> For the complete documentation index, see [llms.txt](https://jwliaomath.gitbook.io/cocofold2/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://jwliaomath.gitbook.io/cocofold2/reference/release_validation.md).

# Validation scope

## Recorded evidence (2026-09-14)

Real Protenix-v1 tests were completed in the development server environment, with additional installation and smoke testing in a fresh Linux environment. These GPU tests were run separately from CI. The fresh-installation checks did not repeat every earlier model test.

| Scope                                                        | Evidence                                                                                                                                  | Limit                                                                                                 |
| ------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
| Single structure, inputs, seeds, records, export and restart | B1–B5 accepted by the author; legacy output exceptions recorded during development                                                        | Legacy files missing R/t require explicit raw export; no bitwise cross-GPU claim                      |
| Small single-state case                                      | B6 smoke and 1000-particle 7ZDT/7ZD5 refine accepted; final structure acceptable to author                                                | Map-derived supervision, SNR=1; RMSD is manual, not an automatic pass threshold                       |
| Contextual 6ZBH 1+3 short two-GPU test                       | B7b real Protenix and weights: 39/39 checks passed                                                                                        | Short continuation/resume and fault checks, not a convergence benchmark                               |
| Contextual 6ZBH, full 3324-particle STAR                     | B8 first four complete epochs: 69/69 checks passed, 416 steps                                                                             | Original target was ten; later epochs and whole-job completion are outside acceptance                 |
| Fresh Linux installation                                     | B9 environment report and full public CPU gate passed, as reported by the author; real-weight 7ZDT/7ZD5 smoke report: 10/10 checks passed | Existing B6 inputs reused; no new prediction, long refinement or two-GPU test in this new environment |

The first fresh-environment smoke attempt failed while compiling `fast_layer_norm_cuda_v2` because the selected system GCC was too old. The author loaded `gnu/12.2.0`, selected GCC/G++ through `CC`/`CXX`, and reported a successful retry. The earlier CPU pass is retained separately; it was not rerun for the compiler-only retry. See [compiler setup](/cocofold2/getting-started/installation.md#prepare-the-compiler-in-the-job-environment).

The supplied smoke report confirms finite and updated latent parameters, frozen GMM amplitudes/widths, optimizer progress, complete-epoch checkpoint, matching CIF coordinates/identities and complete records. Its runtime fields report elapsed time 269.49 s, train-step time 9.36 s and peak memory 17,732,314,624 bytes. These are case-specific report values, not hardware requirements or a performance comparison. The fixed-particle mean FRC decreased from 0.28019 to 0.14819; this is report-only and provides no evidence of quality improvement in this short smoke run. Structural RMSD remains a manual check.

For B8 the author reported a structure within 2 Å after the third epoch. This is a manual structural assessment, not an automatically measured release result. The four-epoch audit reported mean distributed step time 4.10223 s and summed step time 1706.53 s on two NVIDIA A100-SXM4-80GB GPUs. A distributed step takes the maximum rank duration. These times exclude initialization, epoch export and file I/O; they are not whole-job wall time. Recorded peak allocated memory was 47965.59/59394.84 MiB for ranks 0/1; reserved memory was 78834/78780 MiB. These measurements describe this case and environment, not minimum hardware requirements or a performance guarantee.

Independent and Contextual paths share public controls but use different conditioning. The validation recorded here does not include a long real-model Independent-component run.

## Continuous validation and release information

Automated CPU tests and website checks run through [GitHub Actions](https://github.com/jwliaomath/CoCoFold2/actions). Real-model GPU tests and manual structural assessments are recorded separately above.

For subsequent changes, run the relevant checks and retain their reports. Documentation-only changes do not require repeating real-weight training. CPU tests using substitute decoders do not establish real-model behavior. Each new platform requires installation validation; see [installation](/cocofold2/getting-started/installation.md).

The manuscript DOI and software version are recorded in [CITATION.cff](https://github.com/jwliaomath/CoCoFold2/tree/main/CITATION.cff). Published versions are available on the [Releases page](https://github.com/jwliaomath/CoCoFold2/releases). The [project website](https://jwliaomath.github.io/CoCoFold2/) provides an overview and tutorials.

## Build the public candidate

```bash
python tools/build_public_release.py --output /path/to/new_delivery
```

The new directory contains `CoCoFold2/`, `CoCoFold2-public.zip` and a SHA-256 `manifest.json`. The builder copies only `tests/public_files.json`, verifies the archive bytes and refuses an existing destination. It does not use Git, upload anything or include local results, weights or private experiments.

The root `.gitignore` follows the same explicit public inventory. Update both when intentionally adding a public file. An ignore rule does not remove files already tracked by Git or erase private data in existing history. Review and publish the independent candidate; do not push the research working directory on the assumption that `.gitignore` alone makes it public-safe.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://jwliaomath.gitbook.io/cocofold2/reference/release_validation.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.