Signal Evaluation
This section groups the documents that explain how AI, signal interpretation, decision support, and replay validation fit together on the current branch.
Core artifacts
- Documentation index
- Algorithms
- AI intelligence
- Decision support playbook
- Temporal feature upgrade status
Integrated flow
- live feeds and structured services create a current snapshot
- AI, event resolution, and graph layers build evidence-grounded context
- decision-support logic maps signals into structured candidates
- replay and historical validation test whether those candidates are coherent
- validation results refine evidence and admission quality over time
Public mock workbench
The public docs include a click-through mock replay workbench. It is not connected to private feeds, but it mirrors the product structure.
- point-in-time datasets
- replay and scenario comparison
- operator decision posture
- hot / warm / cold storage lifecycle
Mock Replay Studio
Scenario and backtest workbench
This public demo uses synthetic point-in-time data shaped like the real replay stack. You can switch datasets, compare scenarios, and inspect how storage, evaluation, and operator posture move together.
Scenario
Middle East energy shock
Escalation lifts oil and shipping stress while safe-haven positioning turns defensive.
Replay curve
ACLED · conflict events · 91% coverageInput datasets
conflict events · 91% coverage
Conflict and protest events anchor the regime shift signal.
news / document stream · 78% coverage
News burst intensity confirms narrative acceleration around shipping routes.
price series · 96% coverage
USO, XLE, GLD, and TLT provide tradable exit points for the replay.
Decision posture
High conviction only when shipping stress and crude momentum confirm together.
Macro overlay prioritizes capital protection over fresh cyclic exposure.
Scenario timeline
ACLED and news spikes land in Redis and feed the current snapshot.
Transmission edges and hedge bias are recorded in the replay frame.
Max-hold fallback closes the position if no earlier clean exit appears.
Data lifecycle
Live conflict/news payloads stay in Redis with short TTL and schema checks.
Replay frames and run summaries persist into PostgreSQL for operator review.
Parquet snapshots archive the scenario window for later point-in-time reproduction.
Current limits
- some probabilistic layers remain practical approximations
- replay quality depends on point-in-time data completeness
- the main branch is not a full autonomous trading stack