›Navigation

Iteration 103 — THE HANDOVER OPERATOR DIRECTIVE ROW 112 · LOOP PARKED

What this page is: the operator asked (2026-07-14): finish all actionable metric evaluations, then hand over — every metric in plain words, the architectural decision record for live expiry, and the implementation flowchart — then stop. All five actionable batches are closed (57 of 68 feed candidates; the 11 in B-05 are gated behind the operator's own Frontier-#4 review). This page is the handover. The loop is now PARKED until further operator direction.

0 · The handover prompt (copy-paste to resume in any future session)

Resume the Matrix Admission Campaign (opendeviationbar-py, repo ~/eon/opendeviationbar-py, worktree .claude/worktrees/matrix-admission-loop, branch loop/matrix-admission-campaign-cycle-3, PR #599). GROUND FIRST: findings/evolution/audits/2026-07-02-matrix-admission-paradox-loop/ LOOP-PROMPT.md (hard rules: readonly=2 SELECT-only · 5c/5G capped systemd-run scope · spare-only + watchdog · no synthetic data · append-only · never merge PRs / message anyone / touch prod) · LEDGER.md (rows 0a–117) · FRONTIER.md · BONEYARD.md (60 rows) · ADR-2026-07-14-LIVE-EXPIRY.md · HANDOVER-PROMPT.md (this prompt's twin + the full state snapshot). STATE AT PARK (2026-07-14, iter 103): 16 campaign ADMITs → 23 certified gate slots · 45 kills · B-01/B-02/B-03/B-04/B-06 CLOSED · B-05 (11 items, EXPIRY boundary) GATED on the Frontier-#4 review · STABLE slot filled both substrates (e-ICP-fp) · FAIL boundary certified beyond G0 (PID-red #15, O-info #16 WATCH). THE OPERATOR DECISION QUEUE (work nothing until steered; every item has a ledger row): tri-tier BAN/WATCH/ORTHOGONAL ratification (row 105) · ADR-LIVE-EXPIRY ratification (4 open questions, §5) · Frontier-#4 review → unGATEs B-05 and the ADR's L3 sentinel · tail-invariance pattern ×5 (rows 104–109) · IID-CERTIFICATE LAW ×8 (rows 99/111/113/115) · dcor×SKIT-null matched-pair (rows 113–115) · inverted-mechanism ×10 (row 92 trail; latest 114/116) · RW+MCS pipeline (rows 101/103; BY re-entry rides it) · H-005 full-window re-proposal (row 108) · 03c wild-bootstrap gray zone (row 100) · M4-bar question (rows 80/86) · H-062 W-DRO park (row 76) · 73 cycle-1 draft statuses · STABLE-declaration proposal · boneyard counter reconciliation · PR #599 merge (operator milestone). ON RESUME: run ONE bounded frontier item per firing under the frozen M0–M5 ladder exactly as LOOP-PROMPT.md specifies. If the ADR is ratified: build expiry_tripwire.py per its §4. If B-05 is unGATEd: consume it cheapest-kill-first with the certificate attack mandatory. Otherwise: take the operator's steering verbatim and record it as a LEDGER NOTE row before acting.

1 · Every metric evaluated, in plain words — what problem each solves, what it grounds

Framework: the campaign certifies measurement instruments ("matrices") so that feature statuses — FAIL (redundant today) / FRAGILE / CONDITIONAL (works in some regimes) / STABLE (works in all regimes) — rest only on instruments that survived attack. Verdict grammar: ADMIT / INCLUDE-IF / EXCLUDE; default EXCLUDE; kills are wins.

1a · The 16 admitted instruments (+ 7 pre-campaign grounded slots = 23 certified)

