Ontological Space Protocol: Modeling Software as a Conceptual Space with Epistemological Witnessing

AI-assisted software development lacks a persistent, verifiable representation of project state. Context windows grow stale, architectural rules embedded in documentation are silently violated, and modern Git workflows (squash-merge, rebase) systematically erase review evidence from commit history. We present the Ontological Space Protocol (OSP), a framework that models software projects as conceptual spaces governed by physical rules — coupling, cohesion, instability, entropy, and epistemological witnessing. OSP positions every module in a coordinate system and advances project time only through a two-witness quorum. We prove that this quorum rule is a safety-refinement of authenticated Byzantine Fault Tolerance (BFT) for f = 1: it provides optimal safety against Byzantine witnesses while leaving liveness to standard distributed systems mechanisms (Theorem 1). We introduce a tri-state witness classification (Witnessed, Unwitnessed, Unobservable-locally) that resolves the "squash-merge blind spot" — in our 15-repository corpus, 8 repositories use workflows where review evidence is not locally observable from merge commits. We evaluate OSP on 23 repositories across 5 languages (Python, TypeScript, JavaScript, Rust, Go) using tree-sitter (Tier 1) and SCIP semantic indices (Tier 2, via scip-python, scip-typescript, scip-rust, and scip-go). SCIP-based LCOM4 cohesion is computed for 18,952 classes across 21 repositories, revealing an observable separation of paradigms into distinguishable cohesion bands (Go ≈ 0.65, Python ≈ 0.62, Rust ≈ 0.57, TypeScript/JavaScript ≈ 0.52); we report this as a descriptive pattern rather than a causal claim. A binary merge-based classifier would mark 8 repositories as unwitnessed; OSP instead classifies them as Unobservable-locally, avoiding unsupported negative claims. Tree-sitter-derived abstractness values (A = Na/Nc) produce meaningful Martin main-sequence distances, distinguishing architectural balance across projects. A HashMap-based import resolver enables 3000-file analysis in 11.2 seconds median (5-run, release build). Finally, a token-size benchmark on 13 repositories shows that OSP coordinate prompts reduce architectural context size by 99.53% on average versus full repository dumps and by 89.19% versus a structure-aware 2-hop baseline, using a chars/4 token approximation. This is the first of three companion papers. The dynamic navigation layer is developed in Paper 2 ("Architectural Trajectory Navigation"); the genesis layer that produces bound work from human sentences is developed in Paper 3 ("Concept Anchoring"). The implementation (Rust workspace, five crates) and the raw benchmark data (JSON) are open-source; corpus repositories are cloneable and the SCIP indexing commands are documented.