Superhuman Data
for Finance

Existing training data asks experts to reconstruct their reasoning after the fact — producing accounts contaminated by rationalization and hindsight. Engram captures reasoning as it happens, preserving the signals that retrospective annotation destroys: the 4-second pause before an insight, the confidence shift after cross-referencing a source, the pace change that preceded a revision. Time is information. The temporal architecture of expert reasoning is a first-class training signal, not metadata.

Each session produces three interlocking training signals: event classification (per-event cognitive typing), reasoning chains (multi-step logic reconstruction), and session synthesis (full-session analysis). The deliberative structure — hesitation, revision, hypothesis rejection — is encoded directly in the trace. Standard SFT requires extensive RLHF to correct errors the training data failed to prevent. When the preference signal is already in the data, RLHF becomes a refinement step rather than a corrective one.

Every training example embeds the screenshot the expert was looking at when they spoke — base64-encoded inline via VISION format. The model learns situated reasoning: cognition grounded against the specific financial data on screen. Voice cross-references visual context, making the data self-verifying by design.

Each session refines the reasoning format itself. As the corpus grows, it accumulates domain-specific structure — branching patterns, edge-case heuristics, decision-point taxonomies — that later sessions build upon. Session 200 is structurally richer than session 10. The result is a corpus that compounds in depth, not just volume, and grows harder to replicate with every session.