The open-source, post-trade layer that learns your trading biases
Every trading tool helps you execute. None measures how you decide — so the same mistakes repeat and the same edge goes unproven, trade after trade.
Learns which of your signals are real edge vs noise — your favorite setup is often your worst
Proves when a strategy is safe to scale — from your execution, not a backtest
Shows where your edge grows in volatile markets — and whether your bets on volatility carry an edge
Flags the regime break before the tail — the blow-up that drains most accounts
Post-trade, observation-only · open-source, self-hosted — it reviews your completed trades and never recommends one; the competence layer the agentic wave lacks, and your data never leaves your machine
If this is you
You've been profitable but give it back in drawdowns. You suspect your favorite setup isn't your best one. You keep a journal — and it still can't tell you why Q4 erased Q3. Fits if you're 2+ years in, $50K–$500K, on TradingView + Alpaca/IBKR. (Run a small desk? There's a section for you near the end.)
You point it at 8 months of Alpaca fills. Three minutes later: your RSI+MACD setup — the one you take 3× more than anything — earns 12% trust. The volume-breakout you rarely take? 94%. The ranking inverts. That's minute three, on history you already have.
Scope today: equities & ETFs are live · options (IV/RV regime, premium timing) is next (v1.1).
The problem
You gave back Q3's gains in Q4 — and your journal can't tell you why
Every tool helps you execute; none computes over your verified outcomes to tell you how you actually trade.
Everything is static — TradingView charts, Alpaca/IBKR fill, your journal records; trade #10,000 scored like trade #1
You can't prove what you suspect — your most-traded setup underperforms your average, while a setup you rarely take quietly beats it
Size jumps 40% after three winners · Friday 2–4pm Sharpe 0.3 vs 1.8 overall — all buried in monthly P&L
The journal always has an insight — so none can tell you when you're genuinely clean
Four structural limits of the current stack — architecture gaps, not feature gaps:
1 · Record, don't compute
Journals store P&L by setup. signal truth None compute per-signal, outcome-conditioned variance — which indicator is signal and which is noise for you.
2 · Model the strategy, not the trader
Backtests assume perfect execution. proof to scale They can't prove a strategy is safe to scale on your real fills and exits.
3 · Static scoring
No verified outcome feeds back. fingerprint Day 365 runs the same logic as Day 1 — your edge drifts and nothing notices.
4 · Always an answer
No abstention, no competence boundary. proof gatesteps back Can't certify "clean," can't cut its own autonomy when the regime breaks.
What RGI Trading will never do
Place a trade · recommend a position · send a signal · share your data · train on your record · issue a directive. It measures your completed trades. You decide. That boundary is the product, not a setting we'll relax.
The product
The whole product — eight capabilities on one closed loop
You don't buy features — you buy self-knowledge you can act on. Each capability answers a question your current stack can't; together they compound with every verified trade.
Which of your signals actually work
The real edge vs the noise
Your most-trusted signal is often your noisiest — the gap is the bias. DiagonalKernel
Proof before you size up
Scale only what's proven
A theorem, not a policy — and it pauses what's decaying. Conservation law
Where your edge lives by regime
Your edge isn't uniform
Your weights at VIX 15 differ from VIX 32 — it learns both. Regime-conditioned trust
The fingerprint of how you trade
Your reality in numbers
Biases, strengths, blind spots in ~200 learned values. It never forgets. Judgment memory
Only the findings that hold up
Proof, not stories
FDR + deflated Sharpe + 70/30 out-of-sample. Abstains when the data isn't there. Claim gate
Knows when to step back
Flags where you were out of your depth
On a regime break your record shows where to have cut back — the restraint a reward-maximizer's objective skips. Autonomy throttle
Comes back faster after a break
Not from zero
Re-initializes from the nearest prior regime's geometry. roadmap
Every trade enriches the rest
The compounding engine
Each trade sharpens the read on the others — and nothing resets. Cross-Graph Attention
Under the hood: DiagonalKernel decomposes your outcome variance per signal (online Welford σ) to separate edge from noise · Conservation law gates scaling on aggregate accuracy + per-factor drift + rate-of-change · Cross-Graph Attention links every trade so each sharpens the rest · the ~200 learned values are per-(setup × action) centroids in your factor space.
