The ETH-ARB-CCTP corridor in 2025: inventory of critical windows and reading of the five documented events

Research note, retrospective calibration. This article documents what an hourly substrate observation grid records on the Ethereum L1 and Arbitrum L2 chains, together with the CCTP bridge between them, over calendar year 2025. The grid is the production v2.0 API contract, reconstructed retrospectively from public on-chain data.

Three claims are not made and should not be read into the text. The article is not a hack detector: smart contract bugs, governance compromises and exchange breaches are out of scope by design, since substrate observability does not target application-layer code. The article is not a forecaster: the regime classification qualifies the substrate state at a given hour, it does not predict the next hour. The article is not a verdict on RWA criticality: it identifies hours where the substrate diverges from its 30-day nominal, but whether each divergent window is operationally consequential for a Real World Asset settlement flow is a separate question, addressed only by cross-referencing with documented academic Tier A sources.

Preamble: the matrix tested in this article

Every hour of the year is described by one combined cross-chain tuple. The notation is explicit:

Each of the four positions in the tuple is a discrete classification, calibrated per chain or per bridge direction, exposed in the production v2.0 API. Concretely:

• (Sx± . Dx±)_ETH is the Ethereum L1 signed regime: a 2-axis code where S is the structural axis (S1 nominal, S2+ slower than baseline, S2- faster than baseline) and D is the demand axis (D1 nominal, D2+ elevated, D2- depressed, D2± composition asymmetry). The S axis on Ethereum aggregates rhythm_ratio, continuity_ratio, beacon_participation. The D axis aggregates sigma_demand, size_demand, tx_demand. The 12 signed codes are S1D1, S1D2+, S1D2-, S1D2±, S2+D1, S2+D2+, S2+D2-, S2+D2±, S2-D1, S2-D2+, S2-D2-, S2-D2±. Per-axis thresholds are calibrated on a P97 / P2 quantile of a rolling 30-day window with multi-confirmation hysteresis (a regime flip takes about 3 hours to confirm).
• (Sx± . Dx±)_ARB is the Arbitrum L2 signed regime, same vocabulary. The S axis adds sequencer_publish_latency as a third structural observable. The D axis adds complexity_value and gas_complexity_ratio as L2-specific demand observables. sigma_demand on Arbitrum is structurally degenerate by the Nitro mechanism and is excluded from classification. Same 12-code vocabulary.
• BS_eth→arb and BS_arb→eth are the CCTP bridge states, one per direction. Each is binary: BS1 nominal, BS2 degraded. The classification reads the per-hour attestation_latency_p90_s against a threshold calibrated on a P97 quantile of a rolling 14-day window. The two directions are calibrated and classified independently.

What is not in this matrix, and is therefore not what this article tests:

• The composite drift.structural and drift.demand blocks (a separate primitive, addressed in the Delta calibration research note).
• The per-metric raw shift, shift_delta, shift_magnitude_delta continuous fields (these are inputs feeding the regime classification, exposed in diagnostic mode, but not themselves part of the classification tested here).
• The v3 precursors[] array (Primitive 3, chain-type-exclusive, exposed per chain after independent calibration; out of scope for the cross-matrix test).

The cross-chain reading rule applied throughout the article is disjunctive: an event qualifies as captured by the matrix when at least one of the four channels diverges substantially from its year-long baseline during the hot window. This rule is deliberate, not a calibration choice. A conjunctive rule (every channel must fire) would miss Pectra (Ethereum fires, Arbitrum stays nominal), USDe (Arbitrum fires, Ethereum stays mostly nominal) and BPO1 (the strongest channel fires only in the 24 hours following the window). The four channels are not redundant; they expose different substrate phenomena and combine by disjunction.

1. What this article rests on

The analysis uses an hourly journal covering 2025 that records, side by side, the state of Ethereum (the base chain), Arbitrum (the scaling rollup built on top of Ethereum), and the CCTP bridge that transfers the USDC stablecoin between them. Each row of the journal is one UTC hour. Out of the 8,760 theoretical hours of the year, 8,281 are populated, giving 94.5 percent coverage.

