Each strip plots the per-ticker % difference (candidate − reference, relative to the paired mean). The solid line is the mean bias; the dashed lines are the 95% limits of agreement (bias ± 1.96·SD). A clean study has the solid line near zero, narrow LoA, and no visible heteroscedasticity.

MAPE / p95 are sample-level dispersion. CCC is Lin's Concordance Correlation Coefficient (1.0 = perfect agreement). Bias is the Bland-Altman mean; LoA is its 95% interval. Wilcoxon p tests H₀: median bias = 0 (significant at p < 0.01). Tolerance is a DCF-materiality threshold, not "acceptable noise". Agreement tier rule: HIGH requires CCC ≥ 0.99, |bias| ≤ 1× tolerance, ≥ 90 % of names within the tolerance band, and no significant Wilcoxon bias (p ≥ 0.01); MEDIUM relaxes to CCC ≥ 0.95 and |bias| ≤ 2× tolerance; otherwise LOW. The ≥ 90 %-within-band condition is a gate for HIGH, not a hidden criterion — a field can pass CCC/bias yet miss HIGH on it. Denominators: the outlier / %-delta columns divide by |reference|; the Bland-Altman strips divide by the paired mean — different questions, not interchangeable. Pairs whose reference is near zero relative to the field scale are split out (percentage meaningless) and counted, not silently dropped. Where a field carries a documented candidate-side artefact (e.g. multi-class share summation), an ex→TIER chip shows the tier with those names removed. The ex-known view is a scalpel for point artefacts, not a cure for systemic bias: where a field disagrees by a broad bias across the whole universe, removing a few named tickers leaves the tier unchanged — shares outstanding stays LOW on the full universe even after excluding every multi-class name, because the disagreement is systemic, not a handful of exceptions.

Continuous colour scale — saturation grows with |delta|. Green ≤ tolerance, amber tolerance–5%, red > 5%. Empty cell = the field is unavailable on one of the two providers (recorded as missing rather than penalised).

Each entry is a ticker × field pair where the candidate disagrees with the SEC EDGAR reference by more than 5 %. The most likely root causes are (i) fiscal-window misalignment between yfinance's vendor TTM and EDGAR's latest 10-Q sum, (ii) a restatement not yet ingested, or (iii) different treatment of stock-based compensation in FCF.