The AI is only as good as what we show it. That sounds simple, but it was not simple enough in the first version.
We kept asking narrow questions. Is this ticker good? Should we analyze this one? Does this chart matter? But traders do not think like that, at least good ones do not. They look at the market, compare situations, and ask what is moving, what is early, what is fake, what is crowded, and what is worth time.
So we built Brain Triage.
What Did Not Work
A strict scanner alone is too dumb. It can find movement, volume, and activity, but it does not know if attention is being wasted.
Example:
activity = relative volume * absolute price change %
This is useful, but it is not enough. A stock can be active because it is clean, because it already made the whole move, because it is a trap, or because the sector is moving and this one is just the loud cousin.
Numbers help us see the room. They do not tell us who to talk to.
What Worked Better
We changed the question. Instead of asking "should we trade this ticker?", we ask:
which market situations deserve deeper attention?
That is Brain Triage. It sees a slate of candidates and returns:
investigate
watch
ignore
cash
That is all. Important line: triage cannot trade, it can only spend attention. The AI gets a meeting, not the company credit card.
The Input Has To Be Small And Sharp
Dumping everything into the LLM is lazy and expensive. The triage input should look like a trader note:
LEADER source=market_mover price=211.40 rel_vol=1.6x
vwap_dist=+0.31% ema9>ema21 mover_rank=2
detector=VWAP Reclaim side=long score=87
That line says a lot: where it came from, why it is moving, what the structure might be, and whether it is close to something actionable.
Then the AI can answer in a useful way:
{
"ticker": "LEADER",
"action": "investigate",
"playbook": "Qualified VWAP Reclaim",
"direction_bias": "long",
"reason": "Top mover with VWAP reclaim and EMA support.",
"missing_evidence": "Needs pullback hold before execution.",
"risk_flags": ["chasing if extended"]
}
This is better than "yes/no." It gives a reason, a missing piece, and a danger.
The Cost Control Part
This also helps with LLM cost. We do not need the expensive brain looking deeply at everything, so the current shape is intentionally capped:
BRAIN_TRIAGE_MAX_ITEMS=12
BRAIN_TRIAGE_MAX_INVESTIGATIONS=3
The system can look broadly, then think deeply only where it matters. This matters because "AI reads the whole market every minute" sounds cool until the invoice arrives and punches you in the face.
Small joke. Also not a joke.
Where The Brain Belongs
Brain Triage is not here to replace rules. It is here to put intelligence in the right place.
Rules are good for spread, stop distance, reward/risk, daily loss, stale data, and position limits. AI is good for context, comparison, novelty, market feel, catalyst quality, and asking whether something is worth attention.
Those are different jobs. When we mix them, the system becomes weird. When we separate them, it starts to look like a desk.
The Full Flow
The system now moves like this:
Market Radar sees broadly.
Brain Triage chooses attention.
Full Analysis creates plans.
Triggers wait for price.
Risk gates block bad execution.
Broker fills tell truth.
Replay teaches what worked.
That is the vision. Not a generic scanner, not blind AI, not a hardcoded robot pretending to be smart. A system with a brain, a notebook, and a bouncer at the door.
Verdict
Brain Triage works as an architecture. It does not prove edge by itself, but it does something more basic: it stops Ladder from thinking about the wrong things all day.
That is a real improvement, because in intraday trading, attention is part of the edge.