The 479 missing hours are not distributed at random. They concentrate on January-February (118 hours), November (111) and December (98). This is worth flagging honestly: the months when Fusaka and BPO1 occurred are also the months when observability is least complete. Practically, duration and coverage estimates for these two events should be read as lower bounds, not as exhaustive measurements.

Three families of observables drive the analysis:

• Thirteen continuous deviation indicators measuring, for each chain, the gap between the current hour and a reference behavior computed over the preceding days. Five indicators cover Ethereum, eight cover Arbitrum. One of them, measuring gas saturation on Arbitrum, is ignored throughout because the Nitro mechanism structurally pins it near 1.0 and it carries no signal.
• Two bridge health labels (BS1 nominal / BS2 degraded), one per direction, calibrated on the time an attestation takes to arrive. The label flips to "stressed" when this time exceeds a bound computed over the preceding fourteen days.
• Six raw measurements of CCTP attestation latency per hour: median, upper-decile, and highest decile, in each direction.

Three precautions apply end to end. January 2025 is excluded from the baseline because the rolling average reference had not yet stabilized at that point. Latency ratios are computed only for hours with at least five CCTP messages transferred, eliminating outliers from very small samples. The comparison reference for each month is the monthly median calculated outside windows of plus or minus six hours around the labeled events, preventing the events themselves from contaminating the baseline.

2. What counts as a critical window

To qualify a moment as "critical", at least two of the four observation channels below must light up simultaneously:

• On the Ethereum side, either one indicator that deviates very strongly from baseline, or at least two indicators that deviate significantly.
• On the Arbitrum side, the same rule applied to that chain's indicators.
• On the bridge side, an attestation latency at least fifty times the monthly median.
• On the bridge health side, at least one direction labeled "stressed".

A "critical window" is a period containing at least one hour that combines three channels or more, with these hours close enough to be considered part of the same episode (clustering interval of forty-eight hours).

With this criterion, 2025 produces eighteen critical windows, for a total duration equivalent to 1.7 percent of the year. Four match events documented by Tier A academic literature for tokenized Real World Asset flows. Fourteen are not tied to any officially documented event.

This second number requires explicit framing. "Non-documented" here means that no Tier A academic source within the scope reviewed for this study describes the window as a critical event for RWA settlement. It does not mean the window was discovered by the substrate grid alone, nor that it represents a hidden incident. Many non-documented windows likely correspond to market events that did not meet the institutional materiality threshold for tokenized assets, and are therefore filtered out by the specialized literature. The proportion of fourteen out of eighteen should be read as an upper bound on the gap between the substrate grid and the academic grid, not as a discovery rate.

Before presenting the inventory, two robustness questions must be addressed. The criterion above rests on five numerical thresholds (the intensity required on indicators, the latency ratio, the clustering interval). Does the result depend on these choices? And what would this count look like in a period without major events, to serve as a comparison point?

3. Is the result robust to the thresholds chosen?

To answer, each of the five thresholds is varied over a reasonable range and the critical windows are recounted in each configuration.

On nine of the ten configurations tested, the share of non-documented windows stays between 76 percent and 83 percent. The only configuration that breaks this pattern is the one where the clustering interval is extended to 72 hours, which absorbs the May and December events into merged super-windows. The principal result of the paper, that roughly four critical windows out of five do not match a documented event, is therefore stable and does not depend on particular calibration choices.

The additional exclusion of February (which sits in the rolling average stabilization zone) does not change anything important: 13 out of 17, or 76 percent. The figure holds with or without February.

4. Inter-arrival independence test

One might worry that critical windows cluster, which would suggest one event mechanically triggers others and would distort the count. This hypothesis is tested by examining the time separating consecutive windows.

The eighteen windows are on average eighteen days apart, with a median of twelve and a half days. The variability of these intervals is consistent with a pure random process (of the Poisson type): the statistical test does not reject this hypothesis, with an error probability well above the usual 5 percent threshold. Concretely, the "October cluster" of four windows in one month is statistically compatible with random variation and does not require a causal explanation.

