iLOVE2D/tbsciencepipeline

What this is. Scripts only. No source data, task assets or answer keys are included. Every rebuild needs the source data, fetched from its original provider under that provider's licence (listed per task below). Paths such as /path/to/scratch and /path/to/source_data are placeholders: set the environment variables named in brackets in the table, or edit the defaults. Data licences: the Gosai et al. 2024 MPRA table is CC BY-NC 4.0 (non-commercial). Replogle et al. 2022 is CC BY 4.0. The edge-thermal source case is private and not distributed; its pipeline is included for transparency only. Author: Tianyu Liu (tianyu.liu@yale.edu).

TB-Science 5-task data-synthesis pipeline

Each task is rebuilt from its public (or private, see edge) source by one SLURM driver:

sbatch <task>/run_pipeline.sh /path/to/<task>      # the Harbor task directory to (re)write

All stages are CPU work and run inside SLURM, never on a login node. Python is a pinned 3.11 venv (numpy 2.1.3, pandas 2.2.3, scipy 1.14.1, scikit-learn 1.5.2, rdkit 2024.3.5); override it with PY=.... Source paths are overridable by environment variables, shown in brackets below.

Every pipeline has the same shape:

source data -> stage (filter/split/reduce) -> agent asset (.tgz, packed deterministically, SHA256 pinned in Dockerfile)
            -> withheld truth (tests/expected.json) -> gates (thresholds) -> checks

The checks gate every build: oracle = 1, no-op = 0, every trivial or adversarial baseline = 0, and an independent implementation = 1 where the truth is a deterministic reduction. harbor_emulate.sh <task> <workdir> {nop|oracle} then replays the two-container Harbor flow with bubblewrap. The agent jail sees /app and /solution; the verifier jail sees only the tests image plus the declared artifacts.

Per task

task	source	stage	truth	checks
edge-thermal-qualification	frozen OpenFOAM conjugate-heat-transfer case phys_bench/reference_private/of_edge_compute_thermal_qualification_v1/source_case.tgz [EDGE_SOURCE]. Private, read-only, never shipped: only the derived case.tgz is in the task.	build_case.py: unpack, add acceptance_criteria.json, reduce fields with two independent implementations (independent_reduce.py), assert the tolerance policy, pack case.tgz	qualification verdicts and limiting quantities -> expected.json	local_check.sh
coldplate-flow-rebalance	phys_bench/harness/assets/of_battery_coldplate_release_review_v4.tgz [COLDPLATE_SOURCE]	build_all.sh -> build_cases.py: repackage mesh, 0/U, 1/U, system/ and requirements.json; built twice and compared with cmp	build_expected.py plus independent_reduce.py -> expected.json, calibration.json	local_check.sh. See authoring/VALIDITY.md: the fields are synthetic, not a converged CFD solution.
cis-regulatory-extrapolation	Gosai et al. 2024 MPRA (Zenodo 10.5281/zenodo.10698014, DATA-Table_S2, md5 a1b36bb5e270b6414e74f2a83ea47e5d) [CRE_SOURCE]	build_split.py: keep 200 bp ACGT rows with all 3 activities, drop chrY, deduplicate by sequence, hold out chr7 and chr13, sample 120,000 train and 10,000 test (seed 20261029)	build_expected.py with thresholds.json -> expected.json and the agent-visible criteria.json; pack_asset.sh packs and pins	local_check.sh: oracle, no-op, 6 trivial baselines, 4 adversarial-from-oracle submissions
perturbation-transfer	Replogle et al. 2022 essential-gene Perturb-seq, K562 and RPE1 h5ad (CC BY 4.0) [PT_K562, PT_RPE1]	build_pseudobulk.py: backed, chunked per-condition sums and sums of squares of log1p-normalised expression	build_asset.py: K562 is the source and RPE1 the target; split the target perturbations, build perturb_split.tgz and expected.json	local_check.sh (oracle, independent_solution.py, trivial baselines), reproducibility.sh (seeds)
orthogonal-rescoring-triage	AmpC Uni-Dock poses with PLIP interactions, from docking_source/dockingout_proteins_ampc (savedir/*.sdf, *_complex_dock.json, receptor) [ORT_WORK]; smina static build	rescore.py (and rescore.sbatch): add hydrogens with RDKit, rescore every pose with smina --score_only under vinardo, vina, dkoes and ad4	build_truth.py: deduplicate to 966 ligands by canonical SMILES, build screen.tgz, withhold size-adjusted Vinardo truth and gates	local_check.sh, sensitivity.py (gate-margin sensitivity), analyze.py. See authoring/VALIDITY.md.

