Does the caller-neutral Harness boundary remain valuable across different workloads and execution strategies?
Test whether one bounded Agent-Run authority can preserve cognition, Provider/Tool continuity, recovery, action admission, and completion proposals across callers without becoming a second Task/domain store or universal planner.
01 / Current position
A hypothesis is not the current judgment.
The dossier preserves the live research position. Dated articles preserve the complete evidence and argument that changed it.
Immutable Run contracts, exact turn capability authority, durable Provider/Tool effect recovery, current cognition projection, and caller-owned semantic completion form a reusable execution boundary while strategy, evidence relevance, domain truth, and final acceptance remain external.
Answered for the current pre-1.0 product boundary. Harness is now an independent, Host-free, caller-neutral durable cognitive execution substrate with its own Journal/CAS and supported HarnessAgentRun path. Real consumer pressure retained exact Run/turn capability separation, WorkingSet/WorkingView versus canonical history, Provider/Tool UNKNOWN recovery, structured completion as proposal rather than domain truth, and bounded capability composition; other candidate machinery such as a universal planner, generic Memory/RAG, Provider-continuation core, capability registry, and automatic knowledge/promotion services was deleted or kept local. The latest research tournament reports no next branch selected: remaining semantic frontiers are blocked on new consumers, owner transitions, or implementation diversity rather than an unresolved proof that Harness has any reusable boundary at all.
02 / Decision boundary
What keeps this Question alive?
After the Host-backed product line was deleted, Harness earns retention only if caller-neutral Run structure and recovery recur independently of one upstream owner and remain cheaper than duplicated Provider/Runtime integration.
Operate and pressure the current boundary through ordinary materially different consumers. Reopen architecture only when a real workload exposes a repeated Run-owned fact the current contract cannot preserve, or when direct Provider + Runtime integration demonstrates the same recovery, authority, cognition, and completion handoff with materially less permanent Harness machinery.
Direct Provider and Runtime composition repeatedly matches the current interruption recovery, bounded action authority, cognition/current-view separation, evidence, and completion handoff across materially different callers at lower permanent cost—or a new workload requires caller/domain semantics to enter Harness for correctness.
03 / Supporting publications
Complete arguments connected to this Question.
4 dated publications currently document this research line.
The Model Finished. The Protocol Wasn't Done.
How a second native Provider forced Ordivon Harness to separate model-visible semantic history from Provider-required opaque continuation state, and why only one new primitive survived a five-frontier productization test.
12 August 2026 ↗Research reportWe Cut 203 Observations to 8. The Agent Still Refused to Edit.
How Harness P5 moved from exhaustive repository traversal toward Agent-owned high-information discovery, why minimizing observation count alone failed, and why the successful result was an evidence-backed refusal to mutate source.
12 August 2026 ↗Research reportWhat Survived When Codex and Hermes Replaced Each Other Mid-Task
How one task remained coherent through Codex→Hermes and Hermes→Codex replacement, three injected faults, and materially different provider lifecycles.
4 August 2026 ↗Architecture decisionWhy Ordivon Needs a Harness—but Not a Universal Harness
Why H5 correctly rejected a common Codex/Hermes lifecycle while a separate Ordivon Harness remains strategically necessary for bare model APIs, local inference, and controlled agent-loop research.
4 August 2026 ↗04 / Source discipline
The dossier is an index, not the evidence authority.
Owns the current judgment, next test, and deletion condition.
Own complete dated arguments, limitations, comparisons, and source links.
Own exact code, tests, releases, receipts, and machine evidence.