This is methodologically reassuring: the non-documented windows are not the mechanical echo of the documented ones, they are independent observations. The Kolmogorov-Smirnov test on N=17 inter-arrival intervals has limited power, so this is best read as "absence of evidence of clustering" rather than as a strong claim that no clustering exists.

5. Four documented events, four distinct signatures

The four discrete events of 2025 do not produce the same image on the journal. This is itself a finding, because it shows that the three observation channels (Ethereum, Arbitrum, bridge) are not redundant.

Pectra, 7 May 2025. Ethereum hard fork modifying account and validator mechanics. On the journal, the signal during the window concentrates on Ethereum: one indicator exceeds eleven times its typical variation, the chain regime flips on 86 percent of hours. Arbitrum remains calm, the bridge too. This is the only one of the four events to manifest on a single channel.

USDe, 10-11 October 2025. Massive liquidation cascade on Ethena's USDe stablecoin. The mirror image of Pectra: Ethereum stays almost inert, Arbitrum presents the most violent deviation of the year (twenty-six times typical variation), and the bridge flips to "stressed" on the Arbitrum to Ethereum direction during 40 percent of hours. A bridge precursor is observable three days before the cascade.

Fusaka, 3 December 2025. PeerDAS activation, modifying blob data availability. This is the only event lighting up all three channels simultaneously at maximum intensity: Ethereum indicator at seventy-eight times typical variation, Arbitrum at eighty-four times, bridge latency at two hundred sixty-three times the monthly median. Seven indicators simultaneously exceed a high threshold. Interestingly, the Arbitrum peak arrives not during the official window but in the twenty-four hours following.

BPO1, 9 December 2025. Adjustment of the blob target parameter on Ethereum (from 6 to 10). During the official window, the signal is modest: three times typical variation on both chains. But in the twenty-four hours following the window, the bridge climbs to one hundred fifty-two times the monthly median latency, and Arbitrum reaches twenty times its typical variation. This is the only event where the post-window signature is stronger than the signature during the window itself.

These four images are not variations on a single theme: they are qualitatively disjoint. Pectra would not have been detected by a bridge sensor. USDe would not have been detected by an Ethereum-only sensor. BPO1 would not have been detected by a system that looks only at the official window. The operational conclusion is that a credible observation grid must combine the three channels with a disjunctive rule: it is enough that one channel lights up, not that all three light up together.

6. Bridge directions stress independently

Across the full year, 347 hours of stress are recorded on the bridge in the Ethereum to Arbitrum direction, and 357 hours in the Arbitrum to Ethereum direction. How many of these hours see both directions stress together? Ten. The chi-square independence test cannot distinguish this co-occurrence from a pure chance effect.

In other words, the two directions of the bridge stress independently. This is not a detail: it means CCTP must not be monitored as a single box but as two distinct queues, failing for different reasons. For a cross-chain routing system, this imposes two separate detectors, not one.

7. The April 2025 case: a bridge signal without a chain signal

Among the non-documented windows, April presents a specific signature that deserves separate examination. There are eleven distinct hours, spread over twenty-four days, where the bridge is stressed in the Arbitrum to Ethereum direction while both chains display nominal regimes. No deviation on Ethereum, no deviation on Arbitrum, no detectable departure from the demand or structural indicators. And yet the bridge signals an attestation delay between eighty-six and one hundred seventy-three times its monthly median.

This profile is unusual because in every other month, bridge stress is accompanied either by strong activity on the Ethereum side, or by a degraded mode on Arbitrum. In April, the bridge alone speaks.

A spontaneous hypothesis would be that this signature reflects the launch of CCTP V2 on Arbitrum, which took place on 2 May 2025. If April corresponded to residual congestion of the end-of-life V1 version, the signature should disappear after 2 May.

The test is easy to run. The hours where the "outbound Arbitrum stress on nominal substrate" signature occurs are counted per month.

The signature does not disappear after 2 May, it goes silent in May and June, then reappears with intensity multiplied by twenty starting in July. For the two months following the CCTP V2 launch, the signature is almost absent (one case per month). This quietness is itself a signal: right after the upgrade, the bridge performs better. But this improvement only lasts two months. From July onward, the number of "bridge-only stress" hours explodes and stays high until November.

