It thinks like a five-year-old child. Curious enough to try. Primitive enough to be cautious. Sometimes it gets burned — and when it gets burned, it remembers what not to do. The longer it lives, the less it touches the hot stove. The more it watches, the more it gets right.
These are facts from a real run, not promises about your account. We publish them because they're what we have. We're still validating.
One run is a sample of one. We are still running variance and structural-cognition validation. Sign up below to receive the full report when it completes.
Most "AI trading bots" are a webhook and a prompt. Cognitive Trader is a full agentic architecture — built so the LLM never decides direction, only learns from outcomes.
No ACT fires without signal alignment.
The rules engine computes
market_direction via
sign(24h_change_pct) (threshold 1.5%)
while the LLM provides
thesis_confirmation. If
trading_goal (0.50) isn't paired with
market momentum (0.70), the Go-side gate blocks
execution. Defense in depth.
W1: Manual fact extraction on trade
close (source_channel='trade_outcome').
W2: Vector similarity retrieval via
Ollama nomic-embed-text. The agent
retrieves the 5 most-similar past trades into the
prompt context. It doesn't just "remember"—it
performs local RAG on its own historical scars.
A hard-coded circuit breaker in the
cognitive_cycle. If 24h PnL (excluding
pnl_unknown) breaches
-$500, the system enters lockdown. It
automatically cancels orphan
STOP_MARKET and
TAKE_PROFIT algo orders via the Binance
API. No human cleanup required.
The ActionThreshold is dynamic (range
0.35–0.85). Every cycle, the
lifecycle engine runs
CalibrateFromOutcomes, adjusting
weights based on a 30-day rolling PnL window.
Realized profits tighten the threshold; the agent
earns its right to be aggressive.
At 23:00 daily, the agent processes the day's
internal_monologue and
paper_trades. It generates
dream_insights that are injected into
the next day's working memory. This is where
"curiosity-driven" parameters like
intrinsic_motivation are calibrated
offline.
In volatile regimes, anchoring on stale prices is
fatal. If all prices in the
market_stats feed are >45s old, the
system enters Manage-Only mode. It
will close or adjust existing positions but strictly
blocks new opens. Real-world safety for live
futures.
There is a lot of dishonest marketing in algo trading. We'd rather lose the sale than oversell. Here's the straight version.
Most bots forget. Every cycle starts from zero — same indicators, same heuristics, same mistakes.
Cognitive Trader writes a short lesson at the end of every closed position: what the regime looked like, what the agent thought, what actually happened. Those lessons are indexed and fed back as context on the next decision. The bot doesn't just trade — it builds an internal book of what worked here last time.
This is the layer worth paying for.
Shorted ETHUSDT into a ranging regime on a "break of structure" that wasn't. ADX was below 18 the whole session. Don't trade trend-continuation logic when the classifier flags RANGING — wait for ADX > 22 or sit out. Next time this pattern shows in RANGING, skip.
Three things: the binary that does the work, the supporting stack it talks to, and the parts you bring yourself.
STAY_SILENT vs
ACT decision cadence.
-$500.
nomic-embed-text embeddings and
local model fallback.
Backtest results alone don't prove a trading system works. There's a hierarchy of evidence — reproducibility, cognitive falsifiability, out-of-sample, paper-pilot, live. Most fintech SaaS markets from the top of the ladder. We tell you exactly where we stand on each rung.
The system splits cognitive judgement (LLM) from execution
discipline (Go). Both halves were stress-tested independently and
the same architecture was validated with TWO different LLM
backends. Headline numbers from a 3-window × N=3 × 100-bar sweep
on BTCUSDT 4h with the LLM cache disabled
(NO_LLM_CACHE=1):
0.74% of capital; Config B at
0.19%. Both within an order of magnitude of zero.+$94.80
(3.6× the deterministic E3-rule baseline). Config B returned
+$77.52 (2.9× E3). Both turn the losing window
positive.+$21.88 vs +$21.49 per window) but
redistributes the alpha into a better risk profile
(Config B's worst window is 74% smaller).
Bit-identical replay infrastructure verified at three bar tiers:
20, 40, and 60 bars. In each tier, Pass 1 (population, real LLM
calls, fresh cache) and Pass 2 (replay, PLAYBACK_MODE=1
STRICT_DETERMINISM=1, cache-only) produced identical
paper_trades tables and 100% cache hit rates
(61 HIT / 0 MISS at 60 bars). The
CachedLLM now errors loudly on a real cache miss
instead of silently re-filling.
