Research bots

Handing the search to a machine, without handing it anything else

The unattended search used to be a browser tab left open. Someone would arm a cycle, leave the laptop running, and come back to whatever it had found. That worked, and it had three things wrong with it: it died when the tab did, nothing stopped several of them from overwhelming the application at once, and there was no way to tell afterwards which strategies a machine had produced and which a person had.

A research bot is that loop made into a program. It draws a set-up, mines it, runs the survivors through the proving suite, keeps what clears and labels it — then does it again, until its budget runs out or its own rules tell it to stop. The interesting part of building it was not the searching, which already existed. It was working out everything a bot must not be able to do.

---

Three questions, decided before any of it was written

None of those is a setting. A bot cannot be granted a broker connection by configuration, because the refusal is at the door rather than in a policy the door consults — and the doors are checked by a test that reads the functions themselves, so a new provider or broker call that forgets to add one fails the build.

---

The honesty floor, which nothing can lower

A bot is given a protocol: how much history to hold back, how many trades a result must rest on, what statistic to cut on, how many strategies it may promote, how long it may run. Most of those you choose. Some of them have a floor, and a protocol that asks for less than the floor is not rejected — it is raised to the floor, and told so in its own record.

• At least a fifth of the history is held back, and the cut is made on a statistic measured somewhere the search could not see. Cutting on the in-sample score is not available at any setting.
• The empirical null runs before the first cut, so the bar a candidate is judged against is established before anything is selected rather than after.
• Fills are next-open, and a result must rest on a minimum number of trades.
• Limits disarm; they never throttle. A bot that reaches a limit stops. It does not quietly continue at a reduced rate, because a machine that keeps going after its budget is spent is one nobody is reading.
• Budgets end a bot between universes, never mid-funnel — stopping halfway through proving a candidate produces a half-tested strategy, which is worse than none.

Above the floor, a bot can also be told to learn — to lean toward the kinds of set-up that have produced survivors before. Every switch in that group is off by default and each one names what it is doing in the bot's own log, because a search that quietly changes its own aim between runs is one whose results cannot be compared to each other.

---

Whose strategy is it

A machine producing strategies into the same store a person uses is a provenance problem before it is anything else. Every record carries an origin stamped by the one function that writes records, and the consequences of that stamp are deliberate:

---

What the first real run actually said

The first bot to run a full life was read back from its own log afterwards, and most of what it produced was wrong in ways that were only visible because the log recorded every decision and its reason. Three of them are worth stating, because they are the argument for keeping such a log at all.

• Twenty-six cycles turned out to be one search, run twenty-six times. Each pass was sent the same starting seed, and the generator's own duplicate check then discarded almost everything it drew. The honesty arithmetic was being divided by trials that had never happened — which made the results look better, not worse.
• Strategies were being promoted that had failed a test. A stage that rejected every candidate reported "nothing was removed", which downstream was indistinguishable from "everything passed". A proof that a row failed is not a proof that it passed, and a record-breaking candidate is precisely the one most likely to have been lucky.
• Most of the survivors lost to buying and holding. Of twenty-four, twenty did — because the funnel had never been made to ask that question. It asks now.

This is the part worth taking from the whole feature. An unattended researcher is not valuable because it works while you sleep. It is valuable because it does exactly what it was told, thousands of times, and thereby makes the flaws in what it was told impossible to miss. Every one of those three defects existed in the manual workflow first. A person running the cycle by hand would have hit all of them and noticed none.

---

What changed because of it

Each of those findings became a rule, and the rules became the defaults a new bot is deployed with. A stored bot keeps the protocol it was deployed under — changing a default never silently changes what an existing bot is doing.

Operating one

• It says what it will run before it runs. The deploy form carries a counter pricing the whole plan — how many engine passes each stage will cost at the protocol's own settings — and previews every coercion and advisory the deploy would record.
• It can be reconfigured after deployment. Pause it, edit it in the same form it was deployed with, and resume or continue — the edit lands between stages, never inside one; the whole protocol is re-checked, so an edit cannot reach a value a deploy could not; and every change is written to its ledger as setting: old → new.
• Its parameters can be kept. A whole protocol and its budget can be saved under a name and redeployed later — with the deadline stored as a number of days rather than a date, because a stored date would deploy a bot that ends on its first tick.
• "System terms" mean what the system does. The funding assumptions a bot mines under follow the same switch every manual backtest follows; a bot that wants different terms names them explicitly.

---

Watching one

A bot's page shows what it is working on, how far through the current stage it is, and how it is doing — the count of universes tried, survivors kept, what cleared the gauntlet and what the empirical null says about the body of its results rather than its best row. Every number comes from the bot's recorded state; nothing is parsed back out of its log, because a display that re-derives a fact from prose will eventually disagree with the fact.

The funnel is drawn rather than listed — each stage a band on a logarithmic scale, because four hundred candidates narrowing to twelve would otherwise be a sliver — and the bank of what it kept is drawn as a distribution against the noise floor rather than as a leaderboard.

The Mithraeum — a research bot, opened

Bots can be paused, resumed, cloned, continued after they have ended, and deleted — and deleting one offers to clean up what it produced rather than leaving orphans behind. What it mined and promoted survives on its own, still labelled.

Nothing on this page is investment advice or a performance claim. A bot produces candidate strategies and evidence about them; it does not produce results, and no figure here is one. An unattended search that promotes a strategy has established that the strategy survived a battery of tests, which is not the same as establishing that it will work. Trading involves risk of loss.