The Living World Model: A Bitemporal Evidence Substrate
A bitemporal evidence substrate makes a world model 'living'. Each fact carries valid time (when it was true) and transaction time (when we believed it). An append-only ledger separates source of truth from a derived memory index. A continuous-learning loop updates calibration via proper scoring rules.
Figure 1. The evidence ledger as the append-only source of truth, and the memory index as a mutable projection optimised for working recall. As-known-at queries probe the ledger directly, decoupling audit from cache.
Memory without amnesia
A world model that cannot remember cannot improve. To make a world model 'living' we propose a bitemporal evidence substrate with two distinguished time axes and a continuous learning loop that updates calibration and mechanism from measured outcomes.
Each fact f carries a valid time interval τ_v (when the fact was true in the world) and a transaction time interval τ_x (when the system believed it). The knowledge state at any past instant can be reconstructed exactly.
Ledger versus memory
The substrate separates two artifacts. The evidence ledger is append-only, hash-linked, and the authoritative source of truth. The memory index is a derived, mutable projection of the ledger, optimised for retrieval latency.
The key audit property is that as-known-at reconstruction queries the ledger directly, not the cache. The cache may be wrong, the ledger cannot be silently rewritten.
Typed experience
Observations, actions, outcomes, refusals, calibration probes, and human reviews are all typed as immutable ledger entries. Each carries its provenance, its ontology version, and its bitemporal coordinates. The substrate composes these typed events into derived passports that downstream layers consume.
The continuous-learning loop
Predictions are paired with the outcomes they shaped. Outcomes are scored using proper scoring rules. The resulting learning event updates calibration and mechanism, both as new ledger entries. The model becomes sharper from measured experience, not from anecdote.
A daily 'Dream Check' consolidates the ledger into the memory index, surfaces drift, and emits a written reconciliation event.
Limitations and status
The bitemporal substrate is implemented and in use. The continuous-learning loop is a designed capability: outcome telemetry from production deployments is the missing piece to make the 'living' property an empirical one rather than an architectural one. We are running outcome instrumentation in three customer deployments.
Papadopoulos, L. (2026). The Living World Model: A Bitemporal Evidence Substrate. Helios Brain, HB-TR-2026-05.