Architecture Decision Record · for Terry to refine

Live Orthogonality Declaration Engine

The grounded matrices are the graders. This engine is the outcome: for each live market regime it runs those matrices over every shipped feature and declares each one BAN / WATCH / ORTHOGONAL — published live, alerted on Telegram, with an append-only expiry trail.

2026-07-15 PROPOSED — for Terry's steer → implement Quality bar: 10/10, no defective version

Context — the forward-predictiveness research this rests on

This engine is the operational face of a multi-month research arc with one question: "a feature is orthogonal today — will it still be orthogonal tomorrow, and how do we declare that?" The chain you're reviewing:

In one line: the forward campaign proved the declarations are predictive, the admission campaign built and proved the graders, and this ADR ships them live. The full trail is in the provenance appendix at the bottom.

The three tiers (grounded matrices behind each)

Decisions locked with you (2026-07-15)

How it runs (five layers)

  ① REGIME LAYER (the clock)
     live_tail rolling window → classify archetype (crash/bull/bear/range/parity/parabola) + session
     cheap always-on regime-turn watcher  ── fires ──►  full re-declaration
          (B-05 detector, de-aligned + certificate-attacked; daily-floor fallback until B-05 unGATED)
                                   │
                                   ▼
  ② DECLARATION LAYER (cheapest-first, early-exit, per feature × live regime)
     BAN screen (cheap) ─► WATCH check (medium) ─► ORTHOGONAL check (heavy chain)
       └── assign the HIGHEST tier the evidence earns; thresholds operator-calibrated
                                   │
                                   ▼
  ③ VALIDITY SPINE  (mandatory — what makes it 10/10)
     every live null is DEPENDENCE-AWARE (block/circular de-alignment OR the certified
     SKIT betting e-process). No declaration publishes unless it passed the certificate attack.
                                   │
             ┌─────────────────────┴─────────────────────┐
             ▼                                             ▼
  ④ PUBLISH                                       ⑤ NOTIFY (Telegram, @obdpy_bot)
     living dashboard board  +                     regime change  →  board summary
     append-only CH table                          tier flip       →  ORTHOGONAL→WATCH/BAN = EXPIRY
     (expiry history queryable)                     (design only — nothing sent without your go)

Resource efficiency (first-class)

Before it's authoritative — the 10/10 gate

Where your steer would help most

You've already set the direction — these are the spots where your opinion and suggestions would sharpen it.

1. 1 Crypto ORTHOGONAL gap — it rests on one instrument (e-ICP-fp) vs forex's nine. Ship with a single-witness flag, or grow the crypto leg first? Our lean: ship as ORTHOGONAL⁻ (flagged) and keep growing the crypto leg.
2. 2 Prioritise the live regime-turn detector? Building the B-05 change-detector family is what lets the board auto-refresh on real regime turns (vs a daily-floor fallback) — worth pulling forward? Its gate is ours to lift; your read on priority is what helps.
3. 3 Anything you'd add — SOTA angles, robustness checks, or scope — to make the live engine stronger.

On our plate before go-live (operator tasks — for visibility, not sign-off)

• Threshold calibration per tier — set each tier's cut-lines on the census distributions (banlist method: compute → evidence → tiers → declare), every line tagged anchored vs chosen. Not done yet.
• unGATE B-05 via the Frontier-#4 review (operator gate) — the precondition for the regime-turn detector.
• Cadence floors (BAN hourly · full suite on-turn + daily) · Telegram channel (dedicated vs ODB-ops) · crypto feature backfill (bar_* entropy/fractal are NULL in cache) · resource envelope (~2c/2G always-on; heavy suite spare-only).

One paragraph for Terry

Provenance & telemetry — the full trail (everything we have so far)

So you review with the exact view. Dashboard = living pages, deployed at odb.epatner.com (clickable). Audit / telemetry = the append-only forensic record + code + raw JSON on GitHub main (terrylica/opendeviationbar-py).

① Dashboard — matrix evaluation

② Dashboard — harvesting

③ Dashboard — forward-predictiveness + supporting probes

④ Audit folders — the append-only forensic record (GitHub main)

⑤ Telemetry artifacts — the raw evidence

PROPOSED · nothing ratified, nothing deployed, no Telegram sent · append-only. SSoT: findings/evolution/audits/2026-05-26-forward-orthogonality-prediction/ADR-2026-07-15-LIVE-ORTHOGONALITY-DECLARATION-ENGINE.md. Sources: tri-tier proposal (LEDGER row 105) · the 23 grounded matrices · the two laws (rows 99–117) · COMBINED-REGIME-SLICING.md · notify/telegram.py · robustness assessment 2026-07-15.