RYS-JEV

RYS-JEV is a layer-duplicated startlux-models/StartLux-Decision-v1.1-27B-Instant.

Blocks 41–48 execute twice (72 blocks total), and the residual update of both copies is scaled by α = 1.25. The duplication and α are baked into the weights, so the checkpoint is a plain 72-layer Qwen3.5 hybrid stack — no runtime layer map is needed.

Engine, materializer and /v1/systemone server: https://github.com/dnhkng/rys-jev

Results

Jev-Bench suite (17 sources, 17,773 test questions, hash split 8,932 development / 8,841 held-out); one-pass letter read-out, no fine-tuning:

model	dev	held-out	Δ vs base	W/L	p
base 27B Instant	80.52	80.40	—	—	—
RYS-JEV (41–49, α=1.25)	80.86	80.99	+0.59 pp	157/105	0.0016
this checkpoint, dev re-run	80.83	—	—	—	—

Held-out per-source gains: civil_comments +2.21, yelp5 +2.07, measuring_hate_speech +1.43, chaosnli +1.10, strategyqa_closed +0.85, mnli +0.41, fever_evidence +0.40; helpsteer2_verbosity −0.41 and helpsteer2_helpfulness −0.97. The checkpoint matches the virtual duplication engine to 0.03 pp and 40/40 sampled argmaxes. The gain is benchmark-specific: on a separate synthetic suite it was +0.06 pp (p=0.88).

Run

git clone https://github.com/dnhkng/rys-jev.git && cd rys-jev && pip install -e .
hf download dnhkng/RYS-JEV --local-dir models/RYS-JEV
rys-jev serve --model models/RYS-JEV --port 8090

curl -s localhost:8090/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": {"ticket": "I was charged twice for order #4411 and the app still shows it as unpaid."},
  "questions": {
    "team": {"type": "choice", "instructions": "Which team should handle this ticket?",
             "criteria": {"billing": "Payments and refunds", "shipping": "Delivery", "technical": "App and login"}},
    "urgent": {"type": "noul", "instructions": "Should this ticket be answered today?"},
    "severity": {"type": "score", "instructions": "How severe is the impact?",
                 "criteria": ["cosmetic", "annoying", "blocks the customer"]}
  }
}'

The model keeps the base model's bidirectional prompt attention in its full-attention layers. Engines that read the prompt causally serve a different model; the engine above applies the non-causal mask and self-checks it against the untouched forward, padding isolation and late-token reach.

Reproduce the checkpoint

rys-jev materialize --src <base-64-layer> --dst models/RYS-JEV --span 41-49 --alpha 1.25

A block's residual update is linear in its output projections, so scaling both copies' self_attn.o_proj (or linear_attn.out_proj) and mlp.down_proj by α is exactly h ← prev + α·(h − prev) on both executions.

Verification

• Materialized vs virtual engine: 40/40 sampled argmax agreement, max probability difference 0.041 (bf16 weight rounding), dev accuracy 80.83% vs 80.86%.
• Package: 12 unit tests pass; the /v1/systemone server, the CLI and the materializer were exercised end to end on this checkpoint and on a synthetic 4-layer fixture (bit-exact duplication and scaling).

Files

model-0000*-of-00016.safetensors (16 shards, BF16, 61.7 GB), config.json (72 layers, rys metadata), tokenizer and chat template, decision_config.json (letter token ids, per-type temperatures), vision tower and MTP head unchanged from the base.

License and attribution

Weights: CC BY-NC 4.0, inherited from the base model — non-commercial use only. Commercial use requires a license from StartLux Labs. This is an unofficial research derivative, not affiliated with or endorsed by StartLux Labs. Engine code: Apache-2.0 (see the GitHub repository's NOTICE).

Base model and technical report: Zhuang et al., StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions, arXiv:2610.17309.