Settings
These are the source of truth. A forward or supervised session reads them when it starts; command-line options override them for that one run. Saving here does not reach into a session that is already running.
Protection
The off-hours guard. Broker stops only exist during regular hours, so this is what watches a held position pre-market and after hours — a SOFTWARE stop, alive only while the runner is.
on
Whether the guard runs at all. ON is the shipped position: without it a held position is naked out of hours, because the resting stop at the broker does not execute outside the regular session.
Consecutive breaching minute bars required before the guard acts. ONE PRINT IS NOT A BREACH — an extended-hours book has no depth, and a single odd trade through the level is what that looks like.
How far below the triggering price the limit sits, as a fraction. The floor, not a target: too tight and nothing fills in a thin book, too wide and the exit gives away more than the gap it was avoiding.
Minutes before the regular open at which ownership returns to the broker stop. The handback must COMPLETE before 09:30, or the day limit and the re-armed stop are both live at the open — which is the double-sell.
Seconds between tape polls while the guard is watching. Minute bars are the input, so polling faster than they close buys nothing but vendor calls.
The market-data feed the guard will accept. ON A PARTIAL FEED, SILENCE IS INDISTINGUISHABLE FROM CALM — IEX carries a fraction of extended-hours prints, so the guard REFUSES TO ARM on it rather than reporting protection it cannot provide. Leave this on sip.
Supervision
Relaunch a crash, kill a hang, and stop rather than loop. A supervisor is a process too — a reboot or a killed supervisor leaves nothing running, and what survives that is the resting stop at the broker.
defaults Nothing has been saved for this section yet.
on
Whether supervision is wanted. With this off the supervise command REFUSES to start rather than quietly running the session unsupervised — run forward if that is what you want.
Restarts allowed inside the breaker window before the supervisor gives up and leaves positions to their resting stops. RAISING THIS IS HOW A SUPERVISOR BECOMES A CRASH LOOP.
seconds
The window the restart count is measured over, in seconds.
seconds
The wait before the first relaunch, in seconds.
What each successive backoff is multiplied by.
seconds
The ceiling on the backoff however many failures precede it, in seconds.
seconds
How long a running child may go without a heartbeat before it is presumed hung, in seconds. THE POINT OF A SUPERVISOR THAT IS NOT A PID CHECK: a wedged runner holds its lock and its positions and does nothing, which to a PID check looks healthy.
seconds
Grace after a launch before a missing heartbeat counts against the child, in seconds. Warm-up fetches several hundred symbols before the first beat; killing a runner that is still seeding guarantees the outage.
seconds
How often the supervisor looks at its child, in seconds.
off
Whether a clean exit is relaunched IMMEDIATELY, without waiting for the open. OFF, and it does not stop the lane: a live lane already starts the next day's session on its own, at the next market open. ON would relaunch the moment the session returned zero — spinning a new session against a shut market every few seconds, which is what the wait-for-open exists to avoid.
Risk
Portfolio limits and execution realism. A strategy decides whether to trade; these decide how much is at stake when it does.
defaults Nothing has been saved for this section yet.
Dollars committed per position.
Entries allowed in one session.
Positions open at once.
Session loss that halts further entries.
Total open-position value as a fraction of equity. 0 disables. A count cap cannot express this — 20 slots at $3,500 is 77% of a $90k account and 175% of a $45k one. LIVE ONLY: the replay has no equity model, so setting it makes live more conservative than its own backtest.
on
Whether a symbol may be re-entered after being closed the same day.
off
Whether a position may survive the session end. Off is the day-trading behaviour every existing lane was measured under; a multi-day lane cannot be expressed at all while it is off.
off
Whether scarce slots go to the strongest candidates rather than the first to arrive. Off means arrival order, and arrival order is alphabetical.
The protective stop applied when an entry decision does not name its own. A backstop, not the strategy's stop.
An entry filling worse than this from the decision price is abandoned, not taken at any price.
Adverse slippage assumed on every entry fill. THIS IS AN ASSUMPTION, NOT A MEASUREMENT — replace it from forward paper fills and re-baseline.
Strategy
Which lane a session runs, and what answers its entry gate.
defaults Nothing has been saved for this section yet.
ai-momentum is the intraday lane with the AI entry gate. trend holds for weeks, needs no catalyst store and no model key, and forces overnight holds and strength ranking on regardless of what Risk says.
mock runs the real prompt and the real parsing against the local endpoint at no cost. anthropic calls the model, and additionally needs a key and the per-run spend arming flag — selecting it here does not spend money.