Determinism

Assets are packed with tar --sort=name --owner=0 --group=0 --numeric-owner --mtime=<fixed> | gzip -n -9, so rebuilding from the same stage gives byte-identical files and the same SHA256 (the build checks this). Splits and models use fixed seeds recorded in each script.

Truth provenance

Truth is never the reference solution's output:

• cis, perturbation: held-out measurements;
• edge, coldplate: deterministic field reductions, agreed by two independent implementations;
• orthogonal: an independent rescoring function the agent does not get.

Each gate's tolerance is at least 3× the measured seed or implementation spread. The calibration numbers are in each task's authoring/.

Data rules

Only derived values are shipped:

• the private edge source case is never shipped;
• the raw Replogle h5ad and the MPRA table are not redistributed;
• the docking inputs are reduced to the poses and receptor needed for the task.

Also included, for synthesising new tasks on another machine

• briefs/: the build rules every task must meet. COMMON_BRIEF.md is the pass bar plus data and compute rules, HARD_BRIEF.md is the difficulty target, and the paperrep family design is in SCREEN.md and BUILD_BRIEF_*.md. Two negative design results are recorded so they are not repeated: calsel_DESIGN_FINDING.md and seqdes_MEASURED_RESULT.md.
• harness/stage.sh, harness/grade.sh: stage a solver sandbox exactly as the agent image would see it, and grade a sandbox with the task's own verifier. Staging checks the Dockerfile sha256 pin and that no label files leak.
• harness/shared/: shared metric and pass-bar tools used by the family generators.
• harness/opus_probe/: difficulty probe. Claude Code runs headless in a bubblewrap jail, one SLURM job per replicate.
• harness/kimi_probe/: script-only solver loop for an OpenAI-compatible endpoint. Set KIMI_BASE and KIMI_KEY. Its results are a lower bound, because the loop is weaker than a shell agent.
• harness/watch_probe.sh, harness/paperrep_collect.py: poll builders' READY.tsv, probe each new instance and apply the acceptance rule. A task is accepted if Opus passes 0/3, or if Opus passes at least 2/3 and Kimi passes 0/3.

All compute is meant to run through sbatch/srun.

Family generators (families/): how the 100+ task set was synthesised

Each family turns one real data source into many distinct Harbor task instances. Every instance meets the same bar: oracle 1, no-op 0, every trivial baseline 0. Instances are recorded in that family's registry.csv, with the measured rewards.

family	problem type	data source	data licence	registry (PASS / total)
admet	molecular property model search vs. leaderboard gate	TDC ADMET benchmarks	CC BY 4.0 (TDC); several upstream licences unspecified	17 / 21
hts	HTS hit triage and ranking under assay artefacts	Tox21 (NCATS), PubChem BioAssay (Butkiewicz)	public domain; curation CC BY 4.0	11 / 25
pharmome	dose-response curve fitting from raw wells	EVE Bio pharmome	CC BY-NC-SA 4.0	21 / 22
depmap	unseen-lineage CRISPR dependency transfer	DepMap 24Q4	CC BY 4.0	15 / 16
perturb	Perturb-seq held-out / transfer prediction	Replogle 2022; Norman 2019; Adamson 2016 (GEARS-processed)	CC BY 4.0; GEO entries without explicit licence	9 / 15
synergy	held-out-drug combination synergy	O'Neil 2016 via DeepSynergy	redistribution unclear	11 / 52
cfd	frozen-field CFD reductions and release decisions	OpenFOAM-lineage cases; FDA nozzle benchmark	GPL-3.0 lineage; US Gov work	4 / 5
docking	orthogonal rescoring triage	public PDB receptor, in-run docking	see registry	1 / 2
claim	verify a paper's/colleague's claims against data	Tox21, DepMap, Perturb-seq, O'Neil, pharmome	per dataset	5 / 5
audit2	find and repair one planted methodological defect	same datasets	per dataset	5 / 5
triage2	multi-screen triage of an unscreened library	Tox21	public domain	4 / 10
defect	planted split-leakage defect (prototype)	O'Neil synergy	redistribution unclear	prototype
hand-built	the 5 tasks in the top-level dirs + ADMET AMES	see above	see above	6 / 6

BLOCKED means the instance was built but measured to have no honest headroom. Those instances are not counted, and their gates were not weakened to make them pass. Answer keys (expected.json, truth.json), packed assets and blocked task dirs are not in this repo. Source data must be fetched from its provider.

Measured difficulty caveat. Claude Code with Opus 5.5 passed 64/66 probe runs on a sample of these families. So the families are good at producing valid tasks but not hard ones. For hard tasks, use briefs/HARD_BRIEF.md, briefs/SCREEN.md (paperrep) and the probe in harness/, and filter every built instance by measured solver pass rate. See DIVERSITY_ROADMAP.md for the problem types the current families do not cover.