Provisional Metric
A metric that is published but temporarily excluded from the MEI while its normalization bounds are being calibrated from real data. It joins the composite only once its calibration window closes.
Rail: Macro · Updated: 2026-07-09
What It Is
A provisional metric is one that MachineEconomy.ai is collecting and displaying, but is deliberately not yet feeding into the composite index, because its normalization bounds haven't been finalized. When a metric can't have its scale anchored from prior data — a genuine "cold start" — the platform doesn't guess the bounds from a single first reading. Instead it opens a defined calibration window, collects real observations across that window, and only then fixes the metric's floor and ceiling. Throughout that window the metric carries a provisional flag, and it is excluded from the index calculation entirely; the composite is computed only from live, fully-calibrated metrics.
The calibration anchor is not the first value observed, which could be a spike or a trough. It is the geometric mean of the whole calibration window — geometric because that average has to live on the same logarithmic scale the normalization uses, and because it's robust to single-day anomalies. Different metrics use different window lengths depending on how noisy their launch behavior is. At window close — a pre-announced calibration event — the platform runs its launch-band check, finalizes the bounds, clears the provisional flag, and the metric becomes a full index input from that point forward.
Where a provisional metric has a reflective twin (a second metric measuring the same underlying thing), the twin carries their shared category in the meantime, so the index still reflects that category while the provisional metric calibrates. Where a provisional metric is the only one in its category, that category is simply, and transparently, absent until calibration closes.
Why It Matters for the Machine Economy
Provisional status is the honest bridge between two of the platform's commitments: that every index input must have published, defensible bounds, and that no unanchored number is ever quietly averaged into the score. A brand-new metric can't satisfy the first commitment on day one — its bounds aren't yet earned from data — so rather than either inventing bounds or excluding the metric silently, the platform shows the metric, marks it provisional, and keeps it out of the composite until its scale is properly calibrated. A reader sees the data and sees the flag, and knows exactly why it isn't yet scored.
In practice, the metrics that pass through a calibration window are the ones whose history has to be reconstructed before their scale can be set — for example, the ERC-8004 identity-registry activity metric and the PyPI agent-framework download basket, each of which is backfilled from its real historical record and then de-provisioned onto finalized bounds. Provisional status is distinct from a data gap: a provisional metric has real, verified data and is on a clear path into the index, whereas a named gap is a construct with no valid metric available at all. It is also distinct from a stub — the platform never lets a placeholder value into the composite. Provisional is the state of "real metric, real data, scale not yet finalized," and it exists so that the transition from new signal to scored input is disciplined and disclosed rather than abrupt.
Related Terms
- MEI (Machine Economy Index) — the composite a provisional metric is temporarily excluded from
- Backcasting — how a provisional metric's history is reconstructed to set its anchor
- Normalization — the bounds a calibration window exists to finalize
- Named Gap — the distinct case of a construct with no valid metric at all