Reasoning effort for the anthropic backend. Low because a per-bar gate is latency-bound and a slow verdict is a verdict about a bar that has gone.
Momentum lane
Entry gates, score floors and exits for the intraday momentum lane. Changing any of these changes which trades it takes, and invalidates its recorded results.
defaults Nothing has been saved for this section yet.
Multiple of today-anchored VWAP the price must clear. 1.001 = by 0.1%.
on
Whether price must also be above the prior session's close.
Ceiling on stochastic %K for the fresh-signal test — turning up from oversold, not turning up from a top.
VWAP deviation that qualifies a bar as extended-and-confirmed.
Relative volume that qualifies a bar as extended-and-confirmed.
Minimum time-of-day relative volume.
Minimum percent change against today's open.
Minimum share of session volume printed on up bars.
The score floor once the entry window is fully open.
The score floor before the entry window opens. Higher on purpose: an early bar has less session behind it to be judged on.
Eastern time the entry window starts opening.
Minutes over which the early floor ramps down to the open-window floor.
Eastern time after which no entry is taken at all.
Minimum regular-hours bars in the current session before a bar is evaluable.
Minimum bars of indicator warm-up history.
off
EXPERIMENT, off by default. Drop a bar in a Bollinger squeeze instead of asking the model about it. Arming this changes which trades the lane takes and invalidates every result recorded for it.
EXPERIMENT, off by default. Drop a session that opened more than this percent above its prior close. ZERO DISABLES IT.
on
Whether the momentum-exhaustion and jerk adjustments are applied to the normalised score. On in the old lane, so on here.
Score adjustment for momentum exhaustion: subtracted below 0.5, added above 1.5, nothing in between.
Score adjustment for jerk: added when positive, subtracted when negative.
Hard stop distance below the fill, as a fraction. Pinned so minimum, maximum and flat are the same number — A STOP THAT CAN VARY IS A STOP THAT CAN WIDEN.
Favourable excursion at which the trailing give-back arms.
Give-back from the peak once the trail is armed.
Favourable excursion at which the structural VWAP give-back exit arms.
Fixed profit target as a fraction. ZERO DISABLES IT, which is how it runs live.
Eastern time everything is flattened. This is a day-trading lane; nothing holds.
Trend lane
Lookback, trend filter and anti-chase guards for the multi-week trend lane. Few knobs on purpose — the effect is robust across lookbacks, and tuning it is fitting noise.
Showing shipped defaults — the stored settings are version 2, and this build writes version 1
Completed sessions in the trailing-return lookback. 60 is about twelve weeks, and is the longest a multi-week replay can warm up without spending the whole run filling the series.
Sessions in the trend filter's moving average.
Sessions in the moving average that triggers the EXIT, separate from the entry filter's. 0 = same as the entry MA. A LONGER exit MA lets a winner run further before a close-below exits it; typical 20-40.
Minimum trailing return to call the trend positive. ZERO IS THE DOCUMENTED RULE and should stay there unless something measured says otherwise.
Largest session-to-session close ratio the warm-up may contain before the symbol is refused as untrustworthy. 2.5 refuses an overnight move bigger than 2.5x either way - the signature of a corporate action the series was never adjusted for, not of volatility. A DATA GUARD, not a strategy filter. 0 switches it off, and off is how a lane comes to rank wreckage first.
Protective stop handed to the harness on entry, as a fraction. Wider than an intraday lane's because the hold is measured in days — a 2% stop on a multi-week position is noise.
How far up its own intraday range a name may be and still be bought. 0.60 refuses the top 40% of the day's range. THE ANTI-CHASE GUARD — the picker was right and every entry still gave the move straight back because it went in near the intraday top.
How far a name may already have run today and still be bought, as a fraction. The range test alone passes a name that gapped 20% and then went quiet.
Sessions held before the moving-average exit may fire. Entry happens because price is above the average, so on the entry day it sits close to it and one down bar would exit immediately.
on
off
off
off
off
off
Minimum Kaufman efficiency ratio for a name to be bought — below it the name is range-bound chop (an oscillator, not a trend) and is refused. ZERO IS OFF: the sign test is unchanged until this is set above 0. ~0.30 is a sensible floor. Buys only names that are clearly trending.
Completed closes in the efficiency-ratio lookback. Only consulted when the efficiency floor is set above 0.