Track Record

The Track Record page renders the canonical published history bundles for each chain — specifically the meta/<chain>/last90d.json artifacts. These are immutable published outputs of the regime classification pipeline, not recomputed or retroactively adjusted values. Each row corresponds to a single daily meta artifact identified by its as-of date, methodology_version, and determinism hash where applicable.

The page serves three analytical functions. First, it provides a regime frequency distribution over the selected window, which characterises the base-rate of each state for a given chain and period. This is the appropriate denominator for evaluating how unusual any current regime reading is. Second, the regime timeline and confidence history visualisations expose persistence and transition structure — how long regimes lasted, how often they flipped, and whether confidence was systematically lower in certain regimes. Third, the transition matrix quantifies the empirical regime transition probability distribution under the current methodology, which is directly relevant for any regime-conditional strategy development.

The regime mix bars are a frequency distribution of published regime labels over the selected window, expressed as a proportion of total published rows. Each segment width is count(regime) / total_rows for that chain and window. The "other" bucket captures any non-canonical labels that survived the pipeline without mapping to a standard vocabulary member — in a well-functioning pipeline this should be near zero.

Reading the frequency distribution analytically: the regime mix is the empirical base rate of each state under the current methodology. If STABLE accounts for 70% of published days on Ethereum over 90 days, that is the empirical prior probability of seeing STABLE on any randomly selected day in that window. This is directly useful for calibrating how unusual a current HEATING or CONGESTED reading is relative to recent history.

The regime timeline is a sequential visualisation of published label assignments, ordered ascending by as-of date. Its primary analytical utility is in exposing regime persistence and transition timing — properties that are invisible in aggregate frequency distributions but structurally important for understanding how the classification behaves in practice.

Two properties are worth inspecting analytically. First, run length: how many consecutive days a given regime label persisted before transitioning. Long runs in a single regime indicate structural stability in the underlying metric space; short alternating runs may indicate the chain is near a classification boundary, where small metric movements flip the label without implying a genuine regime change. Second, the timing of transitions relative to confidence trajectory: if a transition from STABLE to HEATING coincides with a visible improvement in confidence score, it is more likely to reflect a genuine evidence shift than a noise-driven boundary crossing.

Grey blocks (UNKNOWN/DEGRADED) in the timeline should be read as data voids, not as genuine regime labels. Their presence disrupts run-length inference for the adjacent named regimes — a run of STABLE days interrupted by a grey block does not necessarily imply a genuine transition to a new state; it may simply reflect a temporary data quality failure.

The confidence history chart exposes the temporal trajectory of the evidence-strength scalar over the selected window. It is the appropriate quality-weighting layer for any analysis that uses the historical regime labels — labels from high-confidence periods are more epistemically trustworthy than those from low-confidence periods, even if the labels themselves are identical.

Three structural patterns are analytically informative. Persistent low confidence across a chain for an extended period typically indicates a systematic data availability problem — either AWS source data was incomplete, or a metric space change affected coverage. Sudden drops in confidence that recover quickly may indicate transient data quality events. Confidence that is systematically lower for one chain than others in the same period may reflect chain-specific reporting delays or metric coverage differences.

For backtesting purposes, the confidence score is a necessary weighting factor. A naive backtest that treats all regime labels equally regardless of their confidence score will overweight UNKNOWN/DEGRADED periods (where no useful signal exists) and underweight high-confidence periods (where the signal is strongest). The correct approach is to condition the backtest on confidence band — ideally restricting analysis to rows where confidence exceeds the publication floor of 0.40, or better, restricting to the Good band (≥ 0.70) for highest-quality signal extraction.

The transition matrix is an empirical first-order Markov transition count table over the published daily regime sequence, computed per chain and then aggregated across chains for the selected window. Each cell M[i][j] contains the count of consecutive day-pairs where the regime transitioned from state i to state j. The diagonal M[i][i] counts self-transitions (regime persistence).

Dividing each row by its row sum gives the empirical transition probability matrix — the maximum-likelihood estimator of the transition probability under a first-order Markov assumption. This is the appropriate object for any regime-conditional strategy development: if you want to know the historical probability that HEATING is followed by CONGESTED, normalise the HEATING row and read off the CONGESTED column.

Two methodological caveats apply. First, UNKNOWN/DEGRADED transitions contaminate the matrix in a non-trivial way. A STABLE → UNKNOWN/DEGRADED → HEATING sequence is recorded as two separate transitions (STABLE→DEGRADED and DEGRADED→HEATING), not as one STABLE→HEATING transition. If degraded periods are frequent, the transition matrix will overstate the centrality of UNKNOWN/DEGRADED as a hub state and understate direct transitions between named regimes. Filtering to high-confidence rows before computing the matrix is the appropriate correction. Second, the matrix pools transitions across all four chains when "all chains" is selected, which implicitly assumes chain homogeneity in transition dynamics — an assumption that may not hold, particularly between L1s and L2s.

The table renders a chronologically ordered slice of the published meta history bundle. Each row is a daily Meta artifact identified publicly by chain, date, methodology_version, and named-row determinism hash where applicable. Historical corrections should be understood through public revision notices and dataset-level provenance, not through a single required integer field.

The methodology_version field is the critical comparability boundary. Rows with different methodology versions were produced under different classification rules and potentially different threshold parameters — naive comparison across a methodology version change boundary will conflate genuine network state changes with classification rule changes. The methodology changelog at documents all version transitions.

Public provenance should be read through methodology_version, updated_through, dataset revision or publication batch context, and determinism_hash where applicable. This is the canonical public provenance model for archived Meta outputs.

For reproducibility auditing, record the public identity tuple for the row in question, retrieve the same published artifact, and verify the determinism hash for named regime rows.

The interpretation boundary on this page is an extension of the product-level boundary: no price data, no forecasts, no advisory outputs. In the context of historical regime data, this boundary has an additional specific implication — the track record must not be treated as a backtested strategy performance record.

The regime labels published here are descriptive outputs of the meta layer at each date. They do not constitute signals in the trading strategy sense because: (1) they were not generated with a particular return objective in mind; (2) they are point-in-time published outputs that were available with a 1-7 day lag, not forward prices; and (3) the classification rules were not optimised against any outcome variable. Using these labels as the basis of a claimed backtest without those caveats would be methodologically inappropriate.

The legitimate analytical uses of this data are: regime frequency analysis, persistence and transition characterisation, confidence quality assessment, methodology version impact analysis, and — combined with independently sourced price or flow data — exploratory correlation research with appropriate statistical caveats.