The hard question comes after the order
A market operation is easy to describe while it is happening. A quote changes, a condition is met, an order is submitted, and a position appears. The difficult questions arrive later. Why did this position exist? Which archived observations supported it? What assumptions shaped the entry? Which rule was supposed to close it? Was the operator looking at paper mode or live mode when the decision was made?
A system that cannot answer those questions has moved data, but it has not preserved accountability. Cerberus is organized around that distinction. It is a full-lifecycle trading operations system for Kalshi prediction markets, built as three connected heads: Ingest, Backtester, and Dash. Each head owns a different kind of work. Together they make a chain that can be followed forward during operation and backward during review.
Data moves from capture to validation to action. Accountability moves from action back to the evidence and intent that produced it.
Capture is not the same as memory
Ingest captures live WebSocket streams into Parquet archives and tracks coverage in a SQLite catalog. That architecture separates raw market history from the application screens that happen to be open at a given moment. The archive is not merely a cache for another feature. It is the memory head of the system.
Memory needs boundaries. A file can exist while a time range is incomplete. A stream can reconnect while leaving a gap. A replay can run while relying on thinner evidence than an operator expects. Cerberus is designed to preserve raw data, detect gaps, and make replay quality auditable. It does not claim that every interval is complete or that captured data is perfect. That limit is part of the operating model.
Coverage metadata matters because provenance begins before a strategy runs. If the source interval is uncertain, later precision does not repair it. The right response is to surface the uncertainty, not smooth it away.
Validation needs executable assumptions
Backtester is a local, read-only research tool. It replays archived data without placing orders or changing production trading state. This boundary is practical. Research should be able to inspect the evidence without inheriting authority over live execution.
The replay model also uses executable prices. Buys are evaluated at the ask, while exits are evaluated at the bid. That choice resists a common source of false confidence: treating the midpoint as though it were available for an actual fill. The point is not to make a simulation pessimistic for effect. The point is to keep each assumption legible.
Per-trade audit charts connect entry and exit markers with thresholds, raw quote paths, and diagnostics. A summary can show what a rule produced. An audit chart can show how it produced it. That difference matters when an operator is deciding whether a result reflects a durable idea, a data gap, or a modeling choice.
Validation in Cerberus therefore means more than running a strategy over a file. It means preserving enough context to challenge the result:
- Which archive and coverage range supplied the replay?
- Which executable price was available at entry and exit?
- Which threshold or sell rule changed the position state?
- Which diagnostic explains skipped or unresolved cases?
No single answer proves that a Play should move forward. Together, they give a human operator material for judgment.
Action must retain its reason
Dash is the execution head and a live production prototype. Its central object is the Play. A Play ties intent to order behavior, position state, and eventual exit or settlement. The provenance chain runs from Play to buy order, from buy order to entry fill, then to sell rules and exit or settlement. Every position should answer two concrete questions: why does this exist, and what happens next?
That chain is more useful than a flat order log. An order log records events. An operational record explains how those events relate to a decision. When conditions move quickly, that structure reduces the need to reconstruct intent from timestamps and memory. When a position remains open, it keeps the next rule visible instead of treating entry as the end of the workflow.
Human control is part of the architecture
Cerberus is not an autonomous trading bot. A human operator directs it. An AI assistant may analyze data and propose Plays, but it cannot independently place orders or change trading mode. The distinction is enforced through system boundaries rather than left as an expectation.
Confirmation gates make consequential actions explicit. Paper and live modes are separate operating contexts. Reduce-only behavior constrains the direction of risk changes. Kill-switch awareness gives the operator a clear way to reason about halted action. Multi-account support and backend-only credentials keep authority scoped to the systems that need it.
These controls do not remove judgment. They preserve it. An operator still decides whether evidence is sufficient, whether a Play belongs in live mode, and whether changing conditions invalidate the original premise. The system's job is to keep that decision attached to state and make authority difficult to confuse.
Current limits are operational facts
The three heads do not share one maturity label. Ingest is a production-grade recorder in continuous operation. Backtester is a local research tool. Dash is a live production prototype. Those distinctions set expectations for how each component should be trusted and reviewed.
Cerberus also makes no claim of complete market data, guaranteed execution, strategy edge, or trading performance. Replay evidence can expose assumptions and gaps, but it cannot turn a model into certainty. Human controls can narrow authority, but they cannot eliminate operator error. Provenance can explain why an action occurred, but it cannot make the action correct after the fact.
That is the measured value of the system. It creates a route from live action back through rules, fills, Plays, replays, coverage, and raw observations. The route does not promise an outcome. It makes the operation accountable to what was known, what was assumed, and who chose to act.