project-pandora

isolated runner scripts and model-specific glue for synthesis planning.

why this repo exists

these scripts used to live inside project-procrustes, which is the home of RetroCast itself: canonical schemas, adapters, scoring, analysis, and benchmark tooling.

that arrangement worked, but it blurred two different concerns:

• project-procrustes is evaluation and normalization infrastructure
• project-pandora is the execution-plane junk drawer for planner-specific runtimes

the split exists for a few reasons:

• many planner runners are independent of RetroCast as a product, even when they still reuse some RetroCast helper code
• runner environments pull in old, research-grade, or model-specific dependency stacks that we do not want contaminating the main retrocast repo
• keeping these dependencies out of procrustes reduces noisy security alerts and keeps the core repo focused on the canonicalization/evaluation layer
• Daedalus-style execution via isolated directories makes it more natural to treat runners as their own external runtime surface

relationship to retrocast

RetroCast docs live at retrocast.ischemist.com. RetroCast still defines the data conventions most scripts here follow:

• benchmark definitions
• stock files
• raw result layout
• manifests and result serialization
• adapter expectations for downstream ingestion/scoring

so this repo is split out physically, but not fully severed conceptually. some scripts here still import RetroCast helpers for benchmark io, manifests, and serialization.

what lives here

project-pandora holds model-facing runtime code, including:

• planner wrappers and execution scripts
• asset/bootstrap helpers
• one-off preprocessing steps tied to specific models
• per-runner locked environments where needed
• shared RetroCast-style runtime assets under data/retrocast

currently that includes:

• aizynthfinder
• askcos
• directmultistep
• dreamretroer
• multistepttl
• retrochimera
• retrostar
• synllama
• syntheseus
• runtime/synplanner

what does not live here

things that remain in project-procrustes:

• retrocast library code
• adapters and canonical schemas
• scoring/statistical analysis
• benchmark curation and preparation
• general comparison/reporting workflows

status

this repo is currently an extraction, not a total dependency divorce. expect some awkward transitional coupling while runner code is peeled away from shared RetroCast utilities.

data/retrocast/0-assets/model-configs is the shared transitional home for model configs. lightweight yaml/config files are tracked; large downloaded model payloads such as checkpoints, onnx files, hdf5 stocks, pickles, and generated outputs are ignored locally.