Why does QAJAN separate the 'exploratory layer' from the 'preregistered validation layer'?

Any statistical relationship discovered after changing the event definition or pooling additional historical symbols is vulnerable to specification gaming. QAJAN mechanically tags historical 12-symbol pooled runs (n=148 clusters) with layer='exploratory' and enters_verdict=False. A hypothesis can only earn a confirmatory PASSED verdict after accumulating at least 100 independent post-registration events.

Why did hypothesis QH-20261013-06 pass Round-1 (AUC 0.6028) but fail Round-2A (AUC 0.5782)?

Round-1 used concurrent 1-minute (k=0) OLS beta. Because high-frequency trades are non-synchronous (the 1979 Epps effect), 1-minute concurrent beta underestimated true market beta by 4% to 11%, leaving lagged market propagation inside the residual series. When causal Dimson ±15 min beta and 5-minute Epps aggregation were applied, QH-06 AUC dropped to 0.5782 (<0.60) and 60-minute directional agreement measured 50.31% across 110,001 grid points (coin flip).

Why was ARGON-H4 withheld from production approval despite achieving AUC 0.7006 (32d) and AUC 0.6883 (114d holdout)?

Independent methodology auditing by Grok Assistant (findings M1-M6) identified four prerequisites: (1) ARGON-H4 was not yet preregistered in onkayit.jsonl, (2) its 60-minute refractory window was looser than the 4.5-hour tunnel benchmark (design effect deff = 1.959), (3) under the alternative O-sigma event definition its holdout AUC dropped to 0.5718 (winner's curse), and (4) its dOI/VWAP feature calculator had not yet been added to the 5-point look-ahead canary suite.

How does the 10x look-ahead canary test (asof_sizinti_denetle) detect future data leakage?

At 5 checkpoint timestamps across the 110,001-bar grid, future bars strictly after index t are multiplied by 10x, and Dimson regression windows are restricted to u < idx - k_lead. If any feature value or beta estimate at time t changes when future bars are perturbed, the pipeline raises a leakage violation and rejects the run.

QAJAN 6-Gate Hypothesis Factory: Preregistration, Dimson/Epps Orthogonalisation, and Falsification

Q: Why was ARGON-H4 withheld from production approval despite achieving AUC 0.7006 (32d) and AUC 0.6883 (114d holdout)?

Independent methodology auditing by Grok Assistant (findings M1-M6) identified four prerequisites: (1) ARGON-H4 was not yet preregistered in onkayit.jsonl, (2) its 60-minute refractory window was looser than the 4.5-hour tunnel benchmark (design effect deff = 1.959), (3) under the alternative O-sigma event definition its holdout AUC dropped to 0.5718 (winner's curse), and (4) its dOI/VWAP feature calculator had not yet been added to the 5-point look-ahead canary suite.

Q: How does the 10x look-ahead canary test (asof_sizinti_denetle) detect future data leakage?

At 5 checkpoint timestamps across the 110,001-bar grid, future bars strictly after index t are multiplied by 10x, and Dimson regression windows are restricted to u < idx - k_lead. If any feature value or beta estimate at time t changes when future bars are perturbed, the pipeline raises a leakage violation and rejects the run.

Exploratory Clusters (n = 148) vs Preregistered Validation (n = 6), How Epps Bias Falsified QH-06 (0.6028 → 0.5782), and Winner's Curse Auditing

Architecture & Measurement Record · QAJAN Hypothesis Gates + Statistical Auditing · October 19, 2026 · Author: Yusuf Aziz

Measured Evidence Summary (QAJAN Round-2A/2B & Methodology Audit)

Gate 1 — Exploratory vs Preregistered Layer Split: The 12-symbol pooled dataset (n1_pooled = 290, independent 60-min clusters n = 148, controls n0 = 870) is mechanically stamped layer="exploratory" and enters_verdict=False (passed_count = 0); post-registration validation (after 2026-10-13 13:05:33) has n_independent = 6 < 100, triggering UNDERPOWERED.
Gate 2 — Dimson ±15 min & Epps 5 min Orthogonalisation: Concurrent 1-min (k=0) OLS beta underestimated true market beta by 4%–11% across 10/12 symbols; QH-06 dropped from Round-1 AUC = 0.6028 to AUC = 0.5782 < 0.60 (Holm p = 0.0517) with 50.31% directional agreement across 110,001 grid points (coin flip) → FALSIFIED (DUSTU).
Sole Live Candidate (ARGON-H4) & Audit Hold: On canonical tunel37_escape events, 32-day AUC = 0.7006 (n1 = 154) and 114-day holdout AUC = 0.6883 (n1 = 232, p = 0.0005); gated by clustering design effect deff = 1.959 and naive repeated-peek Type-I rate 0.211 until ≥100 post-registration events accumulate.
Self-test & EVOLUTION-2 loop (autonomy.html): pytest 8/8 and 11/11 passed; system-wide 174/243 tools pass 3,129 self-test assertions, and the EVOLUTION-2 shadow loop stands at 194/240 episodes.

