Objective
Test whether replacing the Exp 09 coarse 6-point grid splice with a per-period integer scan (h in [40,220], step 1) reduces pricing error, holding the geometric tail operator and asymptotic decay rate source fixed.
Setup
- Model: 54 free params; tail operator eta_H * r/(1-r) with global r from MCDecayTable (frozen)
- Evaluator: 3-arm bakeoff (F150, G09, O1) with CRN within each cell
- Comparators: G09 at Exp 09 locked thresholds; F150 fixed baseline
- Acceptance: Phase A 3-gate proceed (G1>=0.05, G2>=0.10, G3>=1e-3); Phase B 5-gate checkpoint (C1-C5)
Procedure
- Phase 0: decontaminated 7-theta panel with SHA-256 dedup (4 fixed + 3 rule-selected slots)
- Phase A: Stage-0 oracle sweep (336 cells x 5 seeds x 181 H values) + 3-gate proceed/kill
- Phase B: 125-triple calibration over multiplier grid, lock thresholds into o1_locked_policy.json
- R3 pilot bakeoff: 336 cells x 2 seed schedules x R in {250,500} x S=5 x 3 arms = 20,160 rows
- R3 checkpoint: 5 gates (C1-C5) at R=500 on both seed schedules, any fail = STOP_WRITEUP
- Audit: seed fragility (6 fresh seeds), R fragility (R=250 to R=2000), per-theta breakdown
Results
- Phase A: G1=0.178 (3.55x), G2=0.545 (5.45x), G3=1.94e-3 (1.94x) — PROCEED at R_prod=250
- Audit fragility: 2/6 seeds KILL G3; G3 collapses 11.4x from R=250 to R=2000 (4x faster than sqrt(8) MC noise)
- Calibration winner: (m_rho, m_fit, m_hold) = (0.75, 2.0, 1.0) * anchor; J*=2.2125
- J objective dominated: 0.10 * median_oracle_gap ~ 2.20 vs median_E_logQ ~ 0.012 (180x ratio)
- C1 FAIL at R=500 on both schedules: ratio_A=1.1072, ratio_B=1.0799 (O1 8-11% worse)
- C2 PASS (consistent sign); C3 PASS (O1 ~10% faster); C4 PASS (10x lower chatter); C5 PASS (10x lower fallback)
- Verifier: bit-identical on all 9 calibration fields and 20,160 pilot rows (max float drift 4.26e-14)
- Decision: STOP_WRITEUP_KILL_DETECTOR_FAILS
- No production code modified; F8/F9/F10/F12 all unfired
Analysis
The central finding is that the integer-scan oracle gap closes faster than MC noise as R grows (11.4x shrinkage vs expected sqrt(8) ~ 2.83x), meaning the apparent quantization benefit at low R is noise-driven, not structural. The integer scan has 181 candidates vs the grid’s 6, so its finite-R argmin is more likely to hit a noise trough. The pre-registered J objective silently became an oracle-matcher (180x dominated by the oracle-alignment term), so the locked O1 thresholds optimize for displacement from the Stage-0 integer oracle rather than for E_logQ directly. The operator-floor null N3 — that the geometric tail operator eta_H * r/(1-r) with a single global r is the real bottleneck — is the leading follow-up hypothesis, consistent with Exp 09’s 87.7% non-monotonicity.
Claim updates
- principal-eigenvalue-determines-long-run-asset: Integer-scan splice (181 H values) cannot improve over coarse-grid G09 (6 values) at R=500, because the apparent quantization benefit is noise-driven (oracle gap collapses 11.4x as R grows 8x). This supports the interpretation that the geometric tail operator’s single global decay rate — not the discretization of the splice horizon — is the binding accuracy bottleneck.
Follow-up
- Operator-floor null N3: test state-dependent decay r(x), log-linear two-term tail, or regime-dependent r_k(theta)
- Re-weight calibration objective J (normalize oracle gap by H_range or use price-based penalty)
- Stage-0 oracle at R>=1000 needed for a stable G_floor baseline