DAIRLab/phystwin_models

PhysTwin fitted models: DAIRLab lab sessions

Complete project-produced PhysTwin-style spring-mass fits of DAIRLab/phystwin_dataset lab sessions, one folder per recorded session and fit run, with every saved checkpoint.

License and attribution

MIT (owner decision D9) applies to the project-produced fit packages in this repository: native checkpoints, run records and the generated metadata. It does not relicense the linked DAIRLab/phystwin_dataset recordings, which declare no license, or any third-party code or data.

• The fits follow the PhysTwin method (Jianghanxiao/PhysTwin, MIT); keep its notice wherever its code is redistributed.

Runs

Location	Run	Stages	Status	Parent	BEST / FINAL	Budget (completed/configured)	Code revision
session_20260921_225630_786148/sync0402-0893	lab-a-faithful-photo	cma, photometric_adam	complete	lab-a-faithful-tracking adam/c0012	adam/c0495 / adam/c0500	cma: (completed_generations 100, issued_asks 800, told_asks 800, physical_failures 0); photometric_adam: traversals 500/500, updates 245500/245500	e80a85d1526e (clean state not recorded for every stage)
session_20260921_225630_786148/sync0402-0893	lab-a-faithful-tracking	cma, adam	complete	—	adam/c0489 / adam/c0500	cma: (completed_generations 250, issued_asks 2750, told_asks 2750, center_evaluations 1, physical_failures 0); adam: traversals 500/500, updates 245500/245500	2b53aa2040ed
session_20260921_225630_786148/sync0402-0893	lab-a-initial-20260923	cma, adam	complete	—	adam/c0500 / adam/c0500	cma: (completed_generations 100, issued_asks 1100, told_asks 1100, center_evaluations 1, physical_failures 0); adam: traversals 500/500, updates 245500/245500	0a0403f3a4d6
session_20260921_225630_786148/sync0402-0893	lab-a-rp7-p0-photo100	photometric_adam	complete_preliminary	lab-a-initial-20260923 adam/c0500	adam/c0099 / adam/c0100	photometric_adam: traversals 100/100, updates 49100/49100	4c29ed1e5da3 (clean state not recorded for every stage)
session_20260921_225630_786148/sync0402-0893	lab-a-rt05-cma250	cma, adam	complete	—	adam/c0500 / adam/c0500	cma: (completed_generations 250, issued_asks 2750, told_asks 2750, center_evaluations 1, physical_failures 0); adam: traversals 500/500, updates 245500/245500	07e620c32f40, aeea3a989482 (dirty reported)
session_20260921_225630_786148/sync0402-0893	lab-a-rt05-photo	cma, photometric_adam	complete	lab-a-rt05-cma250 adam/c0500	adam/c0246 / adam/c0500	cma: (completed_generations 100, issued_asks 800, told_asks 800, physical_failures 0); photometric_adam: traversals 500/500, updates 245500/245500	e7e8b0c6d9a5 (dirty reported)
session_20260921_225630_786148/sync0402-0893	lab-a-rt05-photo-proj-c0015	cma, photometric_adam	complete	lab-a-rt05-projadam adam/c0015	adam/c0497 / adam/c0500	cma: (completed_generations 100, issued_asks 800, told_asks 800, physical_failures 0); photometric_adam: traversals 500/500, updates 245500/245500	9f4942d3fdc3 (dirty reported)
session_20260921_225630_786148/sync0402-0893	lab-a-rt05-projadam	adam	complete	lab-a-rt05-cma250 cma/winner	adam/c0500 / adam/c0500	adam: traversals 500/500, updates 245500/245500	38edf32bfea9 (dirty reported)
session_20260921_230724_040591/sync0316-0630	lab-b-photo	cma, photometric_adam	complete	lab-b-tracking adam/c0123	adam/c0500 / adam/c0500	cma: (completed_generations 100, issued_asks 800, told_asks 800, physical_failures 0); photometric_adam: traversals 500/500, updates 156500/156500	8c03605eb6af (clean state not recorded for every stage)
session_20260921_230724_040591/sync0316-0630	lab-b-tracking	cma, adam	complete	—	adam/c0499 / adam/c0500	cma: (completed_generations 250, issued_asks 2750, told_asks 2750, center_evaluations 1, physical_failures 0); adam: traversals 500/500, updates 156500/156500	8c03605eb6af

What a run package contains

• native/: every saved checkpoint and state file of the run, byte for byte, with its run records and logs (CMA generations, candidates and strategy state; Adam snapshots, continuation state and per-update losses; or Deform360 boundary, optimizer, RNG and role records).
• evidence/: launch, parent-selection and qualification records that the run's records cite.
• run.json: stages, actual objectives, optimized and fixed parameters, configured versus completed budgets, labels with their scoring conditions, code revisions and dirty-state reports, capabilities and unknowns — each value names the native record it came from.
• checkpoints.tsv: one row per checkpoint file (stage, step, parent, labels, status, size, sha256).
• containers/: small native and evidence files packed by category into uncompressed tar containers of at most 1 GiB, bytes and logical paths unchanged (scaling plan §3.2, so a run stays within the Hub's file limits). containers/index.json lists each member's native path, offset, size and sha256 and the files kept individual; scripts/phystwin_publish.py fetch restores the native layout, by HTTP byte range for a single checkpoint.
• dependencies.json: model parents in this repository, dataset inputs to be pinned at a verified dataset revision, upstream archive members, and environment paths that are not package content.

Not claimed

• No model here was loaded, rendered, resumed or re-evaluated for this card; clean-destination use is a separate validated step (plan PR4).
• BEST and FINAL are the labels the fits recorded under their stated scoring conditions; packaging selects no preferred model.
• Interrupted-run recovery is qualified by each run's recorded execution identity and inputs; extending a completed run's horizon and full reproduction are not supported or claimed.
• Historical clamp-at-use stiffness runs are not reinterpreted as projected-update runs.