Move reproducibility files to a private repo for author review before public release

reproducibility/vlm_prompts.md

@@ -1,105 +0,0 @@
-# VLM prompts used in the DeepJEB++ quality filter
-
-<!-- ============================================================================
- PUBLIC FILE - this is what gets uploaded to Hugging Face as reproducibility/vlm_prompts.md
- The annotated working copy, with provenance notes, is prompts_extracted.md in this same
- folder and is NOT for release.
-
- Cleared for upload: the scoring-criteria question was answered by Jinsu Ra on 2026-08-22.
- ============================================================================ -->
-
-Stage 1 of the pipeline screens candidate images with a vision-language model. The model is
-asked to describe a rendered bracket, and the description is scored against a fixed vocabulary
-of defect words; images whose score falls in the most defect-like top-*p*% are removed. This
-file gives the exact prompt strings behind every filter variant reported in the paper, so that
-the numbers in Table 2 can be reproduced.
-
-The strings are reproduced verbatim from the generation notebooks and from the scoring manifest
-on our lab server. Nothing here is retyped.
-
-## The adopted filter — LLaVA with negative-word negation (NWN)
-
-```
-Question: Do not use word "no". Describe the red object's surface roughness as details. Do not use "no". Answer:
-```
-
-This is the filter used to build the released dataset. It reaches **72.17%** accuracy on the
-labeled benchmark at the default threshold and **76.10%** at the tuned operating point
-*p* = 29% (false-positive rate 40.7%, false-negative rate 14.0%).
-
-The doubled negation is deliberate. Asking the model not to use the word "no" suppresses the
-negated descriptions ("no visible cracks", "not rough") that otherwise dominate the output and
-collide with the defect vocabulary, since a negated defect word and a present defect word score
-alike under embedding similarity.
-
-## The no-correction baseline — same backbone, no negation instruction
-
-```
-Could you describe this red object's surface condition as detail as possible and some Geometry Complexity? Answer:
-```
-
-This is Table 2's "Vanilla prompt (no NWN)" row at **56.45%**. The gap between this and the
-adopted prompt is the effect of the negation instruction alone: the backbone, the vocabulary and
-the scoring are identical.
-
-## The alternative backbone — BLIP with the same NWN prompt
-
-```
-Question: Do not use word "no". Describe the red object's surface roughness as details. Do not use "no". Answer:
-```
-
-Identical prompt text; only the vision-language backbone differs. This is Table 2's
-"BLIP + NWN prompt" row at **62.67%**, where the filter degenerates to near keep-all.
-
-## Which result file used which prompt
-
-| Reported result | Backbone | Prompt above | Vocabulary |
-|---|---|---|---|
-| 76.10% / 72.17% (adopted) | LLaVA | adopted NWN | `llm_keywords.txt` |
-| 56.45% | LLaVA | no-correction baseline | `llm_keywords.txt` |
-| 62.67% | BLIP | adopted NWN | `llm_keywords.txt` |
-
-All three scoring runs use the same 112-word vocabulary, released here as
-`reproducibility/llm_keywords.txt`.
-
-## Vocabulary
-
-`llm_keywords.txt` holds the **112** defect and manufacturability terms, one per line, that form
-the negative-word set. The description embedding is compared against each term and the *mean*
-cosine similarity is the image's defect score, so the file's length is also the dimension of the
-similarity vector referred to in the paper as the 112-dimensional negative-word similarities.
-
-The file has no trailing newline, so `wc -l` reports 111; the count is 112.
-
-## Scoring criteria
-
-The 112 defect words were generated by prompting a language model with seven engineering
-criteria. All three scoring runs use the same list:
-
-```
-Manufacturability
-Surface Clearence
-Surface Texture
-mono-body
-uni-body
-Smooth Surface
-Good Strength
-```
-
-These are reproduced exactly as the scoring configuration records them, including the spelling
-of "Clearence". Section 3.2.4 of the paper and Figure 3 list the same seven; both write
-"surface roughness" for the criterion the configuration labels "Surface Texture", which is the
-same criterion under two names.
-
-The criteria are the prompt given to the language model that produces the vocabulary. They are
-not applied to images as separate tests: the only quantity computed per image is the mean
-similarity to the 112 words.
-
-## A limitation of this release
-
-The prompts and the vocabulary are fully determined, and the diffusion sampler is deterministic.
-The selection of *which* images condition a given 3D generation is not recoverable: each design
-is conditioned on eight images drawn from its interpolation pair's retained pool, and that
-sampler was seeded with Python's string hash, which is randomized per process and was not fixed.
-The dataset itself is unaffected — the meshes and labels are what they are — but a bit-exact
-replay of the image-to-mesh step is not possible.