Correlation Framework (Category Caps + Capacity Rights)

Last Updated: 2026-01-27 Status: Draft

This framework has two layers:

Enforcement (simple): governance defines arbitrary categories with hard caps; any exposure above a cap is forced to be fully capitalized (100% CRR). “Capacity rights” prevent new deployers from instantly displacing incumbents.
Calibration (model-driven): governance (or governance-controlled automation) uses regular scenario stress calculations to set and update those caps over time.

Core Idea

Governance defines a set of Correlation Categories (examples: “CLOs”, “US-based assets”, “40–50 month duration”, “real estate”).
Each category has a cap as a percentage of the total portfolio (e.g., “CLOs ≤ 10% of USDS supply”).
Every asset is classified into one or more categories at onboarding.
If aggregate exposure exceeds a category cap, the excess portion is subject to 100% CRR (no leverage on the excess).
Each category also has non-transferable capacity rights (“grandfathered slices”) that evolve over time based on first-come-first-serve, and when cap is reached, through sustained payment of the 100% CRR penalty.

Why This Exists

Duration matching (ALDM with Duration Buckets + caps) prevents classic maturity mismatch, but it does not stop the system from concentrating into a single correlated risk type (e.g., "lots of CLO-like credit") by adding more risk capital. This framework makes "we simply don't want more of this risk" enforceable.

Coverage: Category caps are the mechanism for all concentration risk, including cross-chain bridge exposure. Bridge risk (protocol risk, liveness risk, censorship risk) is managed by defining bridge-related categories and setting appropriate caps — not through a separate risk module.

How Caps Are Set (Scenario Calibration)

The cap system is intentionally simple to enforce. The choice of cap values can be updated over time using a scenario stress engine.

Scenario Engine (Concept)

Define a scenario set S (e.g., credit crisis, crypto crash, stablecoin confidence shock). For each scenario s and category c, estimate the loss severity per unit exposure:

L(c, s) = stressed loss per $1 of category exposure under scenario s

Then choose caps such that portfolio losses remain within a governance-defined loss budget:

B(s) = maximum portfolio loss fraction tolerated under scenario s (or equivalent solvency constraint)

This turns “correlation” into a concrete calibration surface rather than an abstract covariance matrix.

Exposure Basis for Calibration

Use the same exposure basis as enforcement:

Matched exposure: notional / duration-value
Unmatched exposure: market value (MTM)

Scenario calibration may run in either of two modes:

Static matching mode: treat matched/unmatched splits as measured for the current week and calibrate caps using those exposures.
Scenario matching mode: allow scenarios (especially run/confidence scenarios) to reduce matching capacity, increasing unmatched exposure and therefore increasing L(c, s) for duration-sensitive categories.

Two Ways to Derive Caps

Method A (Simple, Independent Caps)

Set each category’s cap from a worst-case scenario bound:

cap_percent[c] = min_s ( B(s) / L(c, s) )

Interpretation: “even if the entire allowed slice of category c is held, the portfolio survives each scenario’s budget.”

This method is simple and robust but conservative because it ignores interactions among categories.

Method B (Joint Caps via Optimization)

Choose caps jointly to satisfy scenario constraints while allowing a preferred composition:

For each scenario s:
  Σ_c cap_percent[c] * L(c, s) ≤ B(s)

Subject to additional governance constraints:

floors/ceilings per category
maximum change per update (rate limit)
policy overrides (legal/compliance “never exceed” caps)

Governance Controls for Calibration

Governance should set:

Scenario set S and budgets B(s)
Cap update cadence (daily at settlement vs monthly)
Smoothing / rate limits on cap changes (e.g., ±X% per update)
Emergency freeze: ability to halt cap updates while keeping enforcement active

Definitions

Exposure Measure (Matched vs Unmatched)

Caps are measured on a unified “portfolio exposure” basis, but the measurement differs by treatment:

Matched exposure: measured using notional / duration-value (the "hold-to-maturity" relevant size).
Unmatched exposure: measured using market value (MTM).

Interpretation: if an exposure is treated as forced-sale (unmatched), it consumes category capacity by its MTM size; if it is treated as duration-matched, it consumes by notional.

Category Cap

For category c:

cap_amount[c] = cap_percent[c] × total_portfolio

Where total_portfolio is defined at the system level (e.g., “USDS supply”).

Over-Cap Portion and Penalty

For each category c, define:

excess[c] = max(0, exposure_total[c] - cap_amount[c])

Only the excess portion receives the 100% CRR penalty.

The 100% CRR penalty is applied once (not stacked) even if an exposure is tagged into multiple categories.

Capacity Rights (“Grandfathered Slices”)

To prevent new deployers from instantly pushing old holders out of a capped category, each category maintains a per-Prime non-transferable allocation:

alloc[p][c] = Prime p’s current “in-cap” share for category c
Σ_p alloc[p][c] = cap_amount[c]

This allocation changes gradually over time based on which Primes have been paying the over-cap penalty.

Normalization Period

For a given exposure, the “speed” at which over-cap penalty converts into allocation is governed by:

T = max(asset_SPTP, 3 months)

Intuition: longer-duration assets take longer to normalize; very short durations still have a minimum normalization of 3 months.

Reallocation Rule (Daily Settlement)

At each daily settlement:

Compute each Prime’s category exposure E[p][c] (using matched/notional vs unmatched/MTM).
Compute “penalized amount”:

P[p][c] = max(0, E[p][c] - alloc[p][c])

This is the portion that is currently over that Prime’s in-cap allocation for the category and therefore faces 100% CRR.

Update allocations by shifting capacity toward Primes that have been paying penalties, with rate:

α = (1 epoch) / T

Mechanically, the intent is:

If you are over-cap by P, then over time T you earn roughly P of in-cap capacity.
That earned capacity comes from the existing in-cap allocations, which pushes incumbents into partial penalty and triggers a “fight back” dynamic.

Implementation note: the exact update function can be implemented as an exponential moving average of penalized amounts or as a direct reallocation step. The key invariant is that long-run allocations reflect “who paid the most 100% CRR penalty for how long,” normalized by T.

Enforcement in the Capital Formula

The over-cap portion of an exposure is forced to 100% CRR:

CRR_effective = min(1.0, CRR_base + cap_penalty_addon)

Where cap_penalty_addon is sized such that the over-cap portion’s required capital equals its exposure amount.

Outputs (For Beacon Reporting)

Per category c:

cap_percent[c], cap_amount[c]
exposure_total[c]
utilization[c] = exposure_total[c] / cap_amount[c]

Per Prime p and category c:

alloc[p][c]
E[p][c]
P[p][c] (penalized / over-allocation amount)

Portfolio-level:

total over-cap exposure (non-stacked)
total 100%-CRR-required capital due to category caps

Open Questions

What is the exact "allocation update" function (EMA vs explicit per-epoch reallocation), and where does it live (beacon vs on-chain vs governance process)?
How to compute category exposure for multi-category assets without double counting while still enforcing caps (binding category vs portfolio-wide max penalty)?
Should caps be defined per-Prime, per-Prime-type, or only globally across all Primes?
How are categories versioned and migrated when governance changes category definitions?
Which cap-calibration method should be canonical (independent worst-case caps vs joint optimization)?
What is the minimal definition of L(c, s) required to safely add a new category?