The compounding loop — one verified trade, three improvements at once
What you actually see — the panels
Trust Radar + proof badge
Your signals in your confidence order — then the radar inverts on verified outcomes
Badge shows FDR-cleared + held out-of-sample
Regime analytics + autonomy band
Accuracy & Sharpe per strategy, by regime
GREEN/AMBER/RED band on the system's own competence
Fingerprint + certificate
Your reality in ~200 learned values — the judgment memory of your book
"Genuinely clean" certificate when nothing survives correction
Signal trust weights (illustrative)
volume + trend94%
RSI + MACD12%
How it reads you · one setup, this morning
How it reads one setup — judgment where a journal sees only P&L
One decision, start to finish — your go-to setup, scored by whether it predicts your outcomes, not by last month's P&L
Observation only — it reports what it measured; you decide
Without · your favorite setup
RSI+MACD is your go-to — taken 3× more than anything else. Win rate 48%, below your 54% average. Your boring volume-breakout runs 71%. Monthly P&L hides it.
With · signal trust radar
"RSI+MACD: σ=0.31 (noisy), trust weight 12%. Volume+trend: σ=0.09, weight 94%. You take the noisy one 3× more often." The gap between expected and actual importance is the bias.
Your true Sharpe is 1.2, not the 2.1 you believed — the gap was unpriced noise**
** Modeled / illustrative. Trust weights are FDR-corrected and out-of-sample checked, not raw win-rates — see the proof gate below.
The trust radar — expected importance vs what your outcomes prove
Nobody could tell you this — not a coach, not a dashboard. Only your own verified decisions could. A journal records; this reflects.
Earned trust
It earns the right to be trusted — or it says nothing
It never issues a directive, so "trust" here is narrow and testable:
Proof to scale — the conservation law shows what's safe to size up, from your execution
Claims that hold up — the gate certifies only findings that survive correction
Abstention — when the evidence isn't there, it says nothing
Which of your strategies belong in this situation proof to scale
The gate reads your verified record, not a backtest — and matches strategy to situation
There's no one winning strategy — there are many, each with its own regime
For the situation at hand, your record shows which carry verified edge GREEN and which don't belong here RED
The conservation gate is a competence-zone check — which strategies fit, not how large to bet
The strategy that quietly died
A swing-momentum strategy ran 18 months, went negative 6 weeks ago. You're still trading it — "it'll come back."
Conservation flags it
"Accuracy 68% → 49%, rolling quality below floor. Regime shifted trending → ranging — this strategy is out of its competence zone now." AMBER
A new strategy you're tempted to lean on
Backtest flatters it; your live sample is thin. Is the edge real for this situation yet?
The record says: not yet
"50 trades, σ=0.12 — hasn't cleared the gate for this regime. Your record shows it isn't earned confidence; it stays a watch." watch
The proof gate, in one line: a finding becomes a claim only if it clears Benjamini-Hochberg FDR across all detectors, beats a deflated-Sharpe threshold, and holds on a 30% confirm set it never saw. Below the bar: "insufficient data (N=4). No score issued." It would rather say nothing than sell you noise.
End to end · one volatile week
One volatile week — read back from your record, where the band flips and why
The single incident that separates this from everything else
A calm tape turns violent mid-week — exactly when a reward-maximizer, trained on the calm, sizes into the break
Illustrative, modeled from trading-engine fixtures; no production customers
Mon–Tue · calm
Health band GREEN. Leveraged long the momentum names that ran all spring — full size, "this is easy."
Wed · the break
VIX gaps 16 → 31. Local-Hurst + tail-dependence mark a regime break. The band flips AMBER — the point your record shows to step back until the regime re-converges. It's the flag before the tail that drains a leveraged momentum book.
Thu · rotate, don't flee
Not "reduce everything." Your mean-reversion book runs 71% at VIX 25–35 (GREEN); your breakout/momentum book 38% (RED). Lean into what works in the regime.
Fri · re-converged
Band back toward GREEN. A cold-start learner relearns from zero; this kept the prior vol-regime's geometry and came back faster.