Introduction: Why agent-generated hypotheses fail in production

When an LLM agent is given historical time-series logs and instructed to propose testable quantitative hypotheses, it can generate dozens of mathematically plausible features in seconds. Yet nearly all such features fail out of sample.

Within the AtlasMaster autonomy architecture, the QAJAN 6-Gate Hypothesis Factory prohibits any generator agent from grading its own hypotheses and subjects every candidate to six sequential falsification gates. The October 14, 2026 Round-2A/2B audit (RAPOR.md and the independent KARNE_DENETIM_GROK_ASISTAN.md) documents how a seemingly significant Round-1 hypothesis collapsed once causal orthogonalisation was enforced.

1. The six-gate QAJAN architecture: Exploratory vs preregistered validation

No candidate feature can enter production without clearing six deterministic gates:

Gate 1 · Preregistration and Layer Separation: Before evaluation, the canonical event definition, refractory window (60 min or 4.5 hours), minimum independent sample size (n_independent ≥ 100), and effect threshold (AUC ≥ 0.60) are timestamped in onkayit.jsonl. Whenever an event definition or symbol pool is modified, the pipeline mechanically stamps the output as katman = "kesif" (exploratory), olay_tanimi_durumu = "OLAY_TANIMI_DEGISTI", and hukme_girer = False.
Gate 2 · 10× Look-Ahead Canary (asof_sizinti_denetle): Across the 110,001-point grid, future bars after index t are perturbed by 10× at 5 checkpoints, and Dimson regression windows are restricted strictly to u < idx - k_lead (dimson_beta_sizinti_yok = true).
Gate 3 · Causal Dimson/Epps Orthogonalisation: Common market factors must be stripped using lead-lag Dimson betas and multi-minute aggregation before testing residual events (Section 2).
Gate 4 · Cluster-Adjusted Permutation and Holm Family-Wise Gate: Effective sample inflation from cross-asset clustering (deff = 1.959) and multiple hypothesis testing (m_effective = 9) are corrected via Holm-Bonferroni; Monte Carlo simulation (10,000 runs) proved that naive repeated peeking inflates the Type-I error rate from 0.05 to 0.211.
Gate 5 · Mechanism Directional Alignment: A feature must not merely yield an AUC above 0.60; its theoretical sign must align with 60-minute forward orthogonalised returns above coin-flip probability (50%).
Gate 6 · Cross-Family Blind Methodology Audit: The generator agent (Argon) automatically recuses itself from judging its own report; an independent model family (Grok Assistant) audits the raw JSON and test scripts read-only.

2. The Epps effect and Dimson ±15 min orthogonalisation: How QH-06 collapsed

In Round-1, hypothesis QH-20261013-06 ("15-minute cumulative beta catch-up gap") achieved a raw AUC = 0.6028 (p_perm = 0.00125, Holm p = 0.01125) when residuals were computed using concurrent 1-minute (k = 0) OLS beta.

When Round-2A replaced concurrent 1-minute beta with causal Dimson (1979) ±15 min lead-lag beta and 5-minute Epps aggregation, QH-06 failed on two independent gates: (1) under Dimson ±15 min, accuracy fell to AUC = 0.5782 < 0.60 (Holm p = 0.0517, within-symbol Holm p = 0.2169, stratified AUC = 0.5503), and under 5-minute Epps ±2 bars to AUC = 0.5841; (2) across 110,001 grid points, 60-minute directional agreement with orthogonalised returns measured 50.31% (49.16% raw—a pure coin flip).

