Troubleshooting

processBinaryAiEngine is gated on the binary_ai_engine extension row. With the extension disabled, no engine is ever loaded into memory, no period rolls over, and the settlement hook finds nothing resident. The engine rows can sit at ACTIVE the whole time and look perfectly healthy.

Enable the extension under Admin → System → Extensions. The scheduler picks it up within a cycle, without a restart.

Check binaryAiEngineEnabled and binaryAiEngineGlobalPause on the Settings page. Either one suppresses all steering platform-wide while leaving every engine row untouched.

The enabled switch also fails closed: if the settings store cannot be read at all, the addon behaves as disabled. If Redis or the settings cache is unhealthy, that is your answer.

simulationMode on the engine, or binaryAiEnginePracticeMode on the Settings page. Both are easy to leave on after testing. In simulation the audit log fills with SIMULATION_RUN rows, which is the confirmation you are in it.

Under Finance → Binary → Markets, the pair's source must be ECOSYSTEM. Markets created by the Import action come in as EXCHANGE and can never be steered — the tape belongs to the exchange.

You cannot flip a market's source to fix this if no ecosystem market exists for the pair; the create endpoint rejects ECOSYSTEM without one.

The attached AI Market Maker must be ACTIVE. Activation checks this, but restart recovery does not — so a reboot can revive an engine whose market maker has since stopped, leaving an ACTIVE engine that legitimately refuses to steer.

Start the market maker. The engine resumes on the next settlement.

The steered close has to become the close of the candle covering expiry. Once that minute has rolled over, the candle is history and a client may already have seen it, so the engine refuses and records how many minutes late it was.

This is a symptom of slow settlement, not of the engine. Look at cron health, database latency and how long the binary settlement job is taking. A rising share of this verdict is worth an alert.

The bucket was never analysed. Three usual causes:

• minPositionForOptimization is above your actual stakes. Default 10. Compare it to the median stake on the market.
• adjustmentLeadTimeSeconds is too short. Default 30. A bucket only becomes eligible inside that window; if settlement runs before a tick lands in it, nothing is recorded.
• The order is not RISE_FALL. Nothing else is ever steered.

maxPriceAdjustmentPercent is 0 on the engine. That genuinely means "do not steer" — a zero is honoured, not replaced by the default.

The steered close could not be written into the ecosystem candle series. Usually the market maker addon is not resolvable, or the ecosystem market is missing. The engine refuses to settle on a price nobody could see.

Most fields on the Settings page are defaults for new engines, read only at creation time. The nine that act immediately are: enabled, global pause, practice mode, min win rate, max win rate and the adjustment speed.

To change a running engine, edit the engine.

The five tier bonus fields there are the platform profile. The bonuses actually applied come from the per-engine tier rows under Users → User Tiers. Edit the tier row.

The controller works against the current period, and it only intervenes when the realised rate leaves target ± variance. Two things to check:

• Reset the period after a material change, otherwise the engine spends the rest of the window correcting toward the old number.
• Look at the variance. At ±5% around a 35% target, anything between 30% and 40% is WITHIN_RANGE and no outcome is forced at all.

Active-cooldown lookups are cached for 30 seconds, tier lookups for 60. Give it a minute.

Two clamps apply on load. maxPriceAdjustmentPercent is re-capped at 0.01 no matter what the column says, and the effective target win rate is clamped into the global min/max band on every decision. Both are deliberate.

Returned as a 400 on activation. Start the attached market maker first. This is the same condition the settlement path checks, so activation and steering agree.

The relationship is one-to-one, enforced by a unique index. Pick a different market maker, or edit the existing engine.

A user hit maxSingleOrderExposure on an ACTIVE engine's market. Raise the cap on the engine if that is a legitimate ticket size; otherwise it is doing its job.

Only one experiment runs per engine. If the blocking test has no durationDays, it runs until you stop it — check Analytics → A/B Testing.

You enabled monitoring without naming a provider, or named one outside BINANCE, COINGECKO, CRYPTOCOMPARE. The request is rejected rather than saving a configuration that would report enabled while permanently down.

The hard ceiling. 1% is already an order of magnitude wider than the 0.3% default; beyond it the nudge stops being a nudge and the published candle stops being credible.

Zero would make every period instantly overdue, so the controller resets before it can measure anything and win-rate targeting silently stops working. A row written by a direct database edit can still be in that state.

The database says one thing and the process another. Set it to PAUSED and back to ACTIVE — both operations work whether or not the engine is resident, so this is safe. If it stays detached, the cron is not running.

Restart recovery reloads every DB-ACTIVE engine on the cron's first successful cycle. If initialisation failed transiently, the reconcile pass on the next cycle recovers it. Give it a minute before intervening.

Global Pause on the Settings page is the fastest reversible option — engines stay loaded and lift instantly. Emergency Stop All is the deliberate, individually-acknowledged option: recovery requires reactivating each engine, which re-runs the market maker check.

Every engine update takes an automatic snapshot before it writes. Go to System → Snapshots, compare against current, and roll back. The rollback itself snapshots the state it replaces, so it is reversible too.

Note that status is deliberately not part of a snapshot — restore configuration, then set status yourself.

Normal on a young engine. The model needs enough settled live positions in the window. Nothing is broken and nothing needs chasing.

Three guards, any of which is enough to no-op: confidence below 0.5, the last training run's accuracy below 0.5, or the recommended change below the minimum step. Each returns a readable reason.

Apply Winner refuses on an inconclusive result. Either the arms have not reached the target sample size per arm, or the difference is not significant on the chosen primary metric. Let it run longer, or accept that the arms do not differ.

They report realised platform profit from position rows, not a projection based on an assumed payout. They will not match a back-of-envelope stake × payout calculation, and that is the point — payouts differ by order type and duration and are stamped on each order at placement.

Hourly buckets are in UTC, because position timestamps are stored in UTC. Offset them to your own timezone when you read the chart.

Monitors are started only for engines that are ACTIVE and have enableExternalCorrelation on. A paused or stopped engine has its monitor torn down, as does the global pause and the global disable.

Compare maxDeviationPercent to the volatility of your market. On a thin ecosystem market a 2% threshold will fire all day. Raise the threshold, or raise alertOnConsecutiveDeviations so a single spike does not alert.

A sustained HIGH or CRITICAL deviation on a market you steer means your published price has genuinely left the wider market behind. Pause the engine and investigate before you tune the threshold.