Two readings are possible. Either the signature reflects institutional capital flows leaving Arbitrum toward Ethereum in a second-half seasonality, in which case it tells something about the RWA market, not the bridge. Or it reflects a degraded mode of the Circle validation pipeline that appears under certain load conditions, in which case it tells something about the bridge, not the market. The data provided here cannot decide alone. The most suggestive element is that the signature follows the same seasonality as the global bridge stress rate on Arbitrum outbound, which rises from 1.7 percent in June to 12 percent in October.

What is established, by contrast, is that the signature observed in April is not an effect of the V1 protocol version. If it were, it would disappear after 2 May and not return. It returns, and stronger.

8. Full inventory of the eighteen critical windows

The table below lists every critical window of 2025, with the main parameters. The right-most column indicates whether the window matches an academically documented event for RWA flows.

Two windows stand out by duration among the non-documented set. The 4 to 7 October window lasts seventy-one hours and carries bidirectional bridge stress. It precedes the USDe cascade by six days and has a latency intensity higher than Fusaka itself. It is not tied to any known Tier A publication, but its temporal proximity to USDe suggests it could constitute a market precursor signal, or reflect a second-order event that was not labeled.

The 1 to 3 February window, fifty-two hours, sits in the rolling average stabilization zone and must be read with caution. If excluded from the count, the totals become 17 windows and 13 non-documented, the same 76 percent ratio.

9. Limitations and reading

Three limitations must be kept in mind.

First, the BS1 / BS2 calibration used here is a research calibration, based on a rolling percentile on reconstructed attestation latency. It is not equivalent to the production calibration that uses a Circle Iris probe in milliseconds. Conclusions should be read as retrospective, not as an audit of the production calibration.

Second, the concentration of missing hours in November and December (209 hours out of 479) imposes that duration estimates for Fusaka and BPO1 be considered as lower bounds. If some missing hours corresponded to stress moments, the count under-estimates them.

Third, the principal ratio of four critical windows out of five not being documented should be read as an upper bound on the gap between the substrate grid and the academic grid. A share of the non-documented windows likely corresponds to market events that did not meet the institutional materiality threshold for tokenized Real World Assets, and are therefore filtered by the specialized literature. Precisely quantifying this share would require systematic cross-referencing with regulatory announcement calendars, centralized platform incidents, and protocol publications, which falls outside the scope of the present work.

10. What this work shows

The observation grid built from Ethereum, Arbitrum and the CCTP bridge detects, in 2025, a class of operational events that the Tier A academic literature for Real World Asset flows does not capture by construction. The detection is robust to calibration choices across a wide range of threshold variation, and the distribution of windows in time is consistent with a random process, which rules out a contagion artifact.

The four documented events each express themselves through a qualitatively different signature, which justifies combining the three observation channels rather than aggregating them into a single score. The two directions of the CCTP bridge stress independently and must be monitored separately.

The signature observed in April 2025 on the Arbitrum outbound direction suggests a bridge regime that goes silent during the two months following the CCTP V2 protocol launch, then returns with strongly amplified intensity from July and persists through November. This behavior cannot be attributed to the V1 to V2 migration and requires other explanations, the most plausible hypotheses being either an RWA capital-flow seasonality or a conditional degraded mode of the attestation pipeline. Choosing between the two requires cross-referencing with data outside the substrate, which is not performed here.

The strongest operational takeaway is the disjoint signature finding: Pectra would not have been detected by a bridge sensor, USDe would not have been detected by an Ethereum-only sensor, BPO1 would not have been detected by a system that looks only at the official window. A substrate observability grid that aims to capture the operational substrate events documented for RWA flows must combine multiple channels with a disjunctive rule, and must read the post-window tail as carefully as the window itself.

Methodological note

Latency ratios compare the upper-decile attestation latency of a given hour to the monthly median computed outside labeled windows. The unit "deviations from typical variation" used in the text refers to a robust z-score computed on the median and median absolute deviation of the baseline. The independence test between the two directions of the bridge is a chi-square test on a contingency table, with a probability value of 0.86. The inter-arrival test is a Kolmogorov-Smirnov test against an exponential distribution, with a probability value of 0.62.