Skip to main content

research platform

Every strategy starts out false.

The research layer inside BLKPVNTHR.OS, built as a system in its own right. It exists to try to break ideas, not to flatter them — a candidate has to survive an explicit chain of validation stages before anything downstream will look at it.

Paper-trading research first — live orders are gated.

Open the Research Division (opens os.blkpvnthr.com in a new tab) See the lifecycle Discuss a project
os.blkpvnthr.com/research-division
The Quantitative Research Division: a board of laboratories with readiness scores, above a row showing the stages of the research lifecycle.
The Research Division. Any figures shown are the product’s own demo state — not real finances, and not a track record.

at a glance

What it is

Role
Sole engineer. Research design, engine, and deployment.
Engine
A FastAPI service, deliberately stdlib-first: fastapi and uvicorn are the only runtime dependencies.
Python packages
research_engine, pattern_engine, alpha_discovery, evidence_fusion, strategy_lab
Scale
18 laboratories, each asking one question, each carrying a readiness score.
Deployment
A systemd unit bound to 127.0.0.1:8010, reverse-proxied at /research-api — nothing new is exposed publicly.
Posture
Paper-trading research first; live orders are gated. Execution is kept separate from analysis.

the problem

A backtest is not evidence

Most trading research is a backtest and a good feeling. A curve that looks strong in-sample usually does not survive costs, a different period, or a different universe — and nothing in the ordinary workflow forces it to try. The failure is not statistical so much as procedural: there is no gate a bad idea has to get through, and no record of what was already ruled out.

So the design goal was not “find strategies.” It was to make an unproven idea expensive to keep: give every candidate an explicit set of hurdles, make the evidence and its provenance a first-class object, and treat anything that has not cleared the chain as what it is — a hypothesis.

the lifecycle

The chain a candidate has to survive

This is the sequence the Research Division actually runs. A candidate enters at the top and is dropped at the first stage it cannot clear.

  1. Find something

    • Observation
    • Phenomenon detection
    • Hypothesis generation
    • Counter-hypothesis

  2. Try to kill it

    • Experimental design
    • Statistical validation
    • Economic validation does it survive costs?

  3. Make it prove itself again

    • Historical robustness
    • Paper-trading validation
    • Replication

  4. Only then

    • Publication
    • Market law
    • Deployable strategy paper first

Each stage depends on the stage above it; a candidate is dropped at the first stage it cannot clear.

Reaching the end of the chain makes a candidate deployable in the paper sense. It does not make it a recommendation, and it is not a signal to act on. See the Financial Disclaimer.

laboratories

18 laboratories, one question each

Rather than one monolithic model, the division is split into laboratories that each interrogate a single aspect of market structure. Each carries a readiness score, so it is obvious which parts of the system are actually load-bearing and which are still speculative.

Market data

What is signal and what is noise in the raw tape?

Time series

Is the series stationary?

Stochastic

How do events arrive?

Market structure

How does price form?

Options & volatility

What does the volatility surface imply?

Information theory

How much information is in this feature?

Causal inference

Does X cause Y, or merely correlate?

Graph

What is the market’s topology?

Eight of the eighteen. The rest are visible in the live Research Division.

the engine

Engineering notes

Stdlib-first, on purpose

The research engine’s only runtime dependencies are FastAPI and uvicorn; the core is plain Python. Fewer dependencies means a deterministic result, a fast cold start, and a service that deploys to a plain systemd unit without a container stack behind it.

Evidence carries its provenance

evidence_fusion exists so a claim can be traced back to the run and the data behind it. A result that cannot be traced is not a result.

Discovery is separated from validation

pattern_engine and alpha_discovery propose; strategy_lab and the validation stages dispose. The half that generates ideas is deliberately not the half that grades them.

Execution is a different system

Research proposes; nothing it produces reaches a live venue on its own. The trading surface is paper-first in the deployed product, with a discipline layer applying risk limits before a strategy can progress.

Stack

  • Python
  • FastAPI
  • uvicorn
  • systemd
  • nginx
  • React 19
  • TypeScript
  • Recharts
  • Zustand
  • Alpaca (paper)

maturity

Where this actually stands

This is an actively developed research platform, not a finished product. Some laboratories are production-oriented; others are experimental and carry a low readiness score precisely because they have not earned a higher one yet. It runs paper-trading research first before certifying a strategy for live trading. Live orders are disabled until a strategy consistently averages $100/day net.

The platform publishes no track record. Nothing here is investment advice — see the Financial Disclaimer.

Open the Research Division (opens os.blkpvnthr.com in a new tab) See the whole system All case studies