Data And Intelligence
Canonical snapshots, feature bundles, provider truth, and the boundary between evidence and reasoning.
Data posture
Raw provider payloads are not reliable enough to become agent truth directly. Providers disagree, fields are missing, freshness varies, and textual summaries can hide important gaps.
The repo direction is to normalize external data into typed contracts before handing it to the model layer.
Key contracts
CanonicalAnalysisSnapshotDecisionFeatureBundleMarketContextPack- provider-specific fundamental or news snapshots before canonical merge
Why this matters
The model should not be asked to infer unstated finance truth from vague strings. It should receive:
- structured summaries
- missing-field visibility
- freshness context
- source attribution
- explicit risk flags
Provider aggregation rules
Good aggregation in this repo means:
- prefer canonical output over provider-specific string parsing
- keep missing provider evidence visible
- preserve source attributions
- degrade confidence when the evidence layer is thin
- keep generic fallbacks clearly labeled as fallbacks
Feature preparation
The feature layer is where the repo turns mixed market, provider, and runtime signals into a compact decision context. If this layer is missing or incomplete, downstream confidence should reflect that.
That is why missing context or missing decision_features should be treated as a real provider or evidence gap rather than a cosmetic omission.
Review and persistence tie-in
These structured artifacts are not only for live reasoning. They also matter for:
- review surfaces
- QA
- replay
- debugging disagreement between providers and agents
Safe mental model
Think of the model as a reasoning layer sitting on top of typed evidence, not as the place where missing market truth gets invented.