Once a feature is declared CONDITIONAL (orthogonal in the current regime), how do we get an automatic alarm when that declaration stops holding in live trading โ with a formal guarantee on the false-alarm rate under continuous monitoring? The required trust basis is an anytime-valid / Ville bound: false-alarm probability โค ฮฑ at every time t simultaneously, even though we peek continuously.
This question's deliverable is a buildable specification + validation plan, not a built monitor (building it is out of the evaluator's two write-surfaces, R3 โ operator). Synthesised from the campaign's literature SSoT (SOTA-RESEARCH-2026-06-23.md ยง2.1/2.2/2.4, METRIC-ADDITION-GATE.md ยงA) and the grounded-state record (STATUS-GROUNDING-AND-THRESHOLDS.md โ "G12 live expiry monitor: designed, not built"). Citations re-verified 2026-06-23 (15/15 real).
| # | Layer | Citation | Role | Input | Why this one (failure mode it covers) |
|---|---|---|---|---|---|
| 1 | WATCH โ weighted conformal test martingale | arXiv:2505.04608 (Prinster, Han, Liu, Saria, 2025) | Core drift monitor, covariate-shift-robust | the feature's conformal non-conformity scores under the validated model, per live bar | An unweighted CTM becomes invalid the instant the input distribution P(X) shifts โ the dominant crypto regime-change mode. WATCH re-weights to stay valid under exactly that shift. Strict superset of the unweighted CTM. |
| 2 | e-detectors โ composite quickest change detection | arXiv:2203.03532 (Shin, Ramdas, Rinaldo, 2022) | Localises when the verdict broke | the running martingale / score stream | A test martingale only says "drift happened sometime." e-detectors add changepoint localisation with an explicit detection-delay vs false-alarm tradeoff. |
| 3 | Group-invariance CTM | arXiv:2401.15461 (Lardy, Pรฉrez-Ortiz, 2024 / PMLR 2025) | Turns the static invariance verdict into a continuously-tested object | per-environment / per-regime conditional scores | Ties the alarm directly to the invariance alternative (the orthogonality property itself): martingale wealth grows โ invariance is breaking in deployment. |
One monitor instance per declared (feature, regime-scope) pair (a CONDITIONAL declaration is always regime-scoped). Data flow per live bar:
live bar โ feature value โ conformal non-conformity score (validated model) โ WATCH martingale (re-weighted for covariate shift) โ group-invariance CTM across the regime partition โ e-detector changepoint layer โ wealth Wt
Alarm fires when martingale wealth Wt โฅ 1/ฮฑ (e.g. ฮฑ=0.05 โ Wt โฅ 20). On fire: the feature's CONDITIONAL declaration is suspended pending re-validation, and the e-detector reports the estimated break time.
Anytime-valid (Ville's inequality). For a non-negative martingale started at 1 under the null (verdict still holds), P(โt : Wt โฅ 1/ฮฑ) โค ฮฑ โ the false-alarm probability is bounded by ฮฑ simultaneously across all t, under optional stopping. This is the property a batch re-test structurally cannot give: continuous "looking" does not inflate the error. WATCH supplies it under covariate shift; the group-invariance CTM supplies it for the invariance alternative.
The monitor must pass an offline replay on committed data before any live deployment:
| Check | Pass condition |
|---|---|
| Detection (sensitivity) | fires on known regime breaks (the committed 10-environment partition) with bounded detection delay |
| False-alarm (specificity) | empirical false-alarm rate โค ฮฑ over stable spans |
| Positive control / power | inject a known-expiring feature โ confirm it fires; feed a known-stable stream โ confirm it stays silent (the same shuffled-null discipline as G10) |
Status: E6 replay NOT run โ the monitor is UNBUILT. This matches the grounded state ("G12 designed-not-built"), which is precisely why the probe remains EVALUATING, not OPERATIONAL.
RESOLVED โ as a specification. Q4 asked for a live alarm with a formal false-alarm bound; the buildable design (3 layers + anytime-valid Ville guarantee + alarm rule + E6 plan) is the evaluator's deliverable and it is delivered. What is not done โ and is explicitly out of the evaluator's two write-surfaces (R3) โ is building and E6-validating the monitor; that promotes the probe from EVALUATING to OPERATIONAL and is an operator action.
Do not read this as "a live monitor now exists" or "any feature is being monitored." Nothing is built or running. The only claim is that the design is specified and its guarantee is sound by theorem (anytime-valid / Ville) โ a literature result, not a data result. No metric is promoted to a gate; no feature's status changes.
| Guard | Evidence |
|---|---|
| ClickHouse READ-ONLY (R1) | Read-only access present (--readonly=2); no query required โ literature + committed-artifact synthesis (0 queries, 0 writes). |
| Resource cap (R2) | No compute envelope launched (spec synthesis only; no measurement). Box well under the load1 โค 24 watchdog. |
| Leakage discipline (R7) | No new measurement โ leakage_guard.py N/A. The guarantee is a theorem (Ville), not a fitted number. |
| Write surfaces (R3) | Only this dashboard folder + PR #543 body. The monitor was NOT built (out-of-surface) โ operator. |
1/ฮฑ alarm threshold, then run the E6 replay above. Only after E6 passes does a CONDITIONAL declaration gain a trustworthy live expiry alarm.