Forward-Orthogonality · feature-declaration expiry · robustness review

Feature-Expiry Robustness Assessment

Does a feature declared orthogonal / conditional today have a trustworthy way to know when that verdict expires in live trading? This reviews the design we already have (the G12 live monitor + the feature-status lifecycle) against what we need now — robustness as research.

2026-07-15 Existing design: EVALUATING — not operational Benchmark: campaign laws + SOTA

The one-line answer. The design is research-correct and honestly bounded — the grading side is ~9/10 and the guarantee (anytime-valid / Ville) is exactly right — but the live-expiry monitor is only ~4–5/10: it was designed before the campaign discovered the IID-CERTIFICATE LAW, so its raw conformal null would false-alarm on our autocorrelated bars like the 8 instruments that law killed, and it has never been power-validated (E6 not run). Robustness today is theoretical, not empirical.

0 · The lifecycle we're assessing

A feature is graded per regime through an early-exit gate ladder; which gate first disqualifies it sets the grade. Promoted features (CONDITIONAL/STABLE) ship with a regime scope and a live expiry monitor. When the monitor fires, the verdict expires and the feature re-runs the sealed ladder.

  feature × regime ──► GATE LADDER (G0 orthogonal-today · G3 regime-invariance ·
                       G2 forward-lift · G10 shuffled-null · G11 e-BH)
                              │ first gate to disqualify = the grade
      ┌───────────┬──────────┴──────────┬────────────────────┐
      ▼           ▼                       ▼                    ▼
    FAIL        FRAGILE             CONDITIONAL        STABLE-ORTHOGONAL
   (=BAN)      (=WATCH)             (=WATCH)             (=ORTHOGONAL)
                  └────────── promoted, scoped to regimes R ──────┘
                              │
                              ▼
             G12 / E6 LIVE MONITOR  (post-promotion, streaming)
             CTM/SKIT wealth martingale + BOCPD alarm, e-BH-merged,
             Ville bound:  P(∃t: W_t ≥ 1/α) ≤ α   →  "verdict EXPIRED"
                              │  fires (or: next completed regime slice)
                              ▼
             EXPIRE → feature reverts to DRAFT → re-run sealed ladder
                      → re-promote (new scope)  OR  demote

Grounded grading evidence (strong): persistence 94–95% into next regime · flip-ranking AUC 0.95/0.88 (crypto/forex) · ACI coverage 89.9%/79.9% · FDR-declared sets at q=0.10, PBO 0.00/0.20, deflation passed · ICP 0/92 ⇒ every declaration is regime-conditional · leakage-gated, twice-replicated. Honest ceiling: ~8/10 structural — unconditional invariance is provably impossible.

1 · Q1 — how robust is it? (the scorecard)

Grading a feature today

~9/10

Guarantee choice (anytime-valid / Ville)

~9/10

Live-expiry monitor (the actual expiry)

~4–5/10

Lifecycle automation (auto-expire + re-ground)

~3/10

Read: we can grade a feature and we picked the right kind of alarm — but the alarm itself, the part that says "this verdict has expired," is the weak link, and the automatic re-grounding depends on a detector we haven't built.

✔ What IS robust — keep

Right guarantee. Ville / anytime-valid: false-alarm ≤ α even under continuous watching, no stationarity assumed — the only guarantee that survives peeking.
Right aggregation intent. e-BH → one FDR-controlled, arbitrary-dependence decision.
Honest scope. Every verdict is regime-scoped; the ~8/10 ceiling is measured (ICP 0/92), never overclaimed.
Well-grounded grading. ~9/10 — leakage-gated, twice-replicated, FDR-bounded.
A certified-safe engine already exists. SKIT's betting e-process was proven valid on dependent bars (campaign row 114) — the fix is reachable.

✘ What is NOT yet robust — the gaps

DECISIVE F1 — raw conformal null on autocorrelated bars false-alarms 11–57% (the IID-CERTIFICATE LAW, discovered after this design).
DECISIVE F2 — E6 replay never run → power unproven ("no alarm" = "can't see it").
HIGH F3 — no attribution: can't tell feature-drift from substrate-drift → spurious mass-expiry.
HIGH F4 — multiplicity not closed at the per-monitor layer.
HIGH F5 — auto re-grounding needs an unbuilt regime-turn detector.
MED F6–F8 — covariate-weight/reference contamination · null-calibration rot · ~2K-bar under-power.

Why F1 is the load-bearing gap. The design's spine is exactly the machinery (CTM + SKIT + e-BH) that the later matrix-admission campaign put under attack — and it proved that any monitor whose "how big is chance?" null assumes independence cries wolf at 11–57% instead of 5% on market bars. The design (June 2026) predates that discovery (July 2026), so its default null is the one the campaign banned. Good news: the campaign also certified the fix (the SKIT betting e-process + circular de-alignment), so the hole is closable — but nothing gates until it passes the certificate attack.

