Do market regimes help short-horizon direction models?

• values as written in the Chapter 4 claims ledger
• populations reported separately, never pooled

Direction-model figures exactly as the Chapter 4 results-claims ledger of the CP2 report draft records them: the regime-adaptive model under its original whole-sequence regime labels, and the unchanged global model relisted from its saved metrics, for each instrument on the development folds (equal mean of the four folds, as the file gives it), the fixed holdout and the historical fresh period. Populations are named as the ledger names them; the ledger rows give no calendar dates. Exploratory research results, not a trading recommendation and not investment advice.

• sensitivity analysis of a stated limitation; exploratory; not a new model, not confirmatory
• primary variant: causal_online_viterbi_cold

The original regime labels were decoded over whole sequences, which can use bars after the row being labelled. MR-01 re-decoded them causally (online, no later bars) and compared the two. For each instrument and population the table gives the replication-gate status and the share of rows on which the causal labels equal the whole-sequence labels, as written in the metrics file. Populations are not pooled and instruments are not averaged.

• synthetic method verification only; says nothing about real markets

Synthetic method verification only; says nothing about real markets. The CP2 hidden-Markov configuration was fitted to simulated series whose true regimes are known, over a grid of regime counts, persistence, separation and series length. Recovery is the adjusted Rand index (ARI) between true and fitted states, where 1 is exact. The chart places every grid cell's stored mean ARI by regime separation, for whole-sequence and causal decoding.

• design-sensitivity analysis from committed summaries; approximate; not a re-analysis of the data

How small an effect the committed TimesFM and EDL designs could have detected, computed from their committed summaries. It does not re-test any hypothesis, does not change a committed result and does not say that a result would change with more data; sessions-needed figures are planning magnitudes only.

• containment_qualified: false
• exploratory, descriptive; opened periods; not confirmatory
• evidence_label: reported

Does one pinned TimesFM 2.5 checkpoint forecast the six-bar-ahead endpoint better than assuming no change (last-close persistence)? The primary measure is the paired MAE difference, model minus persistence, in the holdout and fresh regions, reported separately. A positive value means the model was worse. The bootstrap interval is exploratory, not a confirmatory confidence interval.

• containment_qualified: false
• EDL E2 development-period metrics written by the cp2_edl_runner_v1 E2 runner on the owner's PC; exploratory, assumption-conditional diagnostics; not confirmatory
• evidence_label: reported

Exploratory, assumption-conditional diagnostics of the EDL trading-rule candidates on the development period only, in paper-money units (fraction of a notional K). Contrast K01 compares candidate C02 with C01. No holdout or fresh EDL result exists in a committed file. Not confirmatory; no study-wide error rate is certified.