docs: sc300-v4 controlled ablation (scale at fixed mix)

README.md

@@ -29,11 +29,14 @@ This repository contains the **fm-v6** foundation checkpoint and the **scaling s
 | `scaling/sc27/` | 27.36M | 512 / 8 | v4 | 0.13B | 0.6451 | 0.3068 | 0.4296 | 0.5865 | 0.1877 |
 | `scaling/sc80/` | 88.19M | 768 / 12 | v4 | 0.38B | **0.6563** | **0.3600** | 0.4336 | 0.6280 | 0.1861 |
 | `scaling/sc300/` | 306.4M | 1024 / 24 | v5 (1 epoch, reduced budget) | ~1.28B | 0.6080* | 0.3090 | 0.4267 | 0.6064 | 0.1932 |
+| — (local, not released) | 306.4M | 1024 / 24 | v4: 18 real + 300 synthetic (308 logs) | ~1.28B | 0.6538† | **0.3929** | 0.4450 | 0.6206 | 0.1873 |
 | `scaling/ablation/sc80-v5-rmix14/` | 88.19M | 768 / 12 | v5 rmix14 (18 real ×14 + 2,500 synth) | ~0.41B | 0.6158 | 0.3969 | **0.4695** | **0.6456** | **0.2096** |
 
 Notes:
 - `fm-v6` and the `sc*` models use **different corpora and vocabularies** (58 vs 308 seen logs) and are not directly comparable. The scaling curve is `sc5 → sc27 → sc80 → sc300`.
+- † sc300-v4 is a controlled local run (best epoch 8); its checkpoint is not released. Validation is on corpus v4 and comparable with the other v4 points.
 - The `sc*` points are **under-budget** by design (0.22-0.42× the 20-token/parameter target; sc300 uses 4.2 tokens/parameter), so the curve measures the under-trained regime, not Chinchilla-optimal performance. See `SCALING.md`.
+- **sc300 on corpus v4 (local ablation, checkpoint not released)**: a controlled follow-up on the same corpus as sc5/sc27/sc80, 61 epochs (~1.28B positions), best epoch 8. It isolates scale at fixed mix: validation is flat vs sc80 (0.6538 vs 0.6563) while transfer improves (**+3.3pp zero-shot, +1.1pp FT**). The mix remains the stronger lever: sc80 on rmix14 reaches 0.4695 FT, still the best single model. See the data-mix table below.
 - `sc300` is a **negative result** on the default v5 mix: trained on corpus v5 (2,500 synthetic shards + 18 real logs), it did not improve transfer over sc80. Its validation accuracy (0.6080) is measured on a different, synthetic-heavy validation set and is **not comparable** with sc5/sc27/sc80 (v4). Zero-shot on BPIC 2015-1 is 0.3090 (0.2869 without unseen-to-UNK) and fine-tuned test accuracy 0.4267, both below sc80. The **data-mix ablation** below shows the bottleneck is the real fraction: sc300 on rmix14 reaches 0.3969 zero-shot / 0.4448 fine-tuned, but still does not beat the 80M model on the same mix (0.4695). See `SCALING.md`.
 
 ## Quickstart
@@ -67,7 +70,7 @@ The models use a **global vocabulary**: activities are ids 4..776, resources 2..
 
 ## Scaling
 
-`SCALING.md` contains the protocol, token budgets, caveats and the full table. Headline: validation next-activity accuracy rises 0.6026 → 0.6451 → 0.6563 across 6.3M → 27.4M → 88.2M under the same under-budget v4 recipe, while transfer improves mainly in the larger models (zero-shot 0.2977 → 0.3068 → 0.3600). The 306M point (trained on corpus v5) does **not** continue the trend: it is a negative result for scale, suggesting the data mix is the bottleneck. The data-mix ablation confirms it: rmix14 at 80M reaches **0.4695** fine-tuned, and 300M on the same mix does **not** beat it (0.4448).
+`SCALING.md` contains the protocol, token budgets, caveats and the full table. Headline: on corpus v4, validation next-activity accuracy rises 0.6026 → 0.6451 → 0.6563 across 6.3M → 27.4M → 88.2M and then flattens at 306M (0.6538, best epoch 8), while zero-shot transfer keeps improving slightly (0.2977 → 0.3068 → 0.3600 → 0.3929). The scale effect at fixed mix is small; the **real-data fraction is the dominant lever**: rmix14 at 80M reaches **0.4695** fine-tuned and 0.3969 zero-shot, the best single model. The 306M point on corpus v5 (0.3090 zero-shot / 0.4267 FT) was confounded by the synthetic-heavy mix; see `SCALING.md`.
 
 ## Data-mix ablation (corpus v5)
 
@@ -81,8 +84,9 @@ The v5 corpus is 2,500 synthetic shards plus only 18 real logs. We ablated the r
 | sc80 v5 rmix24 | ×24 | 15.7% | 0.3654 | 0.4597 | 0.6423 | 0.1967 |
 | sc300 v5 | default | 0.7% | 0.3090 | 0.4267 | 0.6064 | 0.1932 |
 | sc300 v5 rmix14 | ×14 | 9.5% | 0.3969 | 0.4448 | 0.6075 | 0.1945 |
+| sc300 v4 (local) | v4 | 5.6% | **0.3929** | 0.4450 | 0.6206 | 0.1873 |
 
-**Reading**: the real fraction is the dominant lever. rmix14 recovers +3.6pp fine-tuned and +1.6pp zero-shot over sc80-v4; more real is not monotone (rmix24 is lower). At 300M the mix recovers +8.8pp zero-shot / +1.8pp fine-tuned over the default v5, but the 300M model still does not beat the 80M model on the same mix (FT 0.4448 vs 0.4695): at this under-budget regime, **scale does not add transfer**. The best single model is `sc80-v5-rmix14` (published here as an ablation checkpoint); the best overall result is its late fusion with `process-graph-fm/gfm-sc80` (test 0.4832, see process-graph-fm).
+**Reading**: the real fraction is the dominant lever. rmix14 recovers +3.6pp fine-tuned and +1.6pp zero-shot over sc80-v4; more real is not monotone (rmix24 is lower). At 300M the mix recovers +8.8pp zero-shot / +1.8pp fine-tuned over the default v5, but on rmix14 the 300M model still does not beat the 80M (FT 0.4448 vs 0.4695). At fixed v4 mix the 300M is +3.3pp zero-shot / +1.1pp FT over 80M (validation flat): the scale effect is small, the mix effect is larger. With the corpus fixed at v4, the 300M model beats the 80M by +3.3pp zero-shot / +1.1pp fine-tuned (validation flat), so the earlier v5 negative result was partly confounded by the mix. Still, at this budget the mix lever (+3.6pp FT at 80M) is larger than the scale lever (80M→300M at fixed v4). The best single model is `sc80-v5-rmix14` (published here as an ablation checkpoint); the best overall result is its late fusion with `process-graph-fm/gfm-sc80` (test 0.4832, see process-graph-fm).
 
 ## MoE ablation (archived)