Paper data archive¶
The data behind the AAMAS paper (pltlf-paper) lives in two identical copies on the DoC filesystems, and nowhere else:
/vol/bitbucket/oja24/_archive/pltlf/aamas26/exports/: the archive. ItsREADME.mdis the authoritative description of layout, provenance and verification;_meta/holds the scripts, file lists, checksums (sha256.txt) and deletion lists. The_archivetree only takes additions with the user's express permission./vol/gpudata/oja24-pltlf/exports/: the working copy, reached as~/Projects/pltlf/exportson cpucluster and gpucluster. It holds exactly the archived file set (2,028 files, 7.6 GB). The workspace expires on 2027-08-03.
Both were reconciled on 2026-10-10. Nothing remains on the laptop or CX3.
What counts as paper data¶
- The 99 curated profiles of the eight suites: every learner family as base,
-CERandhc-variants, plus the three Nash-Q variants on Stag Hunt. Each profile directory keeps everything the tables, the learning curves and the integrity audit read:runs.npy,checkpoints/, primary and confirmatory satisfaction reports, NashConv reports,reevaluation*.json,partial_resume_manifest.json,timings*.jsonandlegacy_evaluation/. Rendered plots and*.lockfiles are not kept. paper_figures/: only the 48*_tightened.pngpanels the paper includes, matched by checksum to pltlf-paper'sfigures/experiments/.scripts/regenerate_paper_figures.pyrecreates the other variants and the data links it mirrors into each suite.integrity_audit_final3.json, the final audit (job 12011).
The archive also keeps run provenance under logs/. That covers 191 executed
notebooks of the paper's training and re-evaluation runs, selected from each
notebook's printed resolved config. It also holds the five surviving Slurm logs
and the 14 August campaign launch records.
Deleted on 2026-10-10¶
These were deleted with the user's approval, after the archive verified by checksum:
exports/o/andexports/optimisations/, the screening runs behind the optimisation notes. Their outcomes survive only in those notes.stale-before-09458fb/, the pre-re-sync Harvest and Territory runs.- Earlier integrity audits, suite and profile renders, and unused
paper_figuresvariants. - Every executed notebook in the cpucluster and CX3 checkouts.
New experiments write into the same exports/ tree. Keep paper data
distinguishable from it, and extend the archive only with the user's
permission.