A provenance-first semantic coordinate, query, lineage, and replay substrate.

October 11, 2026 · Shared references, separate identities

A bounded local prototype let two simulated clients retrieve identical references from the same pinned synthetic record set while retaining distinct client identities and observations. A fresh Linux reconstruction passed 63 tests: 35 inherited DLI tests and 28 pilot tests. The recorded recursive path, type, mode, size, and content-hash inventory remained unchanged across reads.

Focused negative tests covered expired or revoked access, incorrect identity and project bindings, incorrect snapshot pins, corrupted state, and excluded-data leakage. Fresh fixture setup wrote only disposable synthetic data; the read phase did not write DLI state.

This is reference-only engineering evidence. Recall uses deterministic token matching, not complete semantic memory. No actual ChatGPT/Nova, Codex, or other provider was connected; no live authentication or MCP service, real-history access, production readiness, or formal Runtime acceptance was established. The network guard was at the Python layer, not OS-level network isolation. Retrieval grants no truth, promotion, or action authority.

The original event remains

In the research model, nothing arrives alone: a record can retain time, source, participating identity, relationships, context, and uncertainty without being dissolved into a consolidated belief. Supported recurrence may suggest a candidate route; a consequential exception can remain as scar evidence—a reason to inspect, not a judgment.

The original encounter, its attribution, and any contradiction remain available. What may become shorter is the path through the evidence, and any durable default still requires a separate governed transition.