Operator Reliability Score
Operator Reliability Score is a comprehensive measure of how consistently an operator proposes the blocks they are assigned. It isolates the one factor operators control — block production — from network luck and market variables. It does not measure yield (which is tracked independently via APR).
The score can be evaluated over two distinct timeframes using the same Bayesian formula:
- 90-day: Reflects recent operational quality. This surfaces infrastructure issues from the last three months while smoothing out day-to-day variance. Covered misconfigurations count as proposed here exactly as in the all-time score, and an uncovered misconfiguration gets an 8-day grace period before it counts against the score — operators settling delegator payouts off-chain have time to cover a slot before it affects their ranking. The slot stays on record as misconfigured throughout and remains in the published assigned count; for the score itself it is simply set aside — counted on neither side of the formula, so it neither helps nor hurts — and once the grace expires without coverage it counts against the score like any other uncovered misconfiguration. For the SCORE the grace applies only to windows longer than the grace itself, so the shorter windows the API can serve (7-day) charge a misconfiguration immediately, otherwise the slot would leave the window before the grace ran out and never be counted at all. APR defers the same 8 days but applies it to every window, including the 7-day: an operator still inside its settle window should not read 0.00% yield on a reward it may yet deliver. One consequence is deliberate. In the 7-day window an in-grace misconfiguration ages out before the grace expires, so it is never deducted from that window's APR, which is why the 7-day APR and the 7-day score can disagree during grace. The grace covers misconfigurations only. Missed slots and slashes are never graced, on either side.
- All-time: Represents the full historical record. It rewards sustained, long-term consistency.
Components of the Reliability Score
The scoring mechanism uses a Bayesian formula to establish a baseline prior and adjust based on actual operator performance.
Score = ((max(0, B_proposed - P_total) + C * R_net) / (B_assigned + C)) * 100
Assigned and Proposed Blocks
B_assigned: Starts with every slot the operator was scheduled to propose (valid, missed, or misconfigured). Pruned slots are excluded entirely, as the chain dropped them, not the operator. Checkpoint verdicts then make two independent score adjustments: each checkpoint the operator built but did not land on L1 adds one charge, while an individual later slot may be separately excluded when its proposer is expressly exonerated. The checkpoint charge applies whether or not a later slot is exonerated. Neither adjustment changes the chain's recorded miss counts.B_proposed: Represents valid slots plus misconfigured slots that have been covered by the operator topping up the registered split. Uncovered misconfigurations are not counted as proposed—the operator failed to deliver usable rewards. A misconfiguration still inside its 8-day grace period is never credited here — for the windowed score it is excluded from both sides of the formula until it is either covered (then it joins B_proposed) or the grace expires (then it counts against).
Slash Penalties (P_total)
Slashes erase earned rewards; they do not just represent a missed opportunity. Each slash is converted to block-reward units and subtracted from the numerator. The numerator floors at zero, meaning heavy slashing drives the score toward zero but never negative.
P_slash_i = slash_amount_i / sequencer_reward_at_event_i P_total = Σ P_slash_i
Each slash is weighted by the block reward in effect at the time it occurred. Future reward changes never retroactively re-weight history (e.g., Ignition-era slashes use Ignition's reward, Alpha-era slashes use Alpha's).
Network Rate and Confidence Constant
R_net(Network Rate): The network-wide proposal success rate within the same window. On a young rollup with L1 inclusion challenges, this can be well below 100%. The Bayesian prior uses this as the "expected" rate for an operator with no track record.C(Confidence Constant): Determines how much evidence an operator needs before their observed rate overrides the network prior.C = 10for individual sequencers. A few dozen slots are enough for the score to reflect actual performance.C = 50for pools. The higher constant keeps the score anchored until the pool has substantial aggregate data.
Benchmark Values
Using C = 10 (individual sequencer) and R_net ≈ 0.77 (current network rate, subject to shift as the network evolves):
Try it
C = 10 (sequencer) · R_net ≈ 0.77Tip: Notice that 1000 slots at a 5% miss rate (94.8) scores higher than 100 slots at the same miss rate (93.4). Sustained high-volume consistency is rewarded — the Bayesian prior has less influence when there is a larger dataset.
