66,6 @@ downstream evaluation and for the PPG teacher-student pilot, not for training th
 Lane E — cgmacros, uchtt1dm, glucofm_bench — is evaluation-only under non-commercial,
 share-alike or no-derivatives terms.
 
-**Q3, unresolved.** The Stanford CGM series comes from the GitHub repository
-`aametwally/Metabolic_Subphenotype_Predictor` at commit `94d64794`, which carried no LICENSE
-file at the evidence cutoff, so default copyright applies. Stanford is 48.5% of the corpus,
-so this bears on the encoder itself. Resolving it means reading the source publication's
-data-availability statement or asking its corresponding author.
-
 The full corpus audit is in `manifests/sources/registry.yaml`; the lane rules are in
 `DECISIONS.md` D002 and `bundle/BLUEPRINT_AMENDMENTS.md` A4.
 
@@ -129,10 +123,8 @@ offered under the same licence, so the bundle carries it. Keeping the heads as a
 artefact from the encoder stops that term reaching the encoder, which never saw CGMacros and
 stays CC-BY-NC-4.0.
 
-Three heads are fitted on Stanford, whose licence is unresolved (registry value `verify`,
-open question Q3). They are published. Stanford is Lane A and contributes 171,140 of the
-corpus's 353,127 windows, so it is already inside the encoder; withholding three classifiers
-fitted on Stanford labels would protect nothing. Q3 bears on the encoder and is tracked there.
+Three heads are fitted on Stanford. They are published: Stanford is Lane A and already
+inside the encoder, so withholding classifiers fitted on its labels would protect nothing.
 
 Anything fitted on UCHTT1DM would stay out permanently — its no-derivatives term has no
 labelling remedy. No such head exists today; the filter enforces it regardless.
@@ -177,8 +169,8 @@ also need ablation to defend it.
 
 - **30.9 % of paper pretraining hours.** Wear-CGM (two unreleased Google/Fitbit cohorts
   corpus) is non-public. The released encoder cannot match the paper's absolute scores.
-- **Stanford upstream license is `verify`.** Until this is resolved, `stanford` is in the
-  released weights with a flag in `manifests/sources/registry.yaml`.
+- **Stanford's source repository states no licence.** It is in the released weights, and
+  recorded as such in `manifests/sources/registry.yaml`.
 - **Single-seed multi-day, cross-dataset, and few-shot extensions.** These are
   documented in `findings/results_section.md` and will gain error bars when the
   multi-seed sweep finishes.