πŸ”„

World Replay System

Alice records every multi-agent simulation in full β€” each character's position, action, dialogue, emotion, and goal at every step β€” so you can replay it step by step in your browser. The current world, Void Realm, is a controlled single-cognition isolation testbed: the observed cognitive agent is Lina, with 4 scripted-probe characters, so believable behavior and emergence can be studied in a clean setting.

Enter the Replay System β†’

Below are representative formal experiments from the three most recent releases (one per release, each minted an Exp.N with category = experiment); as new releases ship, the oldest is replaced by the newest. The 'Observations' on each card are honest notes from that run, not paper-grade conclusions; see the devlog for fuller analysis.

experiment

Void Realm v1.0.5 Β· Exp.2 Fixes Validated (sleep / text / valence)

🌍 Void Realm v1.0.5 · 144 steps · 5 personas

A full-day validation after fixing the three issues v1.0.4 exposed: sleep-on-bed, no internal text-prefix leak, and a genuinely live emotion valence. Confirms all three fixes hold end to end.

Observations
  • All 37 sleep steps stay on the bed (0 drift); all 720 action descriptions are clean with 0 prefix leaks.
  • Valence comes alive: 140/144 steps non-zero (97%). Yet in the isolated scene Lina trends negative β€” the problem v1.0.6 set out to solve.
β–Ά Play this replay
experiment

Void Realm v1.0.6 Β· Exp.3.1 Emotion Ablation Baseline (C0 Full)

🌍 Void Realm v1.0.6 · 144 steps · 5 personas

The baseline condition C0 of an emotion-system ablation campaign (Exp.3.1–3.5, five conditions): emotion fully on, plus two positive sources (action completion, reflection insight). Tests whether 'positive sources' can flip the negative-leaning mood seen in v1.0.5.

Observations
  • The two positive sources fully flip the negativity: Lina's valence is 76% positive, mean +0.45 (the 'no positive source' ablation condition was 97% negative, 0% positive).
  • Regression checks all green: 40 sleep steps with 0 drift, 0 text leaks, valence non-zero 97% of the time.
β–Ά Play this replay
experiment

Void Realm v1.0.7 Β· Exp.5.2 Memory Relation Graph + Dialogue v2

🌍 Void Realm v1.0.7 · 144 steps · 5 personas

The seal-grade formal experiment for v1.0.7: a graph-structured memory (LLM-extracted relational facts) plus dialogue v2 (cross-conversation summaries). The headline result is that 'goldfish memory' is cured β€” on Lina's second encounter with Gus she remembers the prior visit instead of being shocked anew each time.

Observations
  • The memory graph works: Lina accumulates 126 structured facts about Gus, with bi-temporal invalidation (contradicted facts are retired, not deleted).
  • Goldfish memory is gone: on the second visit the prior cross-conversation summary is injected into her prompt β€” she replies 'why bring this up all of a sudden?' rather than being freshly shocked (versus v1.0.6, where the same scene re-shocked her every time).
β–Ά Play this replay
About