Data quality

Freshness

• Every metric, and every calculation method of it, carries a last-refreshed timestamp: GET /v1/freshness (no key). The catalog shows a fresh flag.
• Each metric has a freshness SLA that follows its cadence: sub-hourly and hourly metrics 6 hours, daily 30 hours, weekly 8 days, monthly 35 days. A few metrics have a documented override because their inputs legitimately pause (for example metrics that need official FX bars, which do not print at weekends: 72 hours).
• A metric past its SLA keeps its last value but is flagged STALE, is hidden from the default catalog, and appears on the status page.
• All metrics are checked every 15 minutes. The status page shows 90-day availability (the share of checks in which every live metric was within its SLA) and an incident list; the same data is at GET /v1/status/history. We are alerted automatically when any metric goes stale.

Gaps

• Missing inputs are never filled in. We do not interpolate, carry forward or substitute a proxy silently. If a metric's required inputs are missing, no value is published for that period, its timestamp does not advance, and it becomes STALE once its SLA passes.
• Every value has a provenance label and a data-quality score reflecting input completeness. Declared approximations are named as such in the metric's description.

Revisions

• Policy: a live value, once published, is not edited. Corrections go forward as new values; the affected period is listed as an incident on the status page.
• Historical rows that were reconstructed are labelled backfill_recomputed, backfill_approximate or backfill_model_retro and sit beside, never instead of, live points. See Data provenance and Backtesting.
• When a metric's definition changes, its definition version is incremented and the change is appended to the history at GET /v1/metrics/{id}/definition.

What is public about each metric

For every metric (GET /v1/definitions, no key): the data inputs (providers), the lookback windows used by the calculation, the update frequency, and the definition version with its history. For classic metrics the formula is published as well. For proprietary metrics the weights and tuning are not published; because history is point-in-time (next page), you can still test them honestly.

Lookback windows are read from the implementation and listed as found; when none could be extracted the page says so rather than guessing.

AI, language-model and quantum methods are experimental

Metrics can be calculated by classical code, AI, Claude, ChatGPT, Gemini, a quantum simulator and real quantum hardware. Everything other than the classical calculation is an experiment with a stated method, published so it can be compared with the classical value on the same inputs (method comparison). None of it is presented as, or claims to be, an advantage or a source of returns, and at today's problem sizes quantum values are expected to match the classical value within noise.

Madjik metrics are informational market indicators, not investment advice or a recommendation to buy or sell any asset.