Finance Coupling Kill

Every position is compliant. The portfolio is not.

Two AI mandates allocate independently. Each passes its own position limit and stress-liquidation check. A coupling oracle checks the composition — and finds a constraint violation that no single mandate could detect.

Axiom

Liquidity is not additive. Market depth is shared.

Scenario

AI Strategy Engine proposes 8% NAV to a semiconductor ETF. AI Liquidity Optimizer proposes 12% NAV to the same asset. Neither sees the other. Neither sees the existing 3% correlated exposure. Combined: 23% into a 15% cluster limit.

PHASE 1 Position Oracle — individual asset limits

Subsystem A — Strategy Engine

Semiconductor ETF · 8.0% NAV

COMPLIANT

Position 8.0% NAV

Limit 10.0% NAV

Utilization 0.80

Subsystem B — Liquidity Optimizer

Semiconductor ETF · 12.0% NAV

COMPLIANT

Position 12.0% NAV

Limit 15.0% NAV

Utilization 0.80

PHASE 2 Liquidity Oracle — per-subsystem stress liquidation

Stress Liquidation — Subsystem A

5% NAV/day normal · 2.5% stress

COMPLIANT

Normal 1.6 days

Stress 3.2 days

Limit 5 days

Stress Liquidation — Subsystem B

5% NAV/day normal · 2.5% stress

COMPLIANT

Normal 2.4 days

Stress 4.8 days

Limit 5 days

All four individual checks pass. Each subsystem is within its position limit. Each can liquidate under stress within 5 days. But neither sees the combined portfolio. Neither sees the correlated cluster.

PHASE 3 Portfolio Coupling Oracle — compositional check

Coupling Oracle

Cluster: TECH_SEMICON

UNACCEPTABLE

Proposed (A+B) 20.0% NAV

Existing corr. 3.0% NAV

Combined 23.0% NAV

Limit 15.0% NAV (policy)

Failure mode CONCENTRATION_BREACH

Deciding CONCENTRATION_BREACH — 23.0% / 15.0%

Observed Shared exit 8.0d / 5.0d — also breached, not deciding

Utilization

1.53

153% of cluster limit

Shared exit capacity

8.0 days

Also breached (not deciding)

Check	Result	Detail
Position A (Strategy)	COMPLIANT	0.80 util
Position B (Liquidity)	COMPLIANT	0.80 util
Liquidity A (stress)	COMPLIANT	3.2d / 5d
Liquidity B (stress)	COMPLIANT	4.8d / 5d
Portfolio Coupling	UNACCEPTABLE	1.53 util

✗ EXECUTION BLOCKED

Correlated cluster TECH_SEMICON combined exposure 23.0% NAV exceeds 15.0% limit (utilization 1.53). Individual positions are each within single-asset limits, but the composition violates the portfolio constraint.

This failure exists only at the composition boundary. No amount of internal alignment surfaces this risk.

↓ Evidence pack (JSON)

Control case

Same architecture, smaller positions. 3% NAV from subsystem A, 2% NAV from subsystem B. Combined with 2% existing correlated exposure: 7% into a 15% limit. The system does not reject by default.

PHASE 1–2 Individual checks

Position + Liquidity — Both subsystems

3.0% + 2.0% NAV

COMPLIANT

Position A 3.0% / 10.0% — COMPLIANT

Position B 2.0% / 15.0% — COMPLIANT

Liquidity A 1.2 days stress / 5d limit

Liquidity B 0.8 days stress / 5d limit

PHASE 3 Portfolio Coupling Oracle

Coupling Oracle

Cluster: TECH_SEMICON

COMPLIANT

Combined 7.0% NAV

Limit 15.0% NAV

Utilization 0.47

Shared exit 2.0 days stress / 5d limit

All constraints satisfied. The system does not reject by default. It rejects when composition violates constraints. Same architecture. Same oracles. Different outcome because the inputs are different.

✓ EXECUTION PERMITTED

Combined cluster exposure 7.0% NAV within 15.0% limit. Shared exit capacity 2.0 days within 5 day limit.

All individual and compositional constraints satisfied. Deterministic evidence pack generated for audit trail.

↓ Evidence pack (JSON)

Scenario — Different failure mode

Concentration limit is generous (30%). But market depth is shallow. AI Alpha Engine proposes 7% NAV to an illiquid biotech. AI Momentum Engine proposes 6% NAV to the same asset. Each can exit alone under stress. Together, they cannot exit through the shared market depth of 2% NAV/day within the policy timeline.

PHASE 1 Position Oracle

Subsystem A — Alpha Engine

Biotech ETF · 7.0% NAV

COMPLIANT

Position 7.0% NAV

Limit 10.0% NAV

Utilization 0.70

Subsystem B — Momentum Engine

Biotech ETF · 6.0% NAV

COMPLIANT

Position 6.0% NAV

Limit 10.0% NAV

Utilization 0.60

PHASE 2 Liquidity Oracle — per-subsystem stress

Stress Liquidation — Both subsystems

2% NAV/day stress volume

COMPLIANT

Subsystem A 3.5 days stress / 5d limit

Subsystem B 3.0 days stress / 5d limit

Concentration is fine — 13% is well under the 30% limit. But concentration is not the only coupling constraint. Market depth is shared. Both positions compete for the same exit capacity.

PHASE 3 Portfolio Coupling Oracle

Coupling Oracle

Cluster: BIOTECH_ILLIQUID

UNACCEPTABLE

Concentration 13.0% / 30.0% — COMPLIANT

Shared volume 2.0% NAV/day (stress)

Combined exit 13.0% / 2.0% = 6.5 days

Limit 5 days

Failure mode SHARED_EXIT_CAPACITY_BREACH

Concentration

0.43

Within limit — not the problem

Shared exit days

6.5 / 5.0

Exceeds limit — UNACCEPTABLE

Check	Result	Detail
Position A (Alpha)	COMPLIANT	0.70 util
Position B (Momentum)	COMPLIANT	0.60 util
Liquidity A (stress)	COMPLIANT	3.5d / 5d
Liquidity B (stress)	COMPLIANT	3.0d / 5d
Concentration	COMPLIANT	13% / 30%
Shared Exit Capacity	UNACCEPTABLE	6.5d / 5d

✗ EXECUTION BLOCKED

Concentration is compliant. Individual liquidity is compliant. But combined positions share one exit — market depth.

Each can exit alone under stress. Together, 13% / 2% = 6.5 days exceeds the 5-day policy limit.

This is the finance equivalent of thermal derating. The constraint is not the position. It is the shared resource.

↓ Evidence pack (JSON)

Architecture Transfer

Same pattern. Different physics.

Structural + Thermal

Structure SURVIVE + Thermal SURVIVE

→ Coupling EXTINCT

Heat derates aluminum yield strength

Finance: Concentration

Position COMPLIANT + Liquidity COMPLIANT

→ Coupling UNACCEPTABLE

Correlated cluster exceeds limit

Finance: Shared Exit

Position COMPLIANT + Liquidity COMPLIANT

→ Coupling UNACCEPTABLE

Shared market depth exhausted

The Pattern

Individual compliance does not guarantee compositional safety. The enforcement layer is domain-agnostic. The physics is not.

FCAL · AI proposes. Physics decides.

Fixed inputs for reproducibility. Three deterministic evidence packs (two failure modes + one control). Same architecture as structural domain. No probabilistic estimation.