Model Health Explained

Model Health is a planned concept — a validation layer being designed for a future implementation phase. It is not currently a live, calculated feature of the Syntalium platform. No HEALTHY / WATCH / DEGRADED / DISABLED status is produced or displayed in the current system.

This page describes what Model Health is designed to do, what the health state taxonomy represents, and why this kind of validation matters in a market intelligence system. The intent is to give users and reviewers a clear picture of where the methodology is heading, without overstating what is currently built.

The Model Health layer is designed to answer a specific question: is the evaluation engine currently operating within acceptable reliability bounds? It is intended to be a meta-layer above the market condition evaluation — monitoring the evaluator rather than the market.

Calibration drift

Markets change regime over time. A model calibrated to one period of volatility or structure may produce less reliable outputs in a materially different regime. Model Health would detect when this drift exceeds a threshold.

Input quality

The reliability of evaluation outputs depends on the quality and consistency of the underlying data. Model Health would monitor whether inputs are meeting expected standards across cycles.

Output consistency

If the evaluation engine is producing outputs that are inconsistent with historical patterns in ways that suggest model degradation, Model Health would flag this before it affects published signals.

Automatic adjustment

When health degrades below a threshold, the planned behaviour is for the platform to automatically withhold setup publication — preventing the platform from publishing signals under conditions where reliability cannot be confirmed.

The planned Model Health taxonomy uses four states, each representing a level of reliability assessment for the evaluation engine:

HEALTHY

The evaluation engine is operating within expected reliability parameters. Input quality, calibration, and output consistency are all within acceptable bounds. Setups may be published normally.

WATCH

One or more monitored indicators have moved outside the normal range but have not yet crossed the threshold for degraded status. The engine is functioning but under closer observation. Setups continue to be published, with the WATCH status surfaced to users who check it.

DEGRADED

The evaluation engine has crossed a reliability threshold. Outputs may be less consistent or calibration drift has been detected. In this state, the planned behaviour is to automatically restrict or withhold setup publication until the issue is resolved.

DISABLED

The evaluation engine is not running or has been explicitly halted — for maintenance, data source failure, or other operational reason. No setup publication occurs in this state.

Market intelligence systems face a structural challenge: a model that was valid in one market regime may produce lower-quality outputs in a different regime — and may do so without obvious external indicators. A user relying on setup signals has no direct visibility into whether the engine producing them is currently operating in a period similar to its calibration data.

Honest disclosure

A Model Health layer makes the platform more honest about its own reliability. Instead of silently publishing signals in a degraded state, the system explicitly flags when outputs may be less trustworthy.

Automatic protection

By building health monitoring into the publication logic, degradation becomes a trigger for automatic withholding rather than requiring a user or operator to notice and intervene manually.

Auditability

A health history layer would allow future review of when the engine was in each state, correlating health status with evaluation periods for retrospective analysis.

As of the current release, Model Health is not calculated, not displayed, and not reflected in any API response field. The platform's current reliability mechanisms are:

Freshness monitoring

FRESH / STALE / NO_DATA reflects whether the underlying data is within expected thresholds. Stale data triggers withholding behaviour.

Safe mode

If freshness or data quality conditions fall below safe thresholds, the system enters safe mode and withholds setup publication.

SNAP proof

Every published record carries a SHA-256 hash for tamper-evident archival. This is a record integrity mechanism, not a model reliability indicator.

Model Health, when it is implemented, will be documented here with its actual behaviour, thresholds, and API fields. Until then, this page reflects the planned design only.