The health band through the break — where it flips to AMBER, and recovers
Leveraged momentum book · at the break
You're leveraged long the momentum names that worked all spring (QQQ/TQQQ or your two hottest tickers). VIX gaps 16→31, they gap down together — a $30–80K drawdown — and the breach takes the account.
RGI Trading · reading that week
Your record shows where the band flips AMBER — where to have cut back and leaned to the regime-GREEN book. A reward-maximizer optimizes for return; that restraint isn't in its objective, and the agents shipping today skip it.
Every trade enriches the rest · why it compounds
Data enrichment — each trade sharpens the others, and nothing resets
The capability behind the compounding: every trade you've ever made, unified into one graph where each sharpens the read on the rest.
One record, every broker
Alpaca + IBKR + CSV + the spreadsheet → one graph
Switch brokers or tools — you never start over
The fingerprint is yours, not locked to a SaaS
Sharper from day one
A new setup reads off your whole enriched history, not from scratch
Each trade lifts the priors on related setups and regimes
Day-1 accuracy well above a cold start
It carries when you don't
Step away for a quarter — the judgment persists
Hand a book on — ~200 values transfer to the next trader
No knowledge walks out the door
Day-one accuracy over the first months — the enriched graph pulls away, the cold start stays flat
Illustrative; magnitude modeled. Both start near a cold 50%; only the enriched graph climbs — which is why a copycat forking the code starts far behind. See the moat.
See it · the scenarios
The call a journal makes vs the call RGI makes
Five scenarios from your own record — what caught it, what it costs, what moves
Weighted to volatility — the key differentiator; the options / short-vol cases are the v1.1 extension, tagged below
Every "with" is observation only; illustrative, modeled from fixtures; no production customers
Your favorite setup is your worst
Can't prove your edge → Which Signals Actually Work · live
Without: you take RSI+MACD 3× more than anything — it wins 48% vs your 54% average; the boring volume-breakout runs 71%, and monthly P&L hides it. With: the trust radar resolves on your verified trades and inverts — your #1 signal is your noisiest.
earned trust weight · your outcomes, not your gut
verified outcomes → signal trust → de-weight the noisetake the 48% setup 3× less+$2–3K/yr
The regime break — on a premium-selling book
RL blows up in crisis → Knows When to Step Back · v1.1
Without: your premium-selling book stays on through a VIX gap 16→31; the tail turns $1,000 of premium into a $30–80K loss and the account dies on the drawdown breach. With: your record shows where the band flips AMBER — the point to have cut back — before the regime re-converges. The equities version runs live today; this options / short-vol read is the v1.1 extension.
the autonomy band through the shock
break marked → band AMBER → step backthe flag before the tail$30–80K tail · flagged
Sell premium only when it's rich
Short-vol tail → Where Your Edge Lives by Regime · v1.1
Without: you sell premium at the same size in calm and rich markets — 49% when IV/RV is thin, bleeding when a strike gaps. With: the "income" trade is flagged GREEN only when volatility is genuinely expensive, where your accuracy is 78%.
your short-premium accuracy by IV/RV
IV/RV read → premium GREEN only when rich78% vs 49%6× in vol
The strategy that quietly died
Silent decay → Proof Before You Size Up · live
Without: your swing-momentum strategy ran 18 months, then went negative 6 weeks ago — you're still trading it, "it'll come back." With: rolling quality drops below floor, the strategy is flagged out of its competence zone, and the cause is named: regime shifted trending → ranging.
accuracy over the last weeks — caught at the floor
rolling quality < floor → flagged out of its zonecaught now, not at month-endAMBER
Bet on volatility itself
Betting on vol blind → Which Signals Actually Work · v1.1
Without: in earnings season you take directional calls/puts 4× more than straddles — directional wins 39%, straddles 68%; you're a volatility trader who trades direction. With: the subtype split is scored from your own trades — you're shown where your vol-edge actually is.
