VennLimitations#
Venn is not universally calibrated, a privacy boundary, an authorization system or a validated guardrail. Embeddings can be inverted. Negation, numbers, dates, identity and adversarial paraphrases can defeat a frozen embedding even when nearby states look geometrically similar.
Images/audio are experimental and do not have established parity with text. Media training/use terms need separate review.
Compatibility tests establish wire format, not semantic correctness. Conditional option probabilities can be overconfident, and standard SDKs may drop Venn-specific abstention metadata. A permutation p-value is not a posterior. Population-shift sensitivity depends on the sample size and shift.
The fixed semantic boundary search exposes paraphrase/format flips and especially poor numeric robustness. The space benchmark reports only its headless frozen geometric tasks, with conditional finite-corpus intervals. This repository contains no model weights, training runs or raw evaluation archives. Download the published head for trained-head inference.
Outcome utilities depend on the app's reward definition, observed feedback, exploration and store freshness. They are not calibrated probabilities. Approximate search, duplicate-state merging and exact-key overrides can change the policy; equal vector dimensions do not establish compatible representations. The experiments behind the outcome API used one seed each and do not establish generalization to other applications. HTTP outcome-only serving still requires the head; only the in-process path is headless.