Autonomous Trading Agent

An agent that researches, decides, and executes without approval. Every position states its thesis and its falsifier before the order is placed; every outcome is graded against both.

• Cash is never fully spent. A reserve is always held back, and no trade may reach into it.
• Nothing is bought without an exit level already set. How far that level sits below the price depends on how much the name actually moves — a quiet stock and a violent one are not held to the same leash. Where a name moves too much to be protected at all, the position is sized as though it could go to zero and the exit level is called what it is: a warning, not protection.
• An exit level can only move one way. It can be raised to lock in a gain. It can never be lowered to give a losing position more room, which is the excuse this rule exists to remove.
• A broken exit level is not an automatic sale. The position has to be argued again from scratch, against the same standard it would face as a fresh purchase today. Doing nothing is not an answer.
• Nothing can block a sale. No rule on this page can trap a position.
• No number is taken on trust. Every figure a decision rests on is retrieved that session, with its source. Never an estimate, never a recollection. A run that breaks off ends in a written report rather than a best guess.
• Every purchase says in advance what would prove it wrong. Specifically enough to settle the question later, rather than in terms that could be argued either way once the answer is known.
• Declining is a decision, and it has to give a reason. A no-trade must name the single thing that, if it had been different, would have changed the answer. The record is checked after each session and a missing reason is flagged — this one is audited rather than blocked, because it describes something that did not happen.
• There is no fixed limit on position size, and that is deliberate. Nothing caps how much of the account a single name may become at the point the order is placed; concentration is reported rather than refused, on the view that a rule which has never once changed an outcome is decoration and not a safeguard. One backstop remains upstream: a check run before every order refuses a position that would leave the account effectively a single bet. The largest position ever held reached roughly thirty per cent of the account. Anything of that size has to argue for itself — what happens if it goes to zero, what the account would then have to make back, and why the view is better expressed whole than halved.

The learning loop cannot change any of these on its own. That is not the same as saying they never change — they have been changed, deliberately, and the list above is what stands today.

As few limits as possible, as many as necessary.

• The argument is written before the order, never after. Facts carry their sources. What would prove the trade wrong is stated in advance. So is a plain number for how likely it is to work.
• That record is locked before the order leaves. It cannot be edited afterwards, and any attempt to would be visible. So the confidence graded later is the one that was actually held, not a tidier version written once the answer was known.
• Declining is scored too. A pass is recorded as a prediction that the name would not beat the index, and graded like any purchase. Most of the judgement here belongs to trades that were never made, and almost nothing else measures that.
• The scoring is not flattering, and it is not finished. Every matured prediction is graded against what the market did. On the current record the agent is very slightly better than guessing — and the sample is far too small to mean anything. The tool that produces the number says so itself, and refuses to let a conclusion be drawn from it. That is the honest position today.
• None of it changes the rules above by itself. Lessons are drawn only when the evidence carries them, and they can adjust judgement, never the limits.
• A second model may propose changes; it cannot make them. Once a week it reads the record and suggests how candidates might be found or ranked. It has no access to the broker. A third model attacks the proposal, code scores the argument, and anything still standing has to clear a test against history before it changes behaviour. Nothing reaches real money by two models agreeing with each other.