Add [TimeBraid-6.7B](https://arxiv.org/abs/2609.29792) ([code](https://github.com/CharonWangg/TimeBraid)) to GIFT-Eval-CTX under `eval/timebraid_6p7b_demo_rewrite_v1/`, with **History + Macro + Micro** inputs. This run used the fixed 6.7B DSFP36 SFT-v2.8 checkpoint-28000 and one demo-style Context renderer for every row. We render the supplied effect JSON as natural-language future-step conditions and convert its level/amplitude coordinates to raw units using observed-history statistics. It did not use a forecast mixture or row-wise model selection.
| 78-group relative MASE ↓ | Overall | Short | Medium | Long |
| --- | ---: | ---: | ---: | ---: |
| TimeBraid-6.7B | 0.773281527744585 | 0.6847064982335485 | 0.8719738490515145 | 0.8640818145482951 |
Inference used BF16, FlashAttention 2, and TimeBraid's native preopened numeric forecast route. Missing observed-history values were filled by leading backfill followed by forward fill, using observed history only. The dataset was pinned to …
