A model was never the unit

Separate models are not independent risk. Independence is a joint claim. A folder of single-name greens never measures it. Shared features, shared calm years, and shared exams compress a book into one bet.

Most teams still gate models name by name. CEH-001 looks fine. Cousin A looks fine. Cousin B looks fine. The dashboard is a row of greens. Then one ugly week hits the whole structure family at once, and the diversified book behaves like one bet with many tickers. The quiet idea underneath that week is book coupling: models trained, scored, and promoted as separate objects can still fail together.

What “separate models” quietly pretends

In probability language, two random variables can each look well behaved on their own and still move as a bundle. Correlation, shared factors, common regimes, shared loss shapes: any of those can make the joint story worse than the marginal story. Finance has lived with that forever in portfolios (Markowitz concentration; common-factor risk). Financial AI often forgets it the moment models get their own repos and their own green lights.

Independently trained does not mean independently wrong. If every model in a family optimizes a similar objective, drinks from a similar feature kitchen, sees the same calm years in training, and faces the same single-name eval pack, they are not five opinions. They are five copies of one opinion with different random seeds and slightly different names. That is evaluative compression at book scale: the exam itself squeezes many candidates toward the same shortcuts and the same blind spots.

Teams hear “ensemble” and feel safer. An ensemble of coupled learners is not insurance. If the members share the fragile assumption, averaging them can sharpen the wrong answer. Diversity of filenames is not diversity of failure modes. Model inventory practice in classical MRM (SR 11-7) already treated the firm’s model set as a portfolio of model risk, not a row of unrelated certificates. AI model cards rediscover the same need when each name has an owner and joint failure has none.

Effective N: the arithmetic that flips a dashboard

For N names with average pairwise residual correlation \(\bar{\rho}\), a standard equicorrelation approximation for effective independent bets is:

\[ N_{\mathrm{eff}} = \frac{N}{1 + (N-1)\bar{\rho}} \]

Three names at average correlation 0.70 yield about 1.25 effective bets. Equal-weight portfolio volatility is then about fifty-five percent higher than an independence-sized mandate. Single-name greens hid a concentration breach.

How this shows up in production

1. Single-instrument courage. Each structure clears its own stress pack. Joint stress across the book never runs, or runs as a polite slide that nobody can fail a release on.
2. Shared-driver blindness. One liquidity cue, one funding proxy, one calm-regime quote intensity feature shows up across cousins. One regime flip moves all marks together.
3. Eval pack monoculture. Every cousin is scored on isomorphic calm weeks. The exam never jointly stresses co-breaks. Rankings rhyme. Failures rhyme later.
4. Size that assumes independence. Risk budgets add as if names were orthogonal. Coupling silently multiplies risk.
5. Local patches, no retirement. Each cousin gets a ticket. Nobody retires the coupled set. The book keeps three local fixes and one structural bet.

CEH-001 and cousins under joint stress

CEH-001 and two cousins each clear powered single-name eval. Residual correlations over the promotion window sit near 0.71, 0.66, and 0.74, all above a card limit of 0.35. Effective N is about 1.25. Joint stress week: dual-frame gaps reopen together; midpoint pressure appears as a family habit, not a single model quirk. Single-name dashboards stay mostly green until P&L is read at book level. The diversified story was a spreadsheet of marginal certificates.

Under a shared funding proxy, when the proxy flips, spread-frame marks gap together while curve-frame marks lag on missing prints. Dual-frame disagreement widens as a family. A single-name hard pack that never co-breaks with cousins will still pass. Joint stress is the only exam that asks the blotter’s question.

Money picture: inventory of 40,000 units makes a 0.60 dual-frame gap into real P&L distance. If three cousins invent correlated midpoint pressure together, the book does not have three independent 0.60 problems. It has one structural habit with three tickers. Single-name powered eval can be honest and still miss that.

Why this failure mode survives sophisticated orgs: single-name model risk committees scale organizationally. Each desk owns a name; each name has an owner; each owner brings a certificate. Joint failure has no natural owner. Until coupling is a card with a refuse edge, the org chart itself produces evaluative compression. Multiple testing across names without joint control is the book-scale cousin of best-of-N inflation (Chapter 8): you can correct each name and still overstate book safety if you never model dependence.

What breaks in the trenches

• Model trench: shared feature kitchen unhashed; residual correlations never computed into N_eff.
• Agent loop trench: per-name promote tools with no family ticket.
• Harness trench: joint stress is quarterly slide theater; cannot fail release.
• Pipeline trench: ensemble of cousins averages residual opinions toward 1.80 peace at book scale.

What to lock before automation

• Residual correlation cards with numeric limits (example limit 0.35).
• Joint stress packs with hashes that can fail release.
• Sizing formulas that use N_eff, not headcount.
• Ban on reading a row of single-name greens as diversification.
• Ensemble policies that require diversity of failure modes, not filenames.
• CEH-001 family: dual marks still dual; correlated midpoint pressure is a family refuse.

I do not trust a book of single-name certificates without a joint stress and a coupling map that can fail promote. The next post builds that book harness.

CEH-001 week walked through the stack

Sizing operators can recompute: N = 3, ρ-bar = 0.70, N_eff = 3 / 2.4 = 1.25. Equal-weight portfolio variance equals \(3 \cdot \frac{1}{9}\sigma^2 + 6 \cdot \frac{1}{9} \cdot 0.70\sigma^2 = 0.80\sigma^2\), so portfolio vol is about 0.894σ. Independence fantasy was 0.577σ. Risk multiple 1.55. If the mandate was written to independence, the harness refuses the size and rewrites the lease to the multiple, or refuses the promote until coupling falls.

Worked retirement path. Inventory shows three consecutive hard weeks where shadow dual-frame deltas for CEH-001 and cousins exceed policy and rhyme in sign. Canary hard mix fails on the same shared liquidity cue. Retirement hook fires: family auto-hedge revoked, size to escalate-only, shared feature kitchen scheduled for unbind, new joint exam hash required before any cousin re-promotes. No hero merges three local “watching it” tickets into a shrug.

Promotion ticket fields the book plane adds: coupling_map_hash, joint_exam_hash, N_eff, risk_multiple, retirement_hook_ids. Missing any field is hard fail for multi-name autonomy. Single-name research can still run; single-name spend at book-material size cannot.

Neighbor contract. Eval harness (Chapter 8) certifies single-name statistical claims. Pipeline harness (Chapter 18) binds stages. Scheme harness (Chapter 5) pins discrete rulers. Ship harness (Chapter 26) binds live observation. Book harness is the only one allowed to say the greens do not sum. Promote preconditions list all five families of doors.

Book: Chapter 27 in Rulers Before Models.