model card: add MMLU/GSM8K general benchmarks (base vs adapter) from the candle harness

README.md

@@ -64,9 +64,29 @@ The adapter **improves every active domain with zero regressions.**
 Domains helped: 9 / 9. Domains hurt: 0. A held-out CE regression guard (ceiling = base + 0.15)
 was active for the whole run and never tripped, so the base capability is provably intact.
 
-> Honesty note: these are domain-CE deltas on the Aether holdout, not general-benchmark
-> (MMLU/HumanEval) numbers. The claim is narrow and measured: on the Aether knowledge domains the
-> 10-Sephirot adapter is a consistent, regression-free improvement over the raw 7B base.
+> The numbers above are domain-CE deltas on the Aether holdout. General-benchmark numbers
+> (MMLU, GSM8K) are below.
+
+## General benchmarks (base vs adapter)
+
+Off-the-shelf lm-eval cannot load the native candle build, so these were produced by a
+purpose-built candle harness (`aether-v7-eval`) that scores the SAME frozen Q4 weights twice,
+once with the Sephirot adapter active and once with it off. MMLU is multiple-choice
+loglikelihood over the A/B/C/D answer tokens; GSM8K is greedy chain-of-thought generation with
+final-number extraction.
+
+| benchmark | n | base | v7.1 (adapter) | change |
+|---|---|---|---|---|
+| MMLU (all subjects) | 14,042 | 71.28% | 71.17% | -0.11 |
+| GSM8K | 625 | 67.8% | 77.8% | +10.0 |
+
+Read this the way it reads: **general knowledge is held** (MMLU is flat across the full 57-subject
+set, the regression guard never tripped), and **multi-step reasoning improves** (GSM8K up ~10
+points on a 625-question sample, partly from the adapter following the chain-of-thought and
+final-answer format more reliably). The adapter does not trade away breadth for the domain gains.
+
+(GSM8K is a 625-of-1319 sample: the full run is generation-bound on a single 12 GB card and the
+sample is already statistically tight. MMLU is the complete set.)
 
 ## Training