Automated analysis of the paper
Staged validation history alone cannot push a research equilibrium into criticality. Yanç proves the sharp claim: purely causal validation cannot produce a same-cycle unit mode because the strictly causal block of the equilibrium response has spectral radius zero. The danger enters through the same-cycle settlement share λ. I would hesitate to trade on the fresh-versus-inherited comparison advertised in the abstract, a limitation the paper acknowledges.
Our own construction appears at the end. Three adaptations came first. We replaced the single abstract risky asset with liquid US ETFs. The research-validation-release feedback gain became a calibrated parameter rather than anything recovered from data, and we discretized the continuous-time Gaussian filter on daily bars.
Which set did you redraw?
One risky asset, one unknown constant drift, Gaussian prior, CARA investor. The investor places an ambiguity set around the posterior mean, then plays maxmin against it in the Gilboa-Schmeidler sense using a rectangular within-vintage distortion class.
The wrinkle lies in how that set ages. Update every member of the date-s set through the same realized likelihood, and prior-by-prior Bayesian transport preserves each alternative's displacement: V_t^{-1}(m_t^q − m_t) = q, the same q you started with. At t, the inherited set is m_t + V_t Q_s, which carries the original body forward. Fresh reconstruction instead builds a fixed-level credible region from today's posterior. In one dimension, its natural-coordinate half-width is z_alpha times the square root of posterior precision. Transport keeps the earlier budget k_s, while the fresh natural budget moves with precision at a fixed confidence level. Once learning becomes nontrivial, the two disagree.
Yanç assigns a price to that gap. In support-function space, the loss incurred when evaluator h uses the action selected by evaluator H is a Bregman divergence. It is nonnegative and vanishes exactly when H's action is also optimal under h, or, under strong concavity, when the two actions coincide. On a fixed common portfolio face in the Gaussian-CARA model, the expression is transparent: the ambiguity-budget gap enters quadratically, r_u(k, k̄) = α²v²(k̄ − k)²/(2γσ²). Integrating this source produces the recalibration tax. The closed form requires the common positive face to remain unbroken before T. If the process exits, a boundary continuation term remains, and the source is integrated only to a random exit time.
One consequence is easy to miss. Recalibration leaves the current interior risky position unchanged on a common face: π^SPE = π^PC = α(m − κ(v)v)/(γσ²). Protocol replacement therefore costs continuation welfare while preserving that interior position. A sufficiently large budget change can still move the position by crossing the switching surfaces at m = ±κ(v)v. Within the face, a snapshot audit of the book reveals nothing.
Settlement within the cycle
Suppliers allocate research effort across J stages. Validated capital from earlier stages reaches later ones through a strictly lower-triangular Q, preventing current-stage effort from returning through the validation channel during the same cycle. Current model value enters through λ, the settlement share.
Yanç separates the derivative of the equilibrium response into a strictly causal component V and a noncausal component N. The causal factor satisfies the fractional Volterra bound ||V^n|| ≤ [CΓ(η)S^η]^n/Γ(nη + 1), η = 1 − α.
Its spectral radius is zero.
The bound makes I − V invertible, allowing D_E F to factor as (I − V)(I − K), where K = (I − V)^{-1}N. Causal history can amplify through the resolvent. Singularity still requires N.
The full capacity-constrained equilibrium corresponds one-to-one with roots of a single scalar equation on [0, ȳ]. On a fixed strictly complementary cell with free coordinates I, the residual derivative is 1 − λβ_{R,I}W_R''(y). Here β_{R,I} denotes resolved validation exposure, and W_R'' is the curvature of optimized portfolio value in released precision. Uniqueness follows whenever λβ_R times the supremum of positive curvature stays below one. At a nonsingular equilibrium on a strictly complementary quadratic-cost cell, any differentiable release-dependent outcome satisfies dO/dz = ω_{O,h,I}/(1 − λ_ρ(Δ)χ_I). Holding costs and loadings fixed, an entrywise increase in the nonnegative Q cannot lower the full-interior β_R. A denser validation graph therefore approaches the pole without creating it.
The fresh-inherited crossover
With β_R = 32, the fresh Gaussian benchmark reaches unit gain at λ about 0.0480026. The inherited comparator reaches it at about 0.0122230. Measured by settlement share, inheritance arrives at reduced unit gain roughly 3.93 times earlier. This is the reduced unit-gain boundary; a feasible fold still depends on the lift conditions.
