TorchCraft: The shared four-stage optimisation schedule
Optimisation schedule of TorchCraft
The sequence logits pass through up to four stages: an optional soft warm-up using softmax probabilities, a linearly scheduled interpolation from raw logits to probabilities, temperature annealing from an initial to a minimum temperature, and straight-through hard optimisation on a one-hot representation with gradients taken through the low-temperature softmax. Minibinder and VHH runs used temperature annealing followed by straight-through optimisation; cyclic-peptide runs used all four stages; small-molecule runs began with a ligand-free warm-up before the ligand was added and the interface weight was raised.
Descriptive fact transcribed from the pinned sources with its own locator.
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
6 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Claimed field optimisation_schedule Context-only references | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Version: arXiv:2609.19770v1, submitted 17 September 2026 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Source location Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Context-only references | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Version: arXiv:2609.19770v1, submitted 17 September 2026 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Claimed value The sequence logits pass through up to four stages: an optional soft warm-up using softmax probabilities, a linearly scheduled interpolation from raw logits to probabilities, temperature annealing from an initial to a minimum temperature, and straight-through hard optimisation on a one-hot representation with gradients taken through the low-temperature softmax. Minibinder and VHH runs used temperature annealing followed by straight-through optimisation; cyclic-peptide runs used all four stages; small-molecule runs began with a ligand-free warm-up before the ligand was added and the interface weight was raised. Context-only references | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Version: arXiv:2609.19770v1, submitted 17 September 2026 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Description Descriptive fact transcribed from the pinned sources with its own locator. Context-only references | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Version: arXiv:2609.19770v1, submitted 17 September 2026 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Relationship: subject torchcraft-20261115-pipeline-torchcraft Context-only references | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Version: arXiv:2609.19770v1, submitted 17 September 2026 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
| Name TorchCraft: The shared four-stage optimisation schedule Context-only references | TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1 Version: arXiv:2609.19770v1, submitted 17 September 2026 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: SHA-256 of the PDF as downloaded. The pdftotext -layout text layer every value was read from hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b; it is regenerated from the PDF and is not archived here. |
Sources and history
Release 2026-11-21-c8bea25e5abb · Record review: source checked
1 source record and release history
- TorchCraft: Unified binder design by inverting an all-atom structure predictor (arXiv:2609.19770v1) · Original source · arXiv:2609.19770v1, submitted 17 September 2026
Technical metadata and extraction receipts
Stable ID: torchcraft-20261115-claim-design-schedule
- areas
- protein-structure; molecular-interactions
- contexts
- research
- field
- optimisation_schedule
- review
- artifact sha256: f8145b5b95552dd30996eda2940507c0f32d09adfc60f44f28458e7872788f21; method: transcription; automated-source-review; source-hash-verification; method note: Every quotation and locator on this record was re-read in the refetched artifact named by retrieval_url. All seven pinned artifacts were re-retrieved by the reviewer on 2026-11-16T00:08Z. Six came back byte-identical to the hash on their source record. The Hugging Face page did not, and cannot: three consecutive fetches gave three different hashes. Normalising its per-request hf-sanitized CSS class hashes leaves one further per-request difference, window.requestId, and the page also prints a download counter; with those normalised the markup is identical. Its licence tag, gate text and dataset card matched the recorded quotations word for word, so the page is a per-retrieval stamp, not a changed page. The pdftotext -layout text layer of the refetched PDF hashes to 90f1da5f8a936766087beba0ade3f038f5b69504b9f581b42f2391f5640c5e6b, the value the extractor recorded.; retrieval url: https://arxiv.org/pdf/2609.19770v1; reviewed at: 2026-11-16T01:05:00Z; reviewer: claude; reviewer note: Claude (Opus 5) review agent, independent of the extractor; no human review claimed. Every pinned artifact was re-retrieved and hashed, and every stored value re-read against it. Nothing was reproduced: no design run, score, docking calculation or binding measurement was repeated, and no value was recomputed from source data.
- source locator
- Appendix A.1.2 and A.1.3; Algorithm 1; Supplementary Table S1
- value
- The sequence logits pass through up to four stages: an optional soft warm-up using softmax probabilities, a linearly scheduled interpolation from raw logits to probabilities, temperature annealing from an initial to a minimum temperature, and straight-through hard optimisation on a one-hot representation with gradients taken through the low-temperature softmax. Minibinder and VHH runs used temperature annealing followed by straight-through optimisation; cyclic-peptide runs used all four stages; small-molecule runs began with a ligand-free warm-up before the ligand was added and the interface weight was raised.
Related records
- subject: TorchCraft