01
Describe
Write a setup in plain English (or drop a chart). Sentio parses it into a DSL spec — every step traceable back to your words.
Live research desk · ES · NQ · CME futures
Models that research themselves.
Describe a model in plain English. A multi-agent debate compiles it into a detector, replays it against thousands of past bars, and repairs it when it fails — every revision logged with the evidence that changed its mind.
Then it keeps going. Each setup it fires, winner or loser, is indexed and pulled back into the next debate — so the model that fires tomorrow has already sat through every session before it. A change only survives if the numbers improve.
The research behind the scenes.
A trader writes a setup in a sentence. Agents compile it into a detector and argue over it. The detector is replayed against real bars, every fire is scored against what price did next, and when it fails, the same agents read those outcomes and repair it — keeping only what measurably helped, and logging every change so the trader can follow and undo it. Each stage has a note that says exactly what it does, what it counts, and where it is weak. This is the program.
- Closed loop
Self-improving trading models
A failed replay hands the model back to the agents with every trigger's outcome. They diagnose, edit, and re-validate. A change is kept only if the numbers improved — including the case where the loop refused two repairs.
Note 03 · ReadIn production - Validation
Backtesting with AI, honestly
What counts as a trigger, how an outcome is scored, why an unfilled setup is not a loss, and how the P&L card sizes contracts so a one-tick stop cannot post a fantasy trade.
Note 02 · ReadIn production - Model builder
From a paragraph to a live scanner
Describe a setup in plain English. A multi-agent debate compiles it, catches the edge cases, and it cannot go live until replay validation passes.
Walkthrough · ReadIn production - Journal
Execution graded against the plan
Every logged trade is linked to the setup it came from and scored on what you did versus what the model said: chased entry, widened stop, exited early. Rules-versus-you on one curve.
Discipline layer · ReadIn production - Next
Prop-firm odds for your own model
Take a validated model's real trigger outcomes and play them through a funded-account ruleset — trailing drawdown, daily loss, payout gating — to get pass probability instead of a P&L guess.
In research
Model builder
Describe a setup. Agents debate it into shape.
Every trader has a setup that isn't on the shelf. Sentio's model builder takes plain English (or a screenshot) and hands it to a multi-agent debate — one agent makes the case, another hunts edge cases against it, another tightens the spec — before the DSL ever gets replay-validated and activated. You watch the argument unfold and every revision is kept on the record.
- Attach a chart image — Sentio AI reads the setup and fills in the description.
- Every step is argued for by name; nothing is invented in silence.
- Similar past triggered setups are retrieved from memory and shown as evidence.
- Replay validation must pass before the model can fire live.
The research loop Every model runs through the same loop
Every model runs through the same loop
before it earns a live trigger.
Describe a setup, let the agents debate it, replay it against real bars, and — if it fails — feed the outcomes back in for another round. Nothing goes live on faith; every model has already argued with itself and been tested against thousands of past sessions.
02
Debate
The agents argue the spec against itself — challenging edge cases, pulling similar past setups from memory, tightening the detector before it can go live.
03
Replay
One click sweeps ~3,000 bars of cached history, marks every trigger, and walks each one bar-by-bar so you see what it would have said.
04
Improve
Fails on the numbers? The same agents read every outcome, edit the spec, and re-validate — keeping only what measurably helped, and logging every change.
Three surfaces
Debate. Memory. Loop.
Every model on Sentio runs on the same three primitives: a debate that compiles the intent, a memory that remembers every setup, and a loop that repairs the model when it fails.
Debate
Sentio's agents argue every model against itself — challenging edge cases, pulling in evidence from past setups, tightening the spec before it can go live. You watch the argument unfold, transcript kept for every revision.
Memory
Every triggered setup — winner or loser — is embedded and indexed. The next debate retrieves the most similar past events as evidence, so the model that fires tomorrow has already learned from every session before it.
Loop
When a model fails replay, the same agents read every outcome, edit the spec, and re-validate. A change is kept only if the numbers improved — otherwise reverted. Up to three rounds, each fully logged.
Built-in models
Five specs. All inspectable. All yours to tune.
Every model ships with its rulebook, its parameters, its interpretation notes, and a replay engine. Zero black boxes.
Liquidity reversal
Forever
Liquidity manipulation, inverse FVG, delivery-state confirmation.
Breaker + FVGUnicorn
Sweep, displacement, breaker-block overlap with a fair-value gap.
SFP FVGSweep Failure
Failed sweep with rejection wick; entry off the inversion FVG.
Polarity flipInverse FVG
Parent gap invalidated; retest of the polarity-flipped zone.
Impulse retracement50 of 50
Large impulse move; 50%-retracement entry inside a tight zone.
Model builderCustom
Describe a setup in plain English. Sentio AI compiles it into a spec.
Closed-loop learning
Every model tracks itself. Losers auto-mute.
Every triggered setup is scored against the actual bars that followed. Win rate, average R, best and worst outcomes — all tallied per model, per timeframe, per symbol. When a model starts bleeding, the scanner shuts it up before you lose more.
Real outcomes, not backtests.
Every fire is scored on the price you would have gotten: entry, stop, T1, T2 walked bar-by-bar. A 50-of-50 T1 win scores 1.25R; a Forever T2 win scores whatever its geometry was worth. No fudged numbers.
Losing models go quiet for the session.
Three consecutive losses, or a win rate below 40% over six or more setups today, and the scanner mutes that model until the next session. Only reduces exposure — never enables anything you wouldn't have seen otherwise.
Fails validation? It repairs itself, on the record.
When a custom model fails replay, Sentio reads every trigger's outcome, edits the spec, and re-validates it on the same history — keeping the change only if the numbers improved, reverting if not. Up to three rounds, each logged with what changed and why, and any kept round restorable in one click.
Discipline layer
The scanner won't fire if your rules say no.
Every trader knows their own worst habits. Sentio makes those habits expensive. Write your discipline rules in plain English — daily loss cap, time-window mutes, size reduction after two losses in a row — and the scanner enforces them at the setup level. If a rule is tripped, the setup is gated and the reason is displayed. Nothing sneaks through.
Per-model rules"Mute Unicorn between 3pm and 6pm ET."
Session guardrails"Stop scanning after $500 realized loss."
Behavioural gates"Halve size after 2 losses in a row."
Cooldowns"After a loss, wait one hour before the next trade."
Not advice · by design We don't send you signals.
We don't send you signals.
Here's why — and why it's better for you.
Under US futures law, whoever pushes personalised entry/stop/target calls to your phone must be registered with the CFTC as a Commodity Trading Advisor. Most Discord groups skip it. Sentio takes a different path.
✕Signal groups & Discord calls
- "Long NQ 21,838, stop 21,821, T1 21,870" — pushed to your phone
- The signal seller does the thinking, you react without knowing why
- Requires CFTC / NFA registration — most groups ignore this
- No audit trail. No reasoning. No recourse when it fails.
✓The Sentio way
- Pattern match with the reasoning that produced it, on rules you author
- You size. You decide. You execute. Sentio never places an order.
- CFTC Rule 4.14(a)(9) — the publisher exemption, applied cleanly
- Every rule, every check, every parameter — inspectable + replayable
Get the second opinion. Keep your P&L.
Start free. No card required.