#InstrumentIn plain words — the problem it solvesGroundsRow
1–2k_v verdict-stability counter (crypto + forex)Counts how often a feature's orthogonality verdict FLIPS as data accumulates — an unstable verdict is worth less than a stable one. The incumbent fragility meter.FRAGILE↔CONDITIONAL line15, 18
3–4e-ICP-fp (forex + crypto) — e-value invariant causal predictionAsks "does this feature's predictive relationship survive EVERY regime?" using evidence (e-values) that stays honest even when you keep watching it. Filled the stack's one vacant STABLE-identification slot on BOTH substrates.STABLE ceiling34, 35
5seqICPA second, environment-free way to corroborate invariance — different math, same question, so agreement means more.STABLE (corroboration)39
6IAS ancestryReads WHICH variables sit stably upstream of a feature (its causal ancestry) — invariance with a "who feeds whom" answer attached.STABLE (structure)41
7IPP distributional invarianceCatches features whose whole probability SHAPE morphs between regimes even when averages look calm — found 27 shape-shifters.STABLE (distributional)43
8HSIC-X anchor guardA validity guard: checks that the anchors (instruments) other tests lean on are themselves clean. A test built on a bad anchor is silently worthless.evidence validity45
9Causal DantzigAn invariance meter that tolerates HIDDEN confounders — reads regime-to-regime moment differences instead of assuming you observed everything.STABLE (confounder-tolerant)56
10Staiger–Stock FAnswers "are our anchors STRONG enough for any of this evidence to mean anything?" — weak anchors make invariance evidence uninformative noise.evidence strength64
11Anderson–RubinInference that stays exactly valid even when anchors ARE weak — the safety net under #10.evidence validity64
12Calibration gapMeasures how far a model's confidence drifts from reality per regime — the campaign's strongest forward-prediction lift (+0.0141).CONDITIONAL scoping75
13MMD kernel two-sample"Do these two regimes produce genuinely different feature distributions?" — kernel lens; admitted on a thin margin, cautions recorded.FRAGILE↔COND (heterogeneity)80
14Wasserstein-1 transportSame question as #13 through a transport lens — "how much probability mass must move to turn regime A's distribution into regime B's?" Clean 4× margin.FRAGILE↔COND (heterogeneity)86
15PID redundant share (MMI reduction)Of the information a feature and a shipped feature carry about the NEXT bar, how much is the SAME information? Redundancy about the future, not the present — the FAIL boundary finally gets a forward-looking meter.FAIL line116
16O-information (WATCH)For a feature group: does redundancy or SYNERGY dominate? (Synergy = information that only exists in the combination.) June-2026 WATCH flag travels.FAIL line (group-level)116
GG0, G2, G3, G5, G7, G10, G11The pre-campaign grounded stack (Spearman redundancy G0 among them) — inherited, not re-litigated, subject to the same live expiry once ratified.all statusespre

1b · The 45 kills + parks, grouped by the seven ways instruments die (each group = a reusable law)

Kill mechanism (plain words)Candidates killed
Broken certificate (IID-CERTIFICATE LAW, 8 members) — the instrument's "how big is chance?" formula assumes independent samples; market bars aren't. Fed provably-null data, they cry signal at 11–57% instead of 5%.HAC-t p-source (43.4%, #28 via BY) · Reduced TE χ² (16.2%, #38) · Hoeffding's D "exact" (56.6%, #39 — also a G0 duplicate) · τ* (11.1%, #40) · dcor block-perm (24.2%, #41) · dHSIC (39.1%, #42) · Effective TE shuffle (12.8%, #44) · copula entropy shuffle (39.4%, #45)
Knob-fragility (dial law) — the verdict changes when an arbitrary tuning knob changes; a measurement that depends on a knob nobody can justify isn't a measurement.PCMCI+ (α) · LPCMCI (α, worse) · EILLS (γ dead-weight) · CVaR-DRO (η knife-edge) · MRP (#25, argmin extremes) · RW-as-instrument (#29, coarse readout) · MCS (#31, degenerate at grid scale) · GW-CPA (#26, 0.868 — closest dial miss)
Duplicate gauge / echo — measures what an admitted instrument already measures (wrappers echo ≥0.95; genuinely different geometries don't).StabReg (ρ 0.977 vs e-ICP) · betting (0.982 vs MMD, #16) · energy (0.978 vs W1 sibling, #18) · V-REx (0.982 fate-transfer, twin of an executed readout) · Hoeffding's D (0.977 vs G0 — double death)
Regime-flat physics (TAIL-INVARIANCE PATTERN, 5 sightings) — the reading is REAL but identical in crash and calm: true physics, useless for regime-scoping. "The tails have one law; regimes live in the body."exceedance asymmetry (0.18×, #32) · extremogram clustering (0.16×, #33) · tail-λ (0.01×, #34) · EV-copula χ (0.08×, #35) · Hüsler–Reiss Γ (0.53×, #37 — proven an independent witness, echo 0.23)
Blind / degenerate / noise gauges — can't see what certified instruments see, or reads its own noise as structure.F014 (blind) · LGC-switch (#17, blind) · LGC measure (#19) · CRQA/recurrence (#20 — an autocorrelation gauge wearing a coupling costume) · congruence (#24 — within-regime halves disagree more than regimes do) · DM (#27 — t-normalization cancels the regime signal) · SPA (#30 — marginal blindness) · stability pair (#21/#22)
No incremental value at the admission bar — real, novel, and still useless for predicting which features stay orthogonal (some actively harmful, some inverted).GroupDRO (M4 harmful) · RBO (#23 — harmful + inverted) · SKIT (#43 — null machinery CERTIFIED as a standing datum, instrument inverted) · Sargan–Hansen J (M4 CI ≤ 0; kernel guard keeps the validity role)
Construction/substrate defects & near-misses — wrong for this data, or under-powered rather than wrong.DYNOTEARS (λ-load-bearing ground-truth bias) · CCM (deterministic-dynamics assumption vs stochastic bars) · Universal Inference (split-fragile) · Regime-MCI (#36 — 0.88×, the closest door-law miss; STRONGEST re-proposal) · W-DRO (PARKED — certificate radius void on dependent data until block-recalibrated)

Complete per-candidate reasons: BONEYARD.md (60 append-only rows) and LEDGER rows cited above. Kills #1–15 predate the B-04 numbering and are listed in their batch rows (B-01/B-03 in FRONTIER.md).

