Methodology
How the desk grades itself and where a couple of numbers come from that surprise people. The full reference lives in the repo: docs/METHODOLOGY.md.
Calibration — when is a call graded?
ot reflect grade --days 5
A call is graded once it is ≥ N sessions old (default N = 5, matching the debate's 5-day swing horizon). A verdict made Monday is first gradeable the following Monday. Grading is idempotent — a call is scored once, then carries its outcome forever.
Is the CALL/PUT for day 0 or day +1?
It depends on the horizon the call was made at, carried on every row as
time_stop_days:
- 0DTE (
ot decide --dte 0) — a same-day call, graded on the day's move. - Swing (
--dte 5, the debate default) — entered from the next session and graded at ≥5 sessions.
The desk never mixes the two: each row is graded on its own horizon. "Time stop" is the point at which being flat-but-right becomes wrong.
Two honesty rules
- Seed rows are excluded. The journal ships with a few bootstrap examples
(
source: "seed") so the desk isn't empty on day one — they never count toward the track record or the lessons. - It grades the desk, not your P&L. The table is the desk's own call history (debate + engine reads), currently global — not split by book or account. An "A-share" row is a China name the desk read, not anyone's realized trade.
Calibration-weighted fusion — where the judge trusts a proven analyst more — waits for ~30 real graded calls before it kicks in. Until then, single-digit cells are anecdotes, not statistics.
Dealer gamma — why two "call walls"?
The same name can show two call walls because they're computed over different DTE windows:
| Where | Window | Question it answers |
|---|---|---|
| Macro & Flow card | ≤ 7 DTE | this week's pin risk |
| Ticker ladder / chart | ≤ 30 DTE | the swing-horizon wall |
Both are correct — they're answering different questions. The chart labels its
walls ≤30D so it's unambiguous.
Retention
The decision journal is append-only and kept (the outcome is what changes).
The news archive under data/news-log/ is what ages out — not the journal.