Helmsman did not emerge from one uninterrupted implementation plan. The broader project included experiments that generated useful code and data without justifying the original idea behind them.
The Polymarket exploration is a good example. We built collection and analysis machinery, investigated several approaches, and recorded reasons not to carry some of them forward.
Separate the collector from the hypothesis
Polymarket hosts markets about event outcomes. For a simple binary example, a share bought for $0.60 can pay $1 if its outcome resolves true and $0 otherwise. That example explains the data being studied; it is not a result from our experiments.
Our exploration included a collector for market data and a model-generated graph of proposed relationships between events. Those are different artefacts with different standards of evidence.
The collector can accurately preserve an observation even if the proposed relationship is wrong. Keeping that separation made it possible to retain useful data plumbing without endorsing the analysis built above it.
A proposed relationship is not a demonstrated cause
The graph experiment produced relationships that the project notes explicitly marked as untrustworthy. A fluent explanation for why two events might be connected was not enough to establish the connection.
The lesson was architectural: a model can propose candidates, but the platform needs independent evidence before treating those candidates as established relationships.
The records also identified a mismatch between properties written by the graph builder and properties expected by a downstream reader. That is a separate failure mode. Even a running job can produce no useful work when producer and consumer contracts disagree.
Keep the verdict narrow
Some approaches were rejected in the explored setting. Other ideas remained untested. Those are different outcomes, and neither should be turned into a claim about every market or every possible implementation.
The useful carry-forward was the collection work and the ability to inspect captured data. Reusing that foundation does not require reusing every original research hypothesis.
Other choices we deliberately did not make
Helmsman also records a decision against deep historical model re-scoring as its validation foundation. Hindsight in models and revised source data could undermine the meaning of the result.
For broad calculation sweeps, we chose a smaller replay kernel instead of adopting a larger event-driven engine for a workload that did not need that shape.
These decisions belong in the build story. Recording what we rejected, why, and what would make us reconsider keeps the project from repeatedly paying for the same uncertainty.
Polymarket’s documentation explains the basic outcome-share model. The project conclusions above describe our own historical investigation.