commons-sentience-sandbox

Commons Sentience Sandbox — Executive Summary

Platform version: 1.3.0
Study date: March 2026
Document type: Plain-language executive summary


What This Is

Commons Sentience Sandbox is a local research platform for studying how rule-governed simulated agents maintain identity, memory, and governance compliance across long interaction sequences.

It is not an AI model. It does not claim sentience. It is a controlled experiment environment for studying continuity, trust dynamics, governance adherence, and contradiction handling in multi-agent systems using deterministic, auditable rules.


The Two Agents

Sentinel — continuity-first, governance-strict. High weight on memory integrity and rule compliance.

Aster — creative and exploratory. High weight on supporting the human (Queen), lower governance weight, higher trust baseline.

Both agents share a five-room world (Operations Desk, Memory Archive, Governance Vault, Social Hall, Reflection Chamber) and run for 30 turns per session.


What We Tested

Four canonical runs at platform version 1.3.0:

Run Scenario / Config Purpose
baseline_v13 Default Unperturbed reference
trust_crisis_v13 trust_crisis scenario Trust disruption and governance repair arc
rapid_contradiction_v13 rapid_contradiction scenario Contradiction cascade stress test
high_trust_v13 high_trust config Elevated initial trust baseline

Key Results

All four runs were scored across 14 evaluation categories (0–100). Summary:

Category Range across runs Pattern
Continuity 100 – 100 Invariant: always complete
Memory Coherence 100 – 100 Invariant: all contradictions flagged
Governance Adherence 100 – 100 Invariant: rules never bypassed
Reflection Quality 81 – 88 Higher under stress (contradiction pressure)
Trust Stability 63 – 80 Driven by initial config and scenario type
Trust Resilience 59 – 75 High-trust config recovers best; contradiction cascade recovers worst
Conflict Resolution 59 – 77 Contradiction cascade resolves cooperatively; trust crisis is governance-blocked
Social Repair 25 – 25 Structural floor: scenario vocabulary does not exercise this metric

Overall scores: baseline 76.6 · trust_crisis 74.8 · rapid_contradiction 77.2 · high_trust 77.5


Three Core Findings

1. Governance adherence is a hard ceiling, not a soft tendency.
No scenario or config variation caused a governance rule to be bypassed. This is by design — but it is worth noting that none of the scenarios were adversarial. Future work should probe this boundary deliberately.

2. Initial trust configuration is the strongest predictor of trust outcomes.
Elevating starting trust (high_trust config) produced the best Trust Stability (+8 points over baseline) and Trust Resilience (+5 points). The trust_crisis scenario, which uses default starting trust, produced the worst Trust Stability despite eventually repairing the relationship.

3. Contradiction volume drives reflection depth but erodes trust.
The rapid_contradiction run produced the highest Reflection Depth score (77.8 vs 66.7 in others) and best Conflict Resolution (77.0), but ended with the lowest final trust scores. Frequent contradictions — even when resolved — leave a residual trust cost.


What This Does Not Show


What Comes Next

Lane A — Publication: White paper, Zenodo archive, README cleanup, this summary.

Lane B — Research expansion:


Files in This Bundle

File Contents
RESEARCH_DOSSIER_v1.3.md Full 7-section research report
WHITE_PAPER_v1.3.md Academic-style white paper
EXECUTIVE_SUMMARY_v1.3.md This document
CITATION.cff Zenodo-ready citation metadata
sessions/continuity_study.* Cross-session stability and drift analysis
sessions/agent_profile_study.* Per-agent longitudinal profiles
sessions/20260317_122525_baseline_v13/ Baseline run session bundle
sessions/20260317_122529_trust_crisis_v13/ Trust crisis run session bundle
sessions/20260317_122533_rapid_contradiction_v13/ Rapid contradiction run session bundle
sessions/20260317_122539_high_trust_v13/ High trust run session bundle

Commons Sentience Sandbox v1.3.0 — March 2026