`sdsge-decompile`

sdsge-decompile extracts a .sdsge bundle into a directory. The output layout matches the sdsge-compile input convention, so the result can be edited and recompiled into an equivalent bundle.

sdsge-decompile <bundle> [-o OUTPUT] [--csv] [--force]

Arguments

Argument	Description
`bundle`	`.sdsge` file to extract.
`-o`, `--output`	Output directory. Defaults to `<bundle-stem>/` next to the file.
`--csv`	Re-encode Parquet members as CSV in the output (useful for editing).
`--force`	Overwrite the output directory if it already exists. Without this flag an existing directory raises `FileExistsError`.

Examples

Extract verbatim — Parquet members stay Parquet:

sdsge-decompile my-experiment.sdsge

Decompile for editing — Parquet members are re-encoded as CSV:

sdsge-decompile my-experiment.sdsge --csv -o my-experiment/

What lands where

The extractor writes each member at its compile-input authoring path, not at its archive path. For most kinds these match; the one exception is model_config:

Member kind	Archive path	Extracted path
`model_config` (role=reference)	`model/reference.yaml`	`reference.yaml`
`model_config` (role=dgp)	`model/dgp.yaml`	`dgp.yaml`
`estimation_*`	`estimation/...`	`estimation/...`
`mc_*`	`montecarlo/...`	`montecarlo/...`
`raw_data`	`data/...`	`data/...`
Simulation prefill (inline in manifest)	—	`simulation.json`

Options sidecars

Model Member.options (e.g. compile_kwargs/solve_kwargs) are extracted to <role>.options.json at the directory root. The compile entry point reads these sidecars on recompile, so the round-trip preserves them.

Inline simulation prefill

The SimSpec rides inline in the manifest (it has no member of its own). On decompile it is written to simulation.json at the root so the recompile picks it up via the standard directory layout.

`--csv` re-encoding

With --csv, every Parquet member is decoded to CSV. The extension and manifest.json format field are rewritten to reflect the new layout.

Member	CSV header strategy
`estimation_data`	Observable names from `Member.columns` (semantic headers — user-friendly).
`estimation_trace`	Mechanical `{name}.{j}` expansion from `trace_to_csv`.
`mc_trace`	Same mechanical expansion.
`raw_data`	Existing column names from the Parquet schema.

Checksum field on decompile

The manifest.json written by sdsge-decompile omits checksums. They were SHA-256 over the archive bytes; after extraction (and especially after --csv re-encoding) those bytes change. A subsequent sdsge-compile recomputes them. Treat the decompiled manifest.json as informational, not as a re-compile input — the compile pass auto-detects the layout.

Round-trip

compile → decompile [--csv] → edit → compile produces a bundle equivalent on every reconstructable component:

sdsge-compile my-experiment/                  # → my-experiment.sdsge
sdsge-decompile my-experiment.sdsge --csv     # → my-experiment/ (editable)
# ... edit reference.yaml, observed.csv, estimation/spec.json, etc.
sdsge-compile my-experiment/                  # → updated my-experiment.sdsge

What is preserved across the round-trip

Model YAML, estimation spec/result, observed data, MCMC posterior, MC pipeline/result, simulation prefill, model options (compile/solve kwargs). The created_by and created_at fields are regenerated on each compile.

sdsge-decompile