3. Round-2A/2B results across 11 hypotheses and the ARGON-H4 holdout analysis

Hypothesis ID	32d Exploratory `tunel37_escape` (`n1/n0`)	Raw / Stratified AUC	12-Symbol Dimson `±15m` / Holdout	QAJAN & Audit Verdict
`QH-20261013-06` (Beta catch-up gap)	157 / 471	0.6325 / 0.5913	Dimson `±15m` AUC 0.5782 · dir 50.31%	FALSIFIED (`DUSTU`) — Epps lag artifact
`ARGON-H1` (Kyle-lambda impact)	157 / 471	0.4758 / 0.4737	Pooled AUC 0.4895 (`Holm p = 1.0`)	FALSIFIED (`DUSTU`) — Below chance
`ARGON-H2` (Ownership bifurcation)	0 / 0	—	Missing historical surface (`n = 0`)	UNMEASURED — Deferred until data
`ARGON-H3` (Anti-Hawkes self-excitation)	157 / 469	0.5553 / 0.5396	Pooled AUC 0.5445 · sign flips across halves	FALSIFIED (`DUSTU`) — Fails Holm gate
`ARGON-H4` (VWAP dOI>0 anchor)	154 / 464	0.7006 / 0.6757	114d Holdout escape 0.6883 (`n1=232`) · O-σ 0.5718	SOLE LIVE CANDIDATE (`UNDERPOWERED`) — To preregister
`ARGON-H5` (Critical slowing down τ)	157 / 471	0.4307 / 0.4283	Pooled AUC 0.5075 (`Holm p = 1.0`)	FALSIFIED (`DUSTU`) — Noise level

Source: Projects/ATLAS/task_bus/ARGON_QAJAN_TUR2_1014/RAPOR.md:43-53 + KARNE_DENETIM_GROK_ASISTAN.md:13-20 (October 14, 2026).

4. Cross-family methodology audit (M1–M6) and falsification gate

Although ARGON-H4 demonstrated stable discrimination on canonical tunel37_escape events in both the 32-day window (AUC = 0.7006) and the 114-day out-of-sample holdout (AUC = 0.6883, split halves 0.6838 / 0.6878, p = 0.0005), the independent Grok Assistant audit (M1–M6) blocked premature approval on four grounds: (1) ARGON-H4 was not yet logged in onkayit.jsonl prior to observing the holdout; (2) under the alternative O-σ event definition, holdout AUC fell to 0.5718 (confirming winner’s curse across event definitions); (3) the 60-minute refractory window was 4.5× shorter than the 4.5-hour tunel39 benchmark (43/190 independent events).

Falsification gate: (1) Once ARGON-H4 is wired into the look-ahead canary and preregistered in onkayit.jsonl, if its forward validation over the first ≥100 independent events (or under the 4.5-hour refractory definition) yields AUC < 0.60 or Holm p ≥ 0.05, ARGON-H4 is permanently falsified.

Frequently Asked Questions (FAQ)

Why does QAJAN separate the 'exploratory layer' from the 'preregistered validation layer'?: Relationships found after changing event definitions or pooling historical symbols suffer from specification gaming; therefore QAJAN tags 12-symbol pooled runs (n=148) with layer='exploratory' and enters_verdict=False.
Why did hypothesis QH-20261013-06 pass Round-1 (AUC 0.6028) but fail Round-2A (AUC 0.5782)?: Round-1 used concurrent 1-minute (k=0) OLS beta, which underestimated market beta by 4%–11% due to the Epps effect; applying causal Dimson ±15 min orthogonalisation dropped QH-06 AUC to 0.5782 (<0.60).
Why was ARGON-H4 withheld from production approval despite achieving AUC 0.7006 (32d) and AUC 0.6883 (114d holdout)?: Grok Assistant's M1–M6 audit identified four blockers: missing onkayit.jsonl preregistration, a loose 60-minute refractory window (deff=1.959), holdout AUC dropping to 0.5718 under O-sigma, and missing look-ahead canary coverage.
How does the 10x look-ahead canary test (asof_sizinti_denetle) detect future data leakage?: At 5 checkpoint timestamps across the 110,001-bar grid, future bars after index t are multiplied by 10x; if any feature or beta estimate at time t changes, the run is rejected for look-ahead leakage.

References and Primary Technical Sources

Primary technical literature and open-source specifications supporting the architectural and experimental measurements on this page:

Related Technical Analyses: For the broader verification architecture, read autonomy, English overview, context-blind auditor and falsifiability gate, event-driven task_bus.jsonl coordination, and cephalometric smile design benchmark.