Seventeen shared-resource clusters discovered on IBM's ibm_fez. Fifteen validated across multiple coordinated-motion events.

Seventeen clusters. Two events were the cutoff at the start of the data window; the other fifteen recurred.

A cluster is registered when a group of qubits shows persistent within-group correlation above a noise-floor null distribution. Confirmation requires at least two distinct coordinated-motion events at least 14 days apart, each with within-cluster |r| above 0.5 in a 7-day rolling window. The two single-event clusters at the bottom of the table were registered close enough to the start of the collection window that earlier events could be present in the archive but outside what we pulled.

★ The 9-qubit cluster is the one we wrote up in detail on Substack: How uncertainty evolved around a 9-qubit cluster. T1 co-moves counts events with simultaneous T1 movement; the remainder are readout-only.

PERSISTENT_FAIL holds 71% precision at both 1-day and 7-day horizons — a 9× lift over base rate that doesn't decay with the prediction window.

Per-qubit failure flags typically lose precision as you extrapolate further out. The substrate's PERSISTENT_FAIL flag doesn't. Below is the breakdown of verdict precision against the base rate of each actual state at horizon, across 59,192 scored predictions on all 156 qubits. Lift = precision divided by the base rate at horizon — it's the right way to compare alerting signals, and it's what separates a real alert from a re-statement of the prior.

Predicted verdict	1d precision	1d base rate	1d lift	7d precision	7d base rate	7d lift
PERSISTENT_FAIL	71%	7.5%	9.5×	71%	8.2%	8.7×
DEGRADING	43%	11.4%	3.8×	11%	11.7%	0.9×

Three operational tiers fall out of this. PERSISTENT_FAIL is a real alert at both horizons — 9.5× lift at 1d and 8.7× lift at 7d on the same 71% precision is genuinely unusual for a per-qubit failure flag. DEGRADING is a usable short-horizon signal at 1d (3.8× lift) — useful for next-day scheduling decisions like deprioritizing flagged qubits for sensitive work — but the lift collapses to 0.9× at 7d, where the flag stops carrying signal. TRANSIENT_ANOMALY and DEGRADING extrapolated to 7d are roadmap — the substrate produces them, but their precision needs recalibration before they enter a paid alerting product.

WAIT_ENVIRONMENTAL is a different signal class entirely. It's rare by design — it fires when multiple qubits across the chip show signs of an environmental event simultaneously, and chip-wide events are uncommon. The May 8–16 window, in which 85 of 156 qubits lost ≥30% of T1 over a week, is the canonical example. Per-qubit horizon precision on the flag is noisy by construction because environmental events recover stochastically; the chip-wide co-fire pattern is the right diagnostic. The operational action it implies is wait, not reroute. Some of the apparent DEGRADING → HEALTHY → PERSISTENT_FAIL flapping observed in individual qubits during environmental windows is attributable to those qubits transiting an environmental event that WAIT_ENVIRONMENTAL was already tracking at the chip level.

Two evaluation modes the dashboard tracks separately, and the page names both on purpose. State forecasts — predict which verdict label a qubit will land in at horizon — score 57% overall state-match across all five states at 1d. That's the operational product the alerts are built on. Value forecasts — predict the exact T1 or readout-error number — score 51% overall value-accuracy. The value bar is harder, and serves as a residual diagnostic (Q53's +11.82σ z_residual on May 12 was the canonical example) rather than the basis of the alerting product. Leading with state-match without naming the value-accuracy bar would obscure how the product actually evaluates.

A method that transfers off the quantum chip.

ibm_fez is the public proof. The substrate is built to run on any device that exposes a per-element calibration feed — quantum hardware, classical silicon, HPC cooling fleets, or model-based forecasting systems. The methodology stays the same. The integration is what's bespoke.