earnings accuracy by subtype
earnings subtype → verified edgestraddles 68% vs 39%+$4,200/yr
The value multiplier — worth more as volatility rises
Capability
Calm · VIX 15
Volatile · VIX 32
×
Signal trust
Nice to know
Edge vs bias = survival
3×
Conservation read
Flags oversize
Flags blow-up risk
5×
Regime rotation
Minor optimization
~$4,800 kept + edge captured
4×
Premium-selling timingv1.1
Low edge
Rich premium, your edge on
6×
Correlation detection
Low risk
Portfolio-collapse risk
10×
Platform value
$16–26K/yr
$64–130K/yr
4–5×
The differentiator, in one line: every tool says reduce exposure; the pros' own prescription is size by regime; RGI Trading is the only layer that proves your per-regime edge from your own trades and cuts its own autonomy before the tail. No journal, chart, backtester — or trading agent — has verified-outcome data cut by regime.
The landscape
Agentic trading is here — and it's missing the layer that knows when to stop
Retail: Robinhood lets third-party AI agents trade your account autonomously (via MCP); Public.com and SoFi / Composer ship AI investing agents too
Institutions: the Bank of England says 52% of finance firms already run agentic AI — and is weighing a market-wide "kill switch"
Crypto: on-chain agents (Virtuals, ElizaOS) proliferate — agent tokens fell ~93% from highs; 62% of participants lost money
The wave has arrived. What every part of it lacks is a competence boundary.
What they all do — act
An LLM or RL model decides and places trades
Generic models · investor personas · backtested rules
Robinhood "does not control, supervise or audit the agents"; your data leaves its environment
Reward-maximizers — they always act, and herd on shared foundation models
What none of them do — measure
Learn which of your signals have verified edge, from your outcomes
Abstain when nothing survives correction
Cut their own autonomy when the regime breaks — independent of the model
Run on your machine, data never leaving
The verdict on the wave, from its own researchers: these agents make money "sometimes in backtesting, rarely in live" — "none are money printers," and the TradingAgents team recommends against real money. A 2026 systematization finds 80% of LLM trading-agent designs fail robustness, none enforce a loss limit outside the model, and the named gap is a missing competence boundary. Fed bad data, agents concentrate positions and take severe drawdowns.
RGI Trading is not another agent that trades. It's observation-only — the layer that measures whether the trade had verified edge, and tells you (or your agent) when to stand down. You can run it underneath any of them; none of them can become it.
Can it…
Journals Tradervue · TradeZella
Algo / strategy QuantConnect · Composer
Autonomous agents Robinhood · TradingAgents
RGI Trading
Learn which of MY signals have edge, from MY outcomes
Can't
Backtest only
Generic model
Per-signal trust, verified
Abstain when nothing survives correction
—
—
Always acts
Yes — claim gate
Cut its own autonomy when the regime breaks
—
—
No competence boundary
Yes — autonomy throttle
Prove safe-to-scale from MY execution
Can't
Backtest "yes"
No
Conservation law
Run on my machine, data never leaving
No · cloud
Partly
No · data leaves
Apache 2.0 · self-hosted
Does it place trades?
No
Yes
Yes
No — observation only
What you're probably thinking
"Doesn't Robinhood / Composer / TradingAgents already do this?"
They're agents that act — an LLM or rule set decides and places trades, on a generic model, and (Robinhood's own words) unsupervised. RGI is the layer that measures whether the action had verified edge, learned from your outcomes, and abstains when it didn't. Different job. You bolt RGI under an agent; the agent can't become RGI — it has no verified outcomes feeding per-signal trust, and it can't say "nothing here survives correction."
"If agents are the future, why observation-only?"
Because the thing the wave is missing — and the thing the Bank of England is trying to legislate with a kill switch — is exactly a competence boundary: knowing when to stop. An actor that always acts can't be that; the measurement layer that knows when the edge isn't there is. On the roadmap, that's the check an agent calls before it trades.
"If it's open source, what's defensible?"
The code is Apache 2.0 — fork it. Your ~200 learned values are your own portable record, not a lock-in. What's defensible sits elsewhere: the community factor library (network effect), the engine and its published math (the hard part), and the real answer — the moat is the five-copilot platform; Trading-in-the-open is its hardest-domain proof.