The gaps in full

#	Gap	Breaks	Severity
F1	Generic CTM fed raw non-conformity scores from autocorrelated bars → 11–57% false-alarm (killed 8 instruments).	L-IID (dependence-aware null)	DECISIVE
F2	E6 replay never run — no positive/negative control; sensitivity + specificity unmeasured.	M5 blind-gauge	DECISIVE
F3	No attribution — feature-drift vs instrument/substrate-drift indistinguishable.	alarm fatigue / spurious re-cert	HIGH
F4	One α-monitor per (feature × regime) × hundreds → family-wide false-alarm explosion; e-BH not wired at the monitor layer.	L-IID / L-MAGIC (aggregation)	HIGH
F5	Interim "re-ground after next completed regime slice" needs a live slice-completed detector — unbuilt (B-05 gated).	auto-expiry inert	HIGH
F6	WATCH covariate weights estimated on ~2K windows; rolling CTM can self-heal toward the new regime.	estimation-error / drift-to-accept	MED
F7	The dependence-aware null's own inputs (block-length / tail-index) can drift; nothing re-checks them.	live L-IID	MED
F8	~2K bars/regime — effective sample size can be too thin; no under-power veto.	false negatives	MED
F9	wold_R crypto inversion (0.278) · forex PBO watch. (forex ICP wiring defect — later REPAIRED by the admission loop.)	grading hygiene	LOW

2 · Q2 — what specific robustness parts to add

Priority-ordered; each closes a numbered gap above.

Tier 1 — non-negotiable · close the decisive holes (F1, F2, F3)

Dependence-aware null on every monitor stream closes F1 — use the certified SKIT betting e-process and/or circular-shift / block de-alignment of the score stream. Nothing gates until it passes the certificate attack (≤10% false-alarm on de-aligned true nulls — the test that killed the 8 instruments).
E6 replay validation gate closes F2 — positive control: inject a known-expiring feature → must fire within a bounded delay; negative control: known-stable stream → must stay silent. Wrap in the leakage guard (T1–T4). Probe stays EVALUATING until it passes.
Attribution race closes F3 — two coupled betting e-processes: one watches the feature, one re-scores a sealed anchor regime to watch the substrate. Force-expire only on a feature verdict; an instrument verdict suppresses expiry.

Tier 2 — structural robustness · (F4, F5, F6)

e-value merge + e-BH at the monitor layer closes F4 — merge per-(feature × regime) monitors by weighted e-value mean (valid under arbitrary dependence) → e-BH → one family-wide FDR bound. Replaces ad-hoc "2-of-3 trip" voting.
Regime-turn sentinel + changepoint localization closes F5 — build the "slice completed" detector from the B-05 family (de-aligned + certificate-attacked) so auto re-grounding fires and re-cert replays the right localized slice. (needs unGATE-ing B-05 via the Frontier-#4 review.)
Fixed-reference conditional CTM closes F6 — freeze the reference to the sealed promotion slice so the window can't self-heal; feed whitened / block-standardized scores.
Null-cleanliness auditor — a pairwise-betting exchangeability test on the whitened stream, certifying de-alignment actually produced an exchangeable stream (guard against re-introducing the 8-instrument failure).

Tier 3 — lifecycle + small-sample robustness · (F7, F8, honest limits)

Null-calibration clock closes F7 — re-calibrate the null when its inputs exit their own block-bootstrap CI (matched-null, gate-legal).
ESS under-power veto closes F8 — withhold a reading when effective mass at ~2K bars is too thin, so a false negative isn't read as "still orthogonal."
Two-channel split — covariate-drift = telemetry that never gates; concept-drift = the sole expiry trigger.
Fail-closed + renewal-before-expiry overlap + append-only EXPIRED — un-recertifiable ⇒ revert to DRAFT; incumbent stays authoritative until re-cert passes; EXPIRED rows supersede, never delete.
Re-anchor at active-set swap — orthogonality is relative to the active set; re-run G2/G3 when the shipped set changes.

3 · Bottom line for Terry

Keep: the anytime-valid guarantee, the e-BH aggregation intent, the honest regime-scoped grammar, and the whole well-grounded grading side — these are SOTA-correct.

Fix before it can be trusted live, in order: (1) a dependence-aware null that passes the certificate attack, (2) the E6 positive/negative-control validation, (3) an attribution race so a regime turn doesn't spuriously expire still-valid features. Everything else is structural hardening on top.

The single sentence: our expiry design is right, but its default statistics were written before we learned that market memory breaks naïve tests — so the expiry monitor must be re-nulled for dependence and power-proven (E6) before any feature's verdict is allowed to expire on its say-so.

Research assessment · no code, no production change, nothing ratified · append-only. SSoT twin: findings/evolution/audits/2026-05-26-forward-orthogonality-prediction/FEATURE-EXPIRY-ROBUSTNESS-ASSESSMENT-2026-07-15.md. Sources: PHASE-3-PROBE-DESIGN.md (G0–G12) · STATUS-GROUNDING-AND-THRESHOLDS.md §2 · SOTA-RESEARCH-2026-06-23.md §2 · CAMPAIGN-GROUNDED-SUMMARY.md · probes/forward.html · matrix-admission LEDGER rows 99–117 (the two laws) · robust-expiry SOTA workflow wf_88195813-778. Citation-verification: pre-2026 backbone established; 2026-dated SOTA pending the standing arXiv re-verification gate before wiring.