manifold-agent

This goes to the bot as a GitHub issue. It answers publicly, usually within a few minutes, and leaves the thread open so you can reply and keep going. It remembers who it has spoken to, across markets and across threads.

Each tick

Reacts before it explores: filled limit orders, then held markets that moved 8 points, then replies, then a scan for new markets. It does the most urgent single thing it can afford, writes its state to a git branch, and exits.

What it looks at

Markets with 25+ traders, M$3k+ volume, resolving inside a month. Every filter is a cheap API read, so a market has earned the model's attention before the model is invoked.

Two models

A cheap model forecasts every candidate blind, without being shown the price. Only markets where its number disagrees with the market, or where it flags something worth reading properly, are passed to the expensive model. Most candidates die here for the price of one small call.

Then two calls

A research pass with live search that is forbidden from stating a probability, then a structured decision that reads the market, the whole comment thread, the research, its standing notes, and its own memory. The market price is withheld from that call too: shown the price, it hands the price back.

Evidence first

The response schema forces evidence for, evidence against, and resolution risk to be written before the number. Left alone, small models put every answer on 0.7 or 0.3, and a flattened distribution destroys the only signal worth measuring.

Sizing

Fractional Kelly on the gap between its estimate and the price, shrunk for distant resolution dates, capped at 10% of net worth and 5% of market volume, then trimmed until a dryRun bet shows it moves the price under 5 points. Nothing the model writes can raise a stake: sizing is computed in code after it has spoken, which is also why nobody can talk it into a bigger position.

What it keeps

A compressed narrative of what it has done, a note on each market it has looked at, every conversation it has had, and a short list of standing notes that sit in front of every decision. It writes most of those notes itself after a trade. The only other source is the person who runs it. Anyone else can argue with it and change its mind about a checkable fact, but a market full of anonymous accounts holding positions does not get to write rules into the thing placing the orders.

The point

Not profit. Once enough people run LLMs on the same markets the model consensus is the price. Every decision logs the model's probability beside the market's, which after a few hundred resolutions is a calibration score. That is the actual output.