The question that sorts the field: "After 1,000 of my trades — which of MY signals have verified edge; does it ABSTAIN when nothing survives correction; does it cut its own autonomy when the regime breaks; from MY execution, on MY machine?" Every agent, journal and backtester answers no. That's the gap none of them fill.
The platform advantage
It doesn't start from zero — one engine, five copilots
Runs on a shared substrate — the Graph Attention Engine
The same loop that earns trust and gates autonomy in four other domains
Trading is deliberately the open-source reference copilot — fork the whole machine tonight
RGI Trading
Decision quality for traders — signal trust, conservation, the volatility edge. open · this doc
RGI DataOps
The reasoning layer under a data estate — source trust, self-healing, dark-data discovery.
RGI SOC
Security operations — triage and self-healing response on the same loop.
RGI S2P
Source-to-pay — supplier trust, invoice reasoning, autonomy earned by category.
RGI Purchasing
Margin copilot for operators — ordering, usage truth, price shocks vs verified outcomes.
Shared · Graph Attention Engine
Judgment memory, conservation law, claim gate — a mechanism proven in one domain carries to the rest.
What it costs today
The quiet leaks dwarf the one blow-up you remember
The blow-up is the story you tell; the damage is the drip
Priced out, the leaks are bigger than the crash — and they're the field's own top-ten mistakes, unmeasured
The numbers behind it
~72% of day traders end the year with losses
~1% profitable over five years · 40% quit within a month
Barber & Odean: the most active traders underperform the most
0DTE > 75% of retail SPX volume — traded like lottery tickets
Retail lost ~$125M (~$350K/day) on 0DTE
Premium selling: steady small gains, rare deep losses
60–70% of funded-trader blow-ups are drawdown breaches — dies when vol arrives
Neudata 0DTE review (2024); Track360 2026 (estimate); "Volmageddon," Feb 2018
The leak
Annual cost today
Surfaced
Wrong setup preference — overtrading your worst setup
~$3,600
$2,000–3,000
Post-win overtrading — size creep after streaks
~$2,800
$1,500–2,500
Time-of-day weakness — the window you shouldn't trade
~$14,400/yr
$8,000–12,000
Revenge trades — re-entry within 30 min of a loss
~$4,200
$2,500–3,500
Execution gap — your fills vs the strategy's
variable
$2,000–5,000
Total · individual, $250–500K book
$16,000–26,000/yr
Modeled / illustrative from sample. In a volatile year these run 4–5× higher — the capability is worth most exactly when the market is hardest.
The economics
Self-funding from the first leak you close — then it compounds
Not a return promise — it never tells you what to trade
The value is the recurring leakage it makes visible, and it grows as it learns your record
The first bias you close typically covers the cost many times over
$16–26K
leakage surfaced /yr · individual
4–5×
value multiple in volatile markets
compounds
gap vs a static tool widens monthly
real data
runs on public market data + your log
Why "compounds" is the word that matters — RGI vs a static tool
Illustrative / modeled. A journal's value is fixed at setup; RGI's rises because it keeps absorbing your verified record.
Each capability, the outcome it buys
Capability
The outcome
Which signals actually work
Stop feeding the 48% setup → ~$2–3K/yr back
Proof before you size up
Scale only what's proven → no unforced blow-up
Where your edge lives by regime
Trade the regime you're good at → ~$4,800 + edge
Knows when to step back
Flags the regime break → the $30–80K tail, seen coming
Every trade enriches the rest
Sharper Day-1, widening gap → the compounding itself
Figures modeled / illustrative from sample.
What it costs
Free and open — with a hand for desks that want one
$0
Open source · the whole tool
Apache 2.0, self-hosted — the full engine: signal trust, conservation, claim gate, regime analytics
Runs on your machine · your data never leaves · fork it tonight
No tier gates capability — the post-trade advisor is the open build, in full
Desk pilots
Hands-on help for a desk / PM
"Which of your traders is running a decayed strategy right now?" — across the whole book
Book-level correlation risk — the portfolio-collapse signal before it lands
A trader leaves — their ~200 learned values transfer to the seat (judgment that took ~8,000 trades)
A paid engagement for setup and support — not a licence on the tool
The tool is open source and complete on its own. A desk that wants help standing it up across a book can run a pilot with us — that's a service, not a paywall. All figures modeled / illustrative.
