CMM.X Observer
predictive_claim = NONE · Diagnostic + verification + history layer. predictive_claim = NONE. No edge, no alpha, no direction, no trading signal. Under prospective test.
Wohin · whither — learning, honestly

Learning skill measured by a proper scoring rule — not a forecast.

Mechanism validation — historical walk-forward is SUGGESTIVE, not proof. Skill measured by proper scoring rule (log-loss / Brier) vs climatology & persistence baselines. Not a forecast, no edge, no direction.

T1–T4 learning targets as-of

T1 · SHAPE
SUGGESTIVE
Beats baselines on log-loss (calibration, not accuracy). Not proof.
T1_VOL_REGIME h1: Brier-skill vs climatology +0.064; accuracy≈climatology (calibration win, not accuracy).

source: reproduce_leg/…/SEER_LEARNING_REPORT.json

T2 · STRUCTURE_REGIME
TIES_CLIMATOLOGY_ACC · WINS_LOGLOSS
Ties climatology on accuracy (baseline-easy) — the real win is calibration (log-loss).
mean LEARNED acc 0.9518; Δacc vs climatology -0.0003; Δlog-loss vs climatology -0.0168; vs persistence -1.6793.

source: reproduce_leg/…/LEARN_T2_REPORT.md

T3 · ENERGY_STRESS
REFUTED_DEGENERATE
Refuted as defined — near-constant target (HIGH band never realized); no model can express skill.
Δacc +0.0002; Δlog-loss vs climatology -0.0020; degenerate single-class runs 8/20.

source: reproduce_leg/…/LEARN_T3_REPORT.md

T4 · ELEMENTS
MAJORITY_BEATS_CLIMATOLOGY
22/30 elements beat climatology on log-loss (73.3%). Strongest signal in the set — still suggestive, not proof.
accuracy secondary 17/30; mean ΔLL +0.0872, median +0.0628; OI-intensity cells (tiny n) lose and are shown.

source: reproduce_leg/…/LEARN_T4_REPORT.json

Convention: a proper scoring rule (log-loss / Brier), lower is better; skill = baseline − learned. T2 accuracy ties climatology because the regime is near-constant (baseline-easy) — the win is calibration, not accuracy. T3 is refuted-degenerate (its a-priori terciles make the target near-constant) — refutation is reported plainly. A calibration improvement over a baseline is not a directional call and is not a trading signal.

Calibration vs baselines (SEER) as-of

Calibration, not accuracy. Skill measured by Brier / log-loss vs climatology & persistence baselines. Not a forecast.

Targethorizonscoredlearned Brierclimatologypersistenceskill vs clim.skill vs pers.
T1_VOL_REGIME 5 3258 0.2783 0.3206 0.6157 +0.0423 +0.3374
T2_STRUCTURE_REGIME 5 3258 0.0899 0.0885 0.1488 -0.0014 +0.0588
T3_ENERGY_STRESS 5 3258 0.0196 0.0205 0.0307 +0.0009 +0.0110
T1_VOL_REGIME 3 3260 0.3119 0.3688 0.6221 +0.0569 +0.3102
T2_STRUCTURE_REGIME 3 3260 0.0860 0.0880 0.1389 +0.0020 +0.0529
T3_ENERGY_STRESS 3 3260 0.0194 0.0201 0.0282 +0.0006 +0.0088
T1_VOL_REGIME 1 3262 0.4055 0.4692 0.6830 +0.0638 +0.2775
T2_STRUCTURE_REGIME 1 3262 0.0765 0.0874 0.1020 +0.0109 +0.0255
T3_ENERGY_STRESS 1 3262 0.0174 0.0196 0.0196 +0.0021 +0.0022

Historical walk-forward — mechanism validation, suggestive only, NOT proof of skill.

Sigma-sync episodes — edge verdict FRESH

SIGMA_SYNC_EDGE_0_NOT_CONFIRMED
independent episodes
46
beats / ties / worse
18 / 1 / 27
beats_rate
0.4
binom_p
0.9324
diagnostic only
YES

crypto assets ~0.57 correlated -> effective independent episodes < raw count; proxy X

Prospective pre-registration ledger FRESH

A predictive claim would require many pre-registered, out-of-sample confirmations accumulated over real elapsed time. None exist yet.

sealed (future-target, append-only)
95
outcomes recorded
23
confirmed out-of-sample
22
inconclusive / pending
0

outcome tally: SUPPORTED: 22 REFUTED: 1

A single SUPPORTED result is NOT a skill claim; pending seals are shown as INCONCLUSIVE_PENDING until their target time elapses.

source: reproduce_leg/…/prereg_ledger/prereg_ledger.jsonl · reproduce_leg/…/prereg_ledger/prereg_outcomes.jsonl