Full 3000-bar replay is deferred — scale-dependent state leaks
(memory ordering when Ollama embeddings are unavailable, the
behavioral_policy.last_calibrated wall-clock
timestamp) remain to be audited. The capability exists; the
3000-bar artifact awaits one focused engineering pass.
Five 3000-bar BTC runs across the development period:
+$10, +$156, +$208,
-$24, +$184. Mean +$107,
four of five positive. These were single-window single-run
measurements before the multi-regime methodology was adopted.
The Tier 1 sweep above supersedes them as the headline
risk-adjusted evidence. Per-run history in the table below.
The cognitive claim is empirically falsifiable through three
ablations. E1 (memory-disabled): chop alpha
collapsed 68% (+$77.52 → +$24.54) and
hold-count on TRENDING regime dropped from 57 → 0 across the chop
window. E2 (random-memory): random synthetic
memory content paralyzes the system via veto over-fire
(0-6 trades per 100 bars vs ~10-18 with organic memory). Memory
is not only load-bearing but its content discipline is
what produces the chop alpha. E3 (hard-veto rule)
is the chop-alpha denominator already published in Tier 1
(Centaur LLM beats the rule by 2.9-3.6×). All three ablations are
complete.
Same Centaur architecture on 2025 BTCUSDT 4h bars never seen
during development (fetched from Binance public REST, 3000 bars
covering 2025-01 through 2026-05). Three windows × N=3:
bear (offset 200, B&H -$203.33) →
strategy -$13.87 = 93% drawdown
reduction out-of-sample. Flat (offset 1500) → strategy
slightly negative, consistent with no-edge regime. Bull
(offset 2800, B&H +$335.38) →
strategy +$160.92 (48% capture, no blowup). The
drawdown-defense claim generalizes outside the dev set.
Live ticks, real-time 4h bars, fresh DB, full Centaur
cognitive cycle driving the patched
internal/tools/binance.go (with -1007
ambiguous-error reconciliation). Gated on
cmd/sanity_trade runtime testnet validation
(CLI committed at b040b308; manual smoke-test
pending).
One real position, $50 or less. Tests real liquidity, counter-party reactions, and the consequences gap between paper and live.
Sustained live performance across varied regimes at non-trivial size. Never "complete" — markets change. What you trust at this tier is graduated confidence, not absolute proof.
| Run | Date | Bars | Trades | Win % | PF | Net P&L | Code |
|---|
Past performance does not predict future results. Sample sizes are small. Numbers above are the data we have, not promises about your account.
Seven white-paper-style topics documenting the structural differentiators — without exposing the competitive edge. Each is implemented and measurable.
Publishing every run including losing ones, every commit hash, every log file, and every trade ledger as a substitute for "trust me" claims. The trust ladder escalates from numbers to live track record.
Six independent sources of non-determinism — wall-clock timestamps, Go map randomness, unstable sorts, LLM temperature, pgvector tie-breaks, environment variability — all gated by a single env flag.
The TWO-KEY VETO BY MEMORY mechanism differs structurally from standard RAG. Memory has refusal authority over the LLM's choice, creating the primitive cognitive loop.
The shuffled-bars experiment: randomizing time-order with a fixed seed to distinguish real cognition from pattern matching. The cognitive thesis is false if shuffled matches chronological within ±$50.
A trade requires both goal-derived signal AND market-direction signal to co-fire. The system default is STAY_SILENT. Differentiates agentic from trigger-happy.
When orders fill at prices different from intent (gap, slippage, partial), stops and targets rescale relative to actual fill — preserving risk geometry, not the original price levels.
The action_threshold recalibrates each cycle from outcomes. You can measure it: overnight #1 logged calibrated action_threshold=0.46 outcomes_considered=48. Veto density grows 10.7× from early to late phase.
Early-adopter pricing — while we're still finding the bugs. Price goes up at 1.0.
Locks in $129/yr as long as you keep your subscription active. You're buying a seat on the build, not a finished product.
Software is provided as-is. You are responsible for your trades. Numbers above are hypothetical performance examples, not forecasts. Not financial advice.
The operating constraints are part of the product: local execution, explicit risk, and no promises about outcomes.
No payment, no commitment. We'll send the full validation report and let you know when access opens.
// we email once when the validation report is ready · unsubscribe anytime
or join the Discord — we're building in public, beta drops there first