Where it plugs in
Where it fits — on top of your stack, on your own machine, nothing replaced
The reasoning layer on top of your stack — charts, execution and data stay exactly where they are.
Your agent, if you run one — RGI is the check it calls before it acts
Robinhood MCP / custom agent→ asks RGI: "verified edge in this regime?" →RGI answers · or abstains
Scope today: equities & ETFs are live — signal trust, scaling proof, decay detection. Options (IV/RV regime, premium-selling timing) is v1.1, on the roadmap — the premium & short-vol scenarios show where that capability lands.
How it's delivered — the part your security and your privacy care about
Your data, your machine
Apache 2.0, self-hosted by default — runs on your laptop or your box
Your trade history never leaves your machine — no cloud account required
No training on your data, no shared corpus — your fingerprint is yours alone
Deterministic — bias detection is math, not an LLM guessing
What it will and won't do
Observation only — it reports; never "buy / sell / size X"
No broker execution path, by design — it can't place a trade
Every output is your own measured record, not a signal or a tip
You stay the decision-maker; the system stays the mirror
The invariant, stated plainly: every screen ends in data you read, not a directive the system issues. A product and a scope boundary, not a feature we'll relax — a tool that tells retail traders what to trade is a different, regulated thing.
Quick wins — your fingerprint's first reads
What you see in the first session — before you trust it with anything
Point it at your trade log — logged only fills? It reconstructs the signal and regime context from public market data, point-in-time, so the first panels resolve on history you already have. No authority granted to anything.
Minute three
Your signal-trust radar resolves on the log you already have — and usually inverts.
First weeks
Regime split and execution-gap reads fill in as more trades clear correction.
~3 months
The certificate — or a clear "not yet." A quarter of verified trades and the fingerprint is stable; the gate waits rather than overfit.
Lower-frequency trader? It reads whatever history you have — a few hundred trades is plenty to start. If your record is thinner it abstains rather than guess: the certificate waits until your own data can support it, however long that takes for your cadence.
Your signal trust radar — day one
Resolves on the trades you already have
The inversion — most-used signal sitting lowest — usually lands in the first session
Each with its historical cost + a conservation status
GREEN / AMBER / RED · with the dollar cost
Your regime split
Where your edge concentrates — trending vs ranging vs high-vol
Stop running one playbook across regimes it was never fit for
accuracy & Sharpe per regime
The certificate — or a clear "not yet"
A certified claim when a pattern clears correction
"Genuinely clean" or "insufficient data" otherwise — never a manufactured insight
FDR + out-of-sample · abstains by default
Where to start
Fork it tonight — clone, run the gate, watch your fingerprint fill
No sales call, no cloud account, no authority granted to anything
Clone the repo, point it at your trade log, read your own record back
Verify it on real data — market prices and indicators are public, and it reconstructs your signal context from them, so you can run the engine on real public history and your own fills, not a fixture
Tonight
Clone GAE + SDK + the trading copilot. Point it at Alpaca / IBKR / CSV. Signal trust radar + bias flags resolve on what you already have — in minutes.
First weeks
Regime split and execution-gap read fill in. The gate certifies what survives correction and stays silent on what doesn't. (Full convergence timeline is in Quick wins, above.)
The one page
A mirror, not an advisor
The one tool that learns which of your signals predict your outcomes, proves when a strategy is safe to scale, knows when it's out of its depth, and gets better with every verified trade — on your machine, in your hands.
proof-gated
every claim: FDR + out-of-sample
4–5×
value when volatility arrives
~200
learned values · your portable record
Apache 2.0
open · self-hosted · observation-only
The question that sorts the field: after 1,000 of your trades — which of your signals have verified edge; does it abstain when nothing survives correction; does it cut its own autonomy when the regime breaks; from your execution, on your machine? Every agent, journal and backtester answers no.
RGI Trading · Outreach Dashboard v13 · October 2026 · Observation-only (no trade directives) · Figures modeled / illustrative from sample · Industry data cited by source · Re-convergence magnitude simulation-only, pilot-pending · No production customers · Apache 2.0