ciaochris/Vers3Dynamics-Civil-Reactor-Expert-3B

feat: add optional Jev decision layer to the grounded serving path

Add usage examples for civil Q&A, classification, and summarization

Enhance model card with HF widgets, grounded tool docs, CI coverage fixes, and Docker deployment

feat: ground numeric answers in the physics engines and measure correctness

Update README.md

docs: fold 9B generalization caveat into existing callout instead of a separate box

feat: scale dataset to 697 examples and eval to 180, fix decay-heat physics bug

eval: independent (non-template) benchmark decomposes format-fit vs real generalization

fix+eval: intrinsic safety eval surfaces overfitting; ship better 3-epoch adapter

docs: record the converged-variant RQ2 result in the technical report

feat: add converged Qwen3-30B-A3B civil-reactor variant (70% perplexity reduction)

feat: reproducible dataset expansion + Tinker fine-tuning and benchmark

eval: run and archive the first safety evaluation; close report TODOs

fix: close newline bypass in safety scanner and correct PWR coolant Cp

fix: implement missing xenon module, correct physics bugs, clean repo-wide lint

fix: measure over-refusal, restore model-card CI, remove fabricated benchmark numbers

docs: add evidence-backed technical report with evidence-gap audit

Update README.md

feat: add comprehensive engineering infrastructure and documentation

Update README.md

Update README.md

Upload test_xenon.py

Upload 3 files

Update README.md

feat: add reactor-type specific Python modules

Close the training and endpoint evaluation loop

Update README.md

Harden the civil reactor expert release for auditability

Publish a real hardware-bounded civil reactor adapter

Use a real 3B-class trainable base

Prevent impossible local GLM-5 training runs

Prepare real LoRA improvement path

Make safe discovery data generation reusable

Keep generated Python cache out of the model repo

Reframe nuclear expertise as civil reactor discovery

initial commit