The Data Pyramid (T0-T5 abstraction tiers)

Last verified2026-06-12

The platform computes hundreds of persisted values plus a stack of derived states, and surfaces a curated subset across the dashboard. The data pyramid is the model that keeps that surface honest: every datum carries one altitude label — how far from the raw source it is, and how close to an action. Six tiers, T0 at the base, T5 at the top.

  T5   ACTIONABLE     the AI brief: direction, stance, downturn read
  T4   REGIME/STATE   regime label, fragility, alignment severity, cascade
  T3   LENS           health score, drawdown lens, base-rate matcher, state flags
  T2   INDICATOR      the Signal-Alignment categories
  T1   SIGNAL         GEX, DIX, breakevens, correlations, the drawdown legs
  T0   RAW            index closes, raw FRED/COT series, option-chain pulls

Once everything is labeled, three decisions become mechanical: what belongs on a card vs in drill-down vs backend-only (display), what earns calculation and validation effort (compute priority), and how the AI’s signal pack and the UI group and order their content (legibility). The goal end-state is a top-of-pyramid read that is legible at a glance, with the ability to drill down through the layers beneath — market-literate audience, no hunting.

The two axes

Every datum is classified on two independent axes. They are not the same thing:

Axis 1 — pyramid tier. Where it sits in the abstraction stack: T0 raw → T5 actionable. A structural fact about the data.
Axis 2 — validation status. What the platform actually knows about it: validated out-of-sample, in-sample-only, heuristic, or pure state-descriptor.

The surface rule ties them together: surface = f(tier, validation). Tier sets the default altitude; validation earns or forfeits it. A T3 lens that fails honest out-of-sample validation gets demoted toward drill-down with an honest label even though its tier says “card.” A strongly-validated, decision-relevant T1 signal can be promoted onto a card. Tier is the default; validation is the override.

The six tiers

T0 — Raw

Ingested values with zero derivation — “what the source said.” Pure data carriers. Index and ETF closes, raw FRED series (EFFR, unemployment, initial claims, core PCE, the Fed stress indices), raw CFTC/COT positioning legs, OFR stress subindices, CBOE put/call ratios, the NAAIM exposure survey, VIX term-structure points, the live per-strike gamma profile.

Default surface: backend-only — the AI’s signal pack and the API carry them; they never earn a card unless reached via drill-down.

T1 — Signal

One read that carries meaning — either raw-that-matters (already meaningful with no calculation, like GEX, DIX, or HY OAS) or one calculation/classification on raw (breakevens, real yields, the rolling correlations, the 252-day net-gamma percentile rank, the VIX term slope, the energy regime, the health-score components, the z-scored drawdown-lens legs, the gamma walls). The widest tier a human ever reads directly.

Default surface: drill-down and detail tables; a few headline signals earn a card slot.

T2 — Indicator

The Signal-Alignment categories — each fuses one or more T1 signals into a single BULLISH / NEUTRAL / BEARISH read versus price direction. Twelve categories (news_sentiment flag-gated; eleven active by default).

Default surface: the Signal Alignment card — this tier’s one canonical home.

T3 — Lens

State lenses — multi-signal composites that output a score or a state but are still a measurement, not the verdict. Distinct from T2: indicators ask “does this agree with price?”; lenses ask “what risk/setup state are we in?” Members: the Health Score, the drawdown-risk lens, the base-rate matcher (“when markets looked like this”), and the deterministic state flags (air pocket, re-entry, squeeze setup, disorderly stress).

Default surface: card-level reads.

T4 — Regime / State

The top classification — “what kind of market is this,” which drives which lower-tier signals lead and which thresholds fire. Members: the market regime label, fragility, alignment severity, and the cascade stage.

Default surface: top-card headlines.

T5 — Actionable

Decision-support synthesis over everything below — the AI brief: headline, market overview, the graded 1/3/5-day direction reads, the drawdown-risk call, the downturn read, and the forward-graded stance. What a human or agent actually weighs before acting.

Default surface: the AI card on The Tape.

How the platform self-labels

The classification is machine-readable, not tribal knowledge. Every metric in the platform’s signal registry carries a pyramid_tier stamp (T0-T5), plus two companion stamps: horizon (the forward window the metric speaks to, e.g. coincident / 1d / 5-10d) and skill_stamp (the prose out-of-sample evidence behind that horizon claim). A horizon claim without evidence fails the build — claims trace to verdicts. The AI brief’s signal pack serializes its sections highest-tier-first, so the model reads altitude cues instead of a flat several-hundred-column blob.

Reading the dashboard with the pyramid

The Tape leads with the top of the pyramid: regime and state headlines (T4), the AI read (T5), the risk lenses (T3).
The Engine carries the wider middle: alignment (T2), the signal grids (T1), with raw carriers (T0) reachable through detail tables and the API.
Tiers can shift only when a datum’s abstraction level genuinely changes; validation changes move its surface, not its tier. A heuristic that earns out-of-sample validation gets promoted in visibility — honestly labeled either way.