2 · ADR — Live Expiry (PROPOSED; full text in ADR-2026-07-14-LIVE-EXPIRY.md)

LayerWhat it doesMachineryStatus
L1 TIME FLOOREvery ADMIT expires at +90 days unconditionally; re-certification = replay the sealed exam on the newest window (cheap, capped, readonly).None new — the sealed exam scriptsready on ratification
L2 DRIFT TRIPWIREFreeze each instrument's admission CI as its certification band; a nightly capped job recomputes the statistic on a rolling window; 2 consecutive out-of-band checks ⇒ expire.None new — admission bands + combo-block bootstrapready on ratification
L3 REGIME-TURN SENTINELA substrate change-point monitor makes "next completed regime slice" DETECTABLE live; its firing expires every certificate scoped to the completed grid.B-05's 11 candidates (CTM-WATCH + e-detectors) — anytime-valid by requirement; the certificate attack is mandatory (IID law)GATED on Frontier-#4 review

Semantics: expiry never deletes — an EXPIRED(layer, reading, date) LEDGER row supersedes the ADMIT; the instrument re-enters FRONTIER Tier-3 with its sealed exam as the re-certification form; dependent status cells degrade assignable → DRAFT until re-admit; kills never expire. Four questions block ratification (§5): the 90-day floor · band width (verbatim CI vs widened) · the Frontier-#4 review · grandfathering of the 7 G-slots.

3 · Implementation flowchart

                       ┌──────────────────────────────────────────────────────────────┐
                       │  ADMIT(substrate, scope, role, expiry) — LEDGER row N        │
                       └──────────────┬───────────────────────────────────────────────┘
                                      ▼
              ┌───────────────────────────────────────────────────────┐
              │ REGISTRY entry (manifest.json reconciliation block)    │
              │ {admit_row, expiry_at=+90d, cert_band=[lo,hi],        │
              │  last_check, checks_out_of_band}                      │
              └───────┬───────────────┬───────────────┬───────────────┘
                      │               │               │
        L1 calendar   │   L2 nightly  │   L3 sentinel │  (inactive until a B-05
        clock ────────┤   tripwire ───┤   e-process ──┤   admission provides it;
        expiry_at hit │   reading exits│  crossing =   │   certificate attack
                      │   band 2× ────┤  "slice done" │   mandatory — IID law)
                      ▼               ▼               ▼
              ┌───────────────────────────────────────────────────────┐
              │ EXPIRY EVENT — any layer fires                        │
              │ • LEDGER row: EXPIRED(layer, reading, date)           │
              │   supersedes: ADMIT row N        (append-only)        │
              │ • status cells citing it: assignable → DRAFT          │
              │ • hub scoreboard decrement + dashboard page           │
              └──────────────────────────┬────────────────────────────┘
                                         ▼
              ┌───────────────────────────────────────────────────────┐
              │ FRONTIER Tier-3 re-entry                              │
              │ re-certification form = the SEALED exam script,       │
              │ newest window, no design drift (else new pre-reg)     │
              └──────────────┬────────────────────────┬───────────────┘
                             ▼                        ▼
                   PASS: re-ADMIT row            FAIL: BONEYARD row
                   (new expiry_at, new band;     (kill recorded; cells
                   cells re-promote to           stay DRAFT until a
                   assignable)                   replacement admits)

4 · Evaluation status at park

BatchBoundaryResult
B-01STABLECLOSED 10/10 — 6 admits · 3 kills
B-03STABLE (2nd)CLOSED 13/13 — 3 admits · 8 kills · 1 coordination · 1 park (W-DRO)
B-04FRAGILE↔CONDITIONALCLOSED 12/12 — 2 admits (#13, #14) · 10 kills
B-06CONDITIONAL-scopingCLOSED 12/12 — 0 admits · 12 kills · 2 durable products
B-02FAILCLOSED 10/10 — 2 admits (#15, #16) · 8 kills · 3 durable products
B-05EXPIRYGATED (11 items) — awaits the operator's Frontier-#4 review; feeds the ADR's L3

▶ Next iteration

Iteration 103 · 2026-07-14 · HANDOVER (row 117) — no verdict; deliverables: this page · ADR-2026-07-14-LIVE-EXPIRY.md · HANDOVER-PROMPT.md · loop PARKED · append-only · campaign at park: 16 ADMITs · 45 kills · 60 boneyard rows · 23 certified gate slots · PR #599 OPEN