chore(epistemic): pivot architecture to research agenda based on Round 3 frontier audit

eval_prompts/reviews/round3/feedback/functorial_geodesics-feedback/feedback-copilotthinkdeeper.md

@@ -0,0 +1,38 @@
+**Short answer:** *The manuscript is conceptually interesting but, at a technical level, is fatally under‑specified: the realization functor, manifold model, stochastic calculus on manifolds, and numerical/algorithmic claims must be rebuilt from first principles before any stability claim can stand.*  
+
+### Major mathematical failures (why the core claim is not credible)
+- **No constructive \(\mathcal{R}\).** You assert a functor \(\mathcal{R}:\mathbf{Set}^{\mathcal{C}^{op}}\to\mathbf{Hilb}\) with metric and differential structure, but give no construction, continuity, or universality property. **Without an explicit embedding (or a theorem invoking a standard realization) you cannot transfer differential or stochastic calculus to categorical objects.**  
+  
+
+
+
+- **Topology and measurability are missing.** Embedding discrete presheaves into a Hilbert space requires choices (basis, topology, sigma‑algebra). **Is \(\mathcal{R}\) linear, continuous, measurable, or isometric?** Each choice changes whether \(\exp\), \(\log\), and stochastic integrals are defined.
+
+- **Manifold model is ambiguous and likely false in practice.** You treat the latent as a finite‑dimensional Riemannian manifold without arguing for finite dimensionality, smooth atlas, or injectivity radius. **High‑dimensional learned latents are typically only approximately low‑dimensional and may lack a global smooth structure; cut loci and non‑unique geodesics break the geodesic‑error calculus.**  
+  
+
+
+
+- **SDE derivation is incorrect for manifolds.** Writing \(de_t = -\kappa e_t\,dt + \sigma e_t\,dW_t\) for geodesic distance ignores Stratonovich/Ito distinctions, curvature corrections, and the fact that distance is not a smooth function at the cut locus. **You must derive the SDE from a manifold SDE (in Stratonovich form), apply Itô’s formula on manifolds, and include curvature/Jacobi field terms.**
+
+- **Stability bound is unjustified.** The Euclidean OU bound \(\kappa>\sigma^2/2\) does not automatically transfer: **sectional curvature, multiplicative noise geometry, and nonlinearity of \(d_{\mathcal{M}}\) modify thresholds**; in negative curvature noise can amplify deviations, in positive curvature it can damp them — you need a rigorous Lyapunov/stochastic stability proof (e.g., Khasminskii‑style) in normal coordinates including curvature terms.
+
+### Concrete technical fixes required (actionable)
+1. **Construct \(\mathcal{R}\)** for a nontrivial class of categories (e.g., finite presheaves) — give an explicit embedding (e.g., representables → feature maps → \(L^2\) basis) and prove injectivity/continuity.  
+2. **Specify \(\mathcal{M}\)**: finite vs infinite dimensional, Riemannian metric, smoothness class, injectivity radius; prove existence of exponential/log maps on the domain of interest.  
+3. **Derive the manifold SDE**: start with Stratonovich SDE for latent state \(X_t\), compute \(d\,d_{\mathcal{M}}(X_t,Y_t)\) via stochastic development and Jacobi fields, show correction terms explicitly.  
+4. **Prove a stability theorem**: linearize in normal coordinates, include curvature bounds (sectional curvature \(K\)), and produce a precise inequality replacing \(\kappa>\sigma^2/2\). Use comparison theorems to bound deviations.  
+5. **Provide counterexamples**: construct simple presheaves and embeddings where the Euclidean bound fails (e.g., manifold with negative curvature or near cut locus).  
+6. **Numerical plan**: give algorithms for computing \(\exp/\log\) robustly in high dimension, quantify numerical error propagation, and show synthetic experiments validating the corrected theorem.
+
+### Experiments and diagnostics to include
+- Toy categorical examples with explicit \(\mathcal{R}\) and measured curvature.  
+- Simulated manifold SDEs (varying curvature) to show how thresholds shift.  
+- Ablations on embedding choices and numerical stability of \(\log\)/\(\exp\).  
+  
+
+
+
+---
+
+**Bottom line:** *Tear it down and rebuild it.* Replace hand‑waving with a concrete \(\mathcal{R}\), a precise manifold model, a Stratonovich‑based SDE derivation with curvature terms, a provable stability theorem, and synthetic experiments that demonstrate the corrected bounds.