The abstract suggests the reverse ordering because inherited transport preserves natural displacement while fresh reconstruction can replace it. Yanç limits the claim carefully. He writes that "the comparison is benchmark-specific: across Gaussian primitives the fresh-minus-inherited curvature can have either sign," then states the crossover exactly. Set x = mq/z and x_c(q) = (8q² + 4)/(5q² + 1). Fresh curvature exceeds inherited curvature if and only if x > x_c(q). Both orderings can occur on the common positive face.
The mathematics works better than the headline. Provenance changes criticality through optimized curvature, yet the paper assigns no general stability sign. It leaves the decision to rebuild confidence sets unresolved. Its usable output is a decomposition of four quantities to measure: protocol-indexed curvature, resolved validation exposure, same-cycle timing, and the financial loading.
The paper locally identifies β_{R,I} through an exogenous perturbation of the predetermined shadow value at λ = 0. Combined with local curvature and release-time records, that perturbation supplies the operational input for the timing diagnostic. I have never seen a research group vary the shadow value deliberately. Validation time D is treated as exogenous, independent of current research and released precision, and outside suppliers' choice. In any real model-risk pipeline, I would attack that assumption first. Yanç sketches endogenous latency and observes that the scalar reduction breaks when latency depends on the full effort vector.
The timing result is worth keeping. Given completion probability p_Δ and decay ρ, the value weight λ_ρ(Δ) lies sharply within [e^{-ρΔ}p_Δ, p_Δ]. If the upper endpoint clears one, the result supplies a subcriticality certificate for every compatible timing law. If the interval straddles 1/χ_I(y) and the loading is nonzero, financial sensitivity has no finite uniform bound. Under the exponential clock, with r = ρ/ν = 0.5 and β_R = 32, the critical normalized deadlines are 0.0498 fresh and 0.0123 inherited. A finite deadline exists only when the curvature index exceeds 1 + r.
Yanç also keeps the meaning of criticality narrow. The paper gives a chain of non-implications: a statewise unit mode need not imply branch criticality, branch criticality need not imply a feasible fold, and a feasible fold need not be financially material. A counterexample breaks the second link. Take J = 2, C the identity, Q zero, loadings (1,1), cost shifter (100,1) and λ = 1. The scalar equation turns at y_f = 0.050727793389 with cost 0.000905444913, yet reconstruction gives the effort vector (−0.01945563, 0.07018342). Its first component is negative. Higher capacity limits cannot fix that failure.
Our implementation
The paper reports no empirical result: no dataset, no sample period, no universe, no Sharpe. The 32 and the 0.5 are chosen normalizations, not estimates.
Our build covers only the inherited branch. For each ETF, it runs a sequential Gaussian drift posterior. Variance is estimated once from a 252-session initialization window. Monthly evaluator resets snapshot the natural budget k_s = z_alpha√p and hold it fixed as posterior learning continues. The posterior mean receives a soft threshold: the position stays flat while the mean remains inside the budget, then takes scaled exposure after clearing it. We impose a Ten per cent cap per name and 4x gross. Rebalancing occurs daily at the close using a one-session signal lag. Every fill incurs commissions of four tenths of a cent a share, subject to a one dollar minimum.
From 2020-01-01 to 2024-07-01, the strategy returned 10.77% in total, with Sharpe 0.60, Sortino 0.83, Calmar 0.30, max drawdown −7.63%, and annualized volatility 3.97%. We implemented the paper's sizing rule; we did not test the paper. Volatility of 3.97% beside a 7.63% worst drawdown shows how rarely the book carried exposure during the window. The window opens on March 2020, and the deepest loss was 7.63%, so the rule avoided most of the stress it was designed to avoid. Freezing variance at the initialization window was our decision and the most consequential choice in 2020. We used One parameter setting (gamma 1.0, z_alpha 1.0) rather than a sweep, and every figure above is ours.
Because our build implements inheritance only, it cannot address the fresh-versus-inherited curvature comparison in Corollary 5.2's benchmark. The equity curve has no bearing on the 0.0480 versus 0.0122 ordering. Our construction exercises the maxmin sizing rule and nothing else from the paper. It leaves the recalibration tax, the Bregman representation, the causal-noncausal separator and the criticality gain untouched.
One result could change my mind
A credible mapping from observable release records, deployment dates and validation lags to β_{R,I} and W_R'' would change the verdict if it avoided perturbing a shadow price. The multiplier 1/(1 − λβW'') could then explain why a fast-deployment research process amplifies the same underlying shock.
The transport identity survives either way. Rebuilding a credible interval at a fixed confidence level as the posterior sharpens changes preferences, carries a price, and remains invisible in the current position on a common face. Firms that reset uncertainty budgets by calendar instead of transport make that change without booking it.
Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.