A world has a history.
When a match result, an item trail or a patch effect is disputed, can your team reconstruct what the authoritative simulation did—and find the first point where two runs differ?
State is not a video frame.
The authoritative simulation advances the world. The client renders, animates and predicts locally. The proposed evidence layer records simulation history and coordinates replay work. Compact world-state changes are kept separately from client presentation, so an investigation can follow the state that mattered.
Reconstructing server state does not establish rendering performance, lower streaming cost or engine-wide compatibility. Those are different tests.
Find the first difference, not just the final symptom.
- NoticeFlag the outcome that needs explanation.
- IsolateSelect the session, shard and tick window.
- ReconstructStart at the nearest checkpoint and apply ordered events.
- CompareCheck state across pinned builds.
- AttributeTrace the first mismatch to its causal parents and the rule in force.
Explore the proposed workflow. Choose an event, then follow it from the authoritative world state to a compact change, the client's rendering and the history that holds the earlier value.
Keep the conditions for reconstruction.
The session record would retain build and configuration identifiers, seeds, initial state, ordered events, checkpoint hashes and the canonical serialization method. A checkpoint is useful only when its inputs and environment can be found again.
If persistent characters enter a later phase, model responses would be recorded as external events. Replaying a stored response is not the same as generating an identical response through fresh inference.
One arena. An explicit set of gates.
The proposed first experiment is a server-authoritative arena: 16 scripted players, one versioned ruleset, fixed 30 Hz ticks and 100 recorded ten-minute sessions. Normal play, collisions, scoring, reconnects and checkpoint recovery are covered with synthetic identifiers and scripted inputs. Integration points are confirmed in discovery; a standalone test is labelled harness-only.
Changed, missing, duplicated and reordered events—and mismatched builds or configurations—must produce an explicit mismatch or refusal. Recovery from a checkpoint is compared with an uninterrupted reference. Controls must pass.
Before building, agree checkpoint match criteria, earliest-fault detection, tick-time overhead and storage per player-minute against your baseline. A second engineer reproduces the report. The evidence package would include the environment manifest, input corpus, divergence report, raw performance data and integration-effort log.
State hashes locate differences. They do not establish input legitimacy, cheat prevention, fairness, production scale or cross-device determinism. Stronger integrity claims require a security design and threat model.
From the question to the test
Bring the session you need to explain.
Bring one bounded workload, your current workflow and the evidence you would need to make a decision.