Pool Scoring
A pool's score is computed by aggregating its sequencers' counts (B_proposed, B_assigned, P_total), then applying the formula once with C = 50. It is not the average of individual sequencer scores. The aggregated B_assigned carries each member's independent checkpoint adjustments: an additional charge for a checkpoint that was built but not landed, and an exclusion only for a slot that was expressly exonerated. These adjustments do not cancel or transfer a miss, even when both operators belong to the same pool. The raw proposal and miss counts shown elsewhere on the dashboard remain the unadjusted chain record.
A large pool that consistently proposes thousands of blocks earns a score that reflects that volume. A pool with 8,000 assigned slots and a 2% miss rate is demonstrably more reliable than a pool with 100 slots and the same rate. However, a pool can score lower than its best individual sequencers because the aggregate absorbs uncovered misconfigurations and misses across the entire fleet.
Performance Thresholds
We calculate color thresholds dynamically using a rolling 90-day distribution of validator scores, ensuring operators are consistently measured against their current peers. We apply the same cutoff to both validators and pools, but the results vary: because pools aggregate multiple operators, they are less susceptible to Bayesian smoothing. Consequently, a cutoff capturing the top 10% of validators impacts roughly 20% of pools. Further differentiation is unfeasible today, as block proposals are the only available operator controlled performance signal; any additional metrics would ultimately derive outside the onchain reality.
- Green: At or above the 90th percentile (top decile performance).
- Orange: Between the network average and the 90th percentile.
- Red: Below the network average (underperforming relative to peers). The average is used rather than the median because the score distribution is bunched high with a low tail — a median cutoff would flag roughly half of all operators red.
Excluded Metrics
The Operator Reliability Score normalizes for external network factors by explicitly excluding specific events from the calculation:
- Pruned slots: Blocks rolled back due to chain-level dynamics, not operator failure. Excluded from
B_assigned. - Checkpoint verdicts: A proposer that builds a checkpoint is obliged to land it on L1. If it does not, that conduct adds one charge to the operator's
B_assigned, regardless of whether the next slot succeeds, is silent, or belongs to the same operator. A verdict may separately exonerate a later proposer that was provably active but had no settled state to build on; only that expressly exonerated slot is excluded from its operator'sB_assigned. A silent later slot keeps its ordinary miss. The charge and any exoneration are independent: no miss moves between operators, and chain-level miss counts, epoch records and network aggregates remain untouched. - Evidence limits on checkpoint attribution: A checkpoint charge or victim exoneration is applied only when the required checkpoint evidence was directly observed. Nodes retain the checkpoint history it depends on for roughly 15 hours and then destroy it, so a miss older than that window can never be re-examined. A miss with no verdict is therefore not evidence that no cascade occurred — only that no evidence survives. Attribution also applies solely to rollup eras that pipeline block building across slots; on Ignition and Alpha a proposer builds within its own slot, so a miss there cannot cascade. See the governance discussion.
- Attestations: Attestation participation is tracked via a separate metric. The score strictly measures block proposal success. Attestations have been meaningless as a metric since Alpha. Even if they do, the rollup does not value attestations. Until there are rewards for attestations, they will not be part of the calculations.
- APR / yield: The score does not factor in operator earnings. The APR also varies depending on luck.
Checkpoint Attribution Cases
Every outcome a missed slot can have. Only one case removes a slot from the score, and only one case adds a charge. Everything else is an ordinary miss.
An excluded slot is never also charged, so a cascade cannot chain down a run of misses. The charge and the exclusion land on different operators and never cancel, even inside one pool.
Cross-Era Consistency
As the Aztec rollup evolves (Ignition, Alpha, future upgrades), the scoring system adapts automatically. The network success rate and color thresholds recalculate continuously. All time scores span the full history across eras, while 90 day scores reflect only recent performance under current network conditions. Currently, the only metric we can utilize from the chain is block proposals. In the future, if attestations and many other metrics become of value for the chain, then we will add them to the scoring here and update the documentation here, but currently, that does not exist.