From a config file to a logged score

Context

The study ends with a list of experiments (P1–P8). Each one is only worth something if it runs the same way as the others: same inputs, same scorer, same record. P0 is that common path, built on top of the Docker pipeline instead of beside it.

What we did

• A model is a config. configs/<name>.json names the module, its parameters and a seed. Adding a model means adding a module and a config, nothing else.
• One command, with or without Docker: generate → official validation → public evaluation → comparison with baseline_smooth50 → record.

python docker/pipeline.py modelo --config configs/baseline_smooth50.json --data-dir <local data> --repetir

• A written data contract: frame index of the tactical video, units, masks for missing data, and H/A/D/X out. Every module reads and writes the same thing.
• Each run gets its own folder with the effective configuration, its hash, the commit, the environment and the CSV hash, so any result on this log can be traced back and repeated.
• 17 tests on invented data: alignment, format, configuration, the four public references, a full synthetic run, errors and concurrent runs. CI runs them without match data.

Result

The pipeline reproduces the official baseline unchanged and is the entry point for P1 and for both candidates. A low score is a warning by default, and a hard failure only when asked (--exigir-referencia).

Next

A clean install by a second member, and the first experiment through it: P1.