compare_bartypes.py

Sentinel/Lab/compare_bartypes.py


Family	Lab (Python)
Version	—
Size	165 lines
Scope	public — ships in `sentinel-suite`
Documented by	no doc tracks this artifact

Rendered from the published copy in sentinel-suite/src/, not the author’s private tree — so this page describes the file you actually have.

What the file says about itself

Compare the Council's decisions ACROSS bar types on one instrument.

Answers the operator question "how much does bar type matter?" from the per-scope
excursion corpus a Market-Replay session produces (one JSONL per scope). It reuses
the Lab's own parsing/label logic (sentinel_lab.dataset + labels) so the numbers are
consistent with train.py -- this is a lens on the SAME corpus, not a second pipeline.

Three views:
  1. PER-BARTYPE -- fire count, cadence, direction balance, conviction, and the
                    first-touch WIN rate (the outcome that actually pays). eff_n is
                    the AFML concurrency-adjusted N -- trust it over the raw count.
  2. CO-FIRE     -- when two bar types fire within a tolerance window, do they AGREE
                    on direction? (decision consistency -- do they see the same call?)
  3. OVERLAP     -- how often do they fire at the same time at all? (do they even see
                    the same setups, or is each bar type its own market?)

Usage:
  python compare_bartypes.py --inst GC [--dir ../Excursions] [--tol-min 3]
                             [--since 2026-07-13] [--until 2026-07-14] [--horizon 15]

Reads only the corpus; never touches bin\\Custom.