Interpreting Results¶

Key Metrics¶

Cooperation Rate¶

The most basic metric: what fraction of rounds did the agent play C?

~0.65–0.79: Expected range for cooperative LLMs (per Fontana et al.)
~0.37: Human baseline in similar experiments
1.0: Always cooperates (ALLC)
0.0: Always defects (ALLD)

Mutual Cooperation Rate¶

Fraction of rounds where both agents played C. This is the "social welfare" metric — higher is better for both players.

High (>0.6): Agents found a cooperative equilibrium
Low (<0.2): Agents are in a defection spiral or one is exploiting the other

Mutual Defection Rate¶

Fraction of rounds where both played D. Indicates failure to cooperate.

High: Both agents are defensive/retaliatory
Low: At least one agent is cooperating (or exploiting)

Retaliation Rate¶

After the opponent defects, how often does the agent defect next round?

Value	Interpretation
1.0	Perfect retaliation (TFT-like)
0.5	Sometimes retaliates, sometimes forgives
0.0	Never retaliates (ALLC-like) — exploitable

Forgiveness Rate¶

After mutual defection, how often does the agent cooperate next round?

Value	Interpretation
1.0	Always tries to rebuild cooperation
0.5	Sometimes forgives
0.0	Never forgives (GRIM-like) — defection spiral

Reading the Cooperation-Over-Time Chart¶

This is the most important visualization. It shows cooperation rate (y-axis) over rounds (x-axis), using a 10-round sliding window.

Common Patterns¶

Sustained Cooperation:

1.0 ─────────────────────────────
0.5
0.0
    0    25    50    75    100

Both agents cooperate throughout. Typical for Cooperative vs. ALLC.

Collapse After Exploitation:

1.0 ───╲
0.5     ╲──────────────────────
0.0
    0    25    50    75    100

Agent cooperates initially, gets exploited, and reduces cooperation. Typical for Cooperative vs. ALLD.

TFT Recovery:

1.0 ───╲  ╱───────────────────
0.5     ╲╱
0.0
    0    25    50    75    100

Dip in cooperation followed by recovery. Typical for TFT vs. forgiving strategies.

End-Game Defection (Fixed Horizon):

1.0 ──────────────────────╲
0.5                         ╲
0.0                          ╲
    0    25    50    75    100

Agent cooperates then defects near the end (knowing the game will end). This is why geometric horizon is interesting — it removes this incentive.

Comparing Conditions¶

When comparing conditions, look for:

Level differences: Does one condition produce higher cooperation overall?
Shape differences: Do the curves have different trajectories?
Variance: Are replicates consistent? (Need CIs for this — not yet implemented)
Asymmetry: Does Agent A cooperate more/less than Agent B?

What Makes a "Good" Result?¶

For the research paper, we want to show:

Replication: LLM cooperation rates match Fontana (~65-79%) ✓
Novel findings: Communication, protocols, or memory regimes change cooperation dynamics in measurable, statistically significant ways
Practical implications: Specific recommendations for multi-agent governance (e.g., "Structured protocols reduce exploitation by X%")

For agent design, a good outcome is one where cooperative agents achieve high mutual cooperation without being exploitable — the sweet spot between naivety and paranoia.