The premium is the last thing you should compute
There is a habit in derivative work of treating the price as the object of interest. Somebody asks what a structure is worth, a model returns a number, and the number goes into a report. The hedge, if discussed at all, arrives afterward as an operational detail. That ordering is backwards. In an arbitrage-free market the price of a claim is not an independent quantity; it is the cost of the portfolio that reproduces the claim's payoff, and the portfolio is the primitive. If you cannot name the portfolio, you have not priced anything, you have expressed a view and formatted it to two decimals.
Beta owes you a named market
CAPM is written so often it has stopped being read: βi = ρimσi/σm. That second form shows what has to be true. Beta is not a property of the asset. It is a property of a pair — the asset and a specific market proxy, with a specific correlation and a specific volatility. Change the market and beta changes, not as a rounding artifact but as a matter of definition, and everybody uses a proxy since nobody actually holds the theoretical value-weighted everything-market.
Work the same equity leg, σi = 26%, market risk premium 5.5%, against two proxies a desk might reasonably pick. Proxy A, a broad large-cap index, σm = 16%, correlation 0.38: βA = 0.38×(0.26/0.16) = 0.6175, required excess 3.40%. Proxy B, a sector-heavy index closer to the leg's actual exposure, σm = 14%, correlation 0.50: βB = 0.50×(0.26/0.14) = 0.9286, required excess 5.11%. Same leg, same day, a spread of 1.71 percentage points of required return produced entirely by which index name went into a config file.
A research process rarely tries one proxy. It tries nine — large-cap, sector, equal-weight, a currency-hedged variant, two vendor composites, three windows — and reports the one whose beta produces the most defensible required return. If each candidate had a 5% chance of clearing a threshold by noise alone, the chance that some candidate clears is 1 - 0.959 ≈ 37%. A 5% test has become a 37% test, invisibly, because only the winner gets described in the writeup. In more than one audit, CEH-001's own denied 1.80 midpoint has been reverse-justified exactly this way: pick a beta somewhere between two proxies, discount the projected cashflows at the implied rate, and land near 1.80. That is not a derivation. It is a beta chosen to hit a target, and the giveaway is that it gets quoted without the proxy attached.
Hedge first, premium second, worked to four decimals
Leave the required-return world entirely and go to the actual derivative. Take the one-period binomial toy: a call struck at 100, stock at 100, up to 110, down to 90, riskless gross return 1.05. Before anything else, check the arbitrage sandwich, one line of arithmetic that costs nothing: 0.90 < 1.05 < 1.10. Holds. Now solve for the replicating portfolio directly — hold Δ shares, borrow the rest — and Δ = (10-0)/(110-90) = 0.50, B = -42.8571, so V0 = 50 - 42.8571 = 7.1429. Solve it the other way, as a discounted risk-neutral expectation with q = (1.05-0.90)/(1.10-0.90) = 0.75, and you land on the identical 7.1429, reached by rearrangement, not by new information. There is no probability of an up-move anywhere in the replication derivation. None was needed. The measure is a device that makes the algebra look like an expectation. It is not a belief.
Now break the sandwich deliberately, because this is the check that costs nothing and saves a mark. Suppose a stale funding curve pushes the riskless return to 1.12, outside [0.90, 1.10]. The formula for q still evaluates: q = (1.12-0.90)/0.20 = 1.10, a probability above one. The pricer does not object. It returns (1.10×10)/1.12 = 9.8214, a plausible-looking number 2.68 points above the correct 7.14, and it is the price of a market containing a free lunch. Nothing downstream can detect this. Only the two-comparison sandwich check can, and only if it actually runs.
Wrong coin, wrong world
Two probability objects live in the same model and they are not interchangeable. The physical measure P describes what the world actually does. The pricing measure Q is the one under which the discounted stock is a martingale — a bookkeeping device, not a forecast. Price the same call under P with p = 0.80 instead of the replicating q = 0.75 and the answer is 7.6190 against the honest 7.1429: rich by 6.67%. Push p to 0.85 and it's rich by 13.33%; to 0.90, by 20.00%. The trap is the first row: at p = 0.75 the wrong method happens to give the right answer, because that is exactly where P and Q coincide. A team that validates its pricer on a calibration case where the estimated real-world probability sits near the risk-neutral one will ship it and be wrong by twenty percent the moment the estimated drift moves — and the error is proportional to the risk premium, so the wrong coin is most wrong exactly where the interesting books are.
The CEH-001 version is how 1.80 gets a respectable-looking derivation. Build a Monte Carlo of the structure's payoff under estimated physical dynamics — real drift, real vol, all individually defensible — discount the expected payoff at the riskless rate, and land above the 1.50 spread-frame mark. It looks like the spread frame was too conservative. It is not a truth. It is a P-expectation discounted at an r that only belongs with Q: a hybrid that corresponds to no market and no portfolio. The tell needs no judgment call: check whether the simulation's drift is still zero after discounting. It never is.
Hedging makes model error measurable
The market price of risk, θ = (α-r)/σ, is the single field that bridges the two measures — and it exposes just how invisible a wrong drift really is. With σ = 20% annual, the standard error of an estimated mean return over T years is σ/√T: 20% at one year, 6.32% at ten years, 2.50% at sixty-four years. Telling a five-percentage-point drift gap apart from noise needs T = (2σ/(α-r))² = 64 years. No production system, no backtest, no monitoring window has that. The drift is not statistically identifiable at any horizon a desk will ever operate over — which is exactly why it has to be a declared card field, not an estimated one.
Here is the payoff for hedging first. Run the same detection calculation on the replication residual instead of the raw return series. Unhedged daily volatility on this book is 1.2649%; a well-hedged residual might run 0.08% — a variance reduction of 250 times, a 15.81× drop in standard deviation. Resolving that same drift gap on the residual series needs just 64 days instead of 64 years. Hedging does not just reduce risk. It makes model error measurable on a timescale a desk can actually govern, which is the deepest reason the premium is the last thing you should compute.
Nine more objects, one refuse culture
CAPM named worlds, the CRR sandwich and multiperiod induction, discounted wealth marks, Radon-Nikodym receipts, the price of risk θ, Black-Scholes mesh honesty, and martingale representation certificates all carry the rule this chapter opens with: a claimed premium without a hedge schedule under the declared measure is incomplete, and the schedule must be predictable with respect to the contract's own filtration. Black-Scholes appears here strictly as a contrast object, never as a silent sized-rights passport. A replication certificate that cannot produce its Δ process gets a band, or it gets refused.
Book: Chapter 15 in Rulers Before Models.