àyò

The premise

Most amateur trading bots are overfit backtests that their authors do not realize are overfit. ÀYÒ is built the other way around. Generating strategies is easy; the hard part is not being fooled by them. So the whole system is organized around a validation layer that a good looking but wrong strategy cannot pass. A language model proposes, and a deterministic gate disposes.

Àyò is the Yoruba seed and pit game of counting and calculated capture, and it also means joy. A fitting name for something that wins by playing a patient, measured game.

Architecture · two clocks

ÀYÒ runs on two clocks that never touch. The slow clock is the brain: Claude researches and writes each strategy as structured data, never as runnable code. The fast clock is the body: a deterministic engine that evaluates the frozen strategy against live prices, so the same inputs always produce the same orders.

Because the brain emits a content hashed artifact, you validate the artifact rather than the model, so a nondeterministic author yields perfectly reproducible behavior. The guarantee is blunt: remove Claude entirely and the body keeps trading safely on its frozen rules. It simply stops getting smarter.

The validation gate

A beautiful backtest is worthless. The engineering value sits in a gate that a plausible but wrong strategy cannot clear. Three layers do the work:

• Walk forward. Strategies are selected on in sample history and judged only on data they never influenced.
• Robustness sweeps. An edge has to survive across many rolling windows and parameter settings, not appear at one lucky split.
• Trials adjusted significance. The more ideas tested, the higher the out of sample bar, so the search cannot mine its way to a fluke.

A momentum book once looked like it beat a passive benchmark on a single split. The robustness sweep took it apart, and the gate refused to deploy it:

robustness sweep · cross sectional momentum
  rolling windows   beat benchmark in 2 / 5 periods
  parameter grid    positive edge in 0 / 12 settings
  verdict           FLUKE, no robust edge

An in sample Sharpe of 3.72 that becomes minus 6.99 out of sample is exactly the trap that empties accounts. The whole job of the system is to see it and say no.

When the scoreboard lied

Refusing to deploy a bad strategy is the easy half. The harder half is not being fooled by your own measurement, because that is where self deception is invisible and where everybody quietly cheats. In July the system was caught doing exactly that, in its own favour.

benchmark scoreboard · self audit
  symptom   the scoreboard was flattering itself
  cause     after hours benchmark prices up to 3% stale
            (769 logged for a market that never traded above ~748)
  effect    overstated the book against the market, all live window
  fix       benchmark sampled in session only, median anchors,
            recorded history sanitised
  result    a materially smaller, honest number. published as is.

The bug flattered the system, which is the only kind that survives a casual review. It was found by auditing a number that looked too good rather than enjoying it, and the corrected figure replaced it the same day. A platform that will not audit its own instruments has no business auditing its strategies.

Autonomous research

A headless Claude runs on the same machine as the trading daemon, so the brain is always reachable. On a schedule it proposes fresh strategies. Every proposal is parsed into typed rules, never executed as code, and forced through the same out of sample gate before anything can go live. Claude proposes, math decides, and real money always needs a human. Here is one cycle:

research cycle · brain proposes, gate decides
  REJECTED  Oversold Mean Reversion Snapback
            in-sample Sharpe 3.72  ->  out-of-sample -6.99
  REJECTED  Trend Momentum Breakout
            out-of-sample Sharpe 1.39  but drawdown 65%

  no proposal survived out of sample. nothing deployed.

Published research

Building the gate raised questions the literature had not settled, so the work produced two preprints. Both are about the same problem this system exists to solve: how do you keep a search process honest when a language model is doing the searching?

Under the hood

Honest status

ÀYÒ is not a proven money maker and it does not pretend to be. It trades paper only, and it has no validated alpha. Every simple price based strategy tested so far has failed honest validation, which is the expected and correct result on efficient large cap markets.

What ÀYÒ is: a complete research and execution platform whose validation layer reliably refuses to deploy mirages, applying current language model systems engineering to a domain that punishes self deception harder than almost any other.

The open question is now running as a live experiment rather than a plan. Signals drawn from information the price does not already contain cannot be backtested honestly, since no clean history of it exists, so each hypothesis is registered in advance and forward tested on paper with the book hedged to market neutral, so the record reflects the selection rather than the weather. That evidence accrues in weeks, and most of the hypotheses are expected to die. Publishing the ones that fail is the point.