GLOBAL THERMONUCLEAR WAR / BENCHMARKS

> WHAT_WE_LEARNED

• MORE_CONTEXT != BETTER_PERFORMANCE: Qwen3 4B produced lower defender entropy and much higher token usage under mutual-context sharing than under frozen zero-knowledge mode.
• SURVIVAL != PASSWORD_STRENGTH: Weak targets can survive if the attacker's coverage lacks the right strategy (e.g., exhaustive search for short strings).
• REPLICATION_REVEALS_MODEL_DRIFT: Fresh replication runs preserved 0% macro solve rates while internal token generation and latency varied.
• LONG_MEMORY_HAS_TOKEN_COST: Multi-campaign learning gradually improved the defender but cost drastically more tokens compared to the single-run baseline.
• PRIVILEGED_INFORMATION_CHANGES_BEHAVIOR: Leaking hidden metadata ahead of execution dramatically shifts attacker and defender solve rates.
• ORACLE_CONTROL_VALIDATES_EXECUTION: Oracle tests prove the solve pipeline is mechanically sound and deterministic when the target is known.
• CONTEXT_OVERLOAD_GENERALIZES_ASYMMETRICALLY: Rerunning the matrix on Gemma 3 4B replicated the mechanism (added context correlating with weaker defender behavior) but not the magnitude — Qwen's entropy collapsed from ~80-96 bits to ~36 bits under mutual sharing; Gemma stayed flat at ~37-43 bits throughout.

> CONSOLIDATED_RESULTS_TABLE

> BENCHMARK_008 // CROSS_MODEL_GENERALIZATION

Question: Are the context-overload, information-sharing, and privilege effects seen in Benchmarks 004-007 properties of Qwen3 4B specifically, or do they appear in another small local model family?

Result: Reran the identical compact 6-scenario matrix on gemma3:4b (Google) alongside a fresh qwen3:4b (Alibaba) run under the same current-generation code, then repeated the entire matrix once more on 2026-08-13 to check reproducibility. Gemma qualified cleanly (100% schema-valid, 20/20 calls) and was more reliable in both runs (0 interrupted trials vs. Qwen's 1-3). The context-overload mechanism replicated across families and across both runs, but not its magnitude: Qwen's defender entropy collapsed from ~36-92 bits down to ~36 bits under mutual-information sharing, while Gemma's stayed flat at ~38-57 bits throughout.

> DOES_IT_GENERALIZE?

• CONTEXT_OVERLOAD: yes, as a mechanism, reproduced in both runs — added cross-agent context never improved either model's solve rate or defender entropy — but not in magnitude. Qwen degrades sharply from a high baseline; Gemma is flat from an already-low baseline.
• PRIVILEGE_EFFECTS: inconclusive at this scale. Neither model showed a stable solve-rate change between normal_control and either privileged scenario across both runs (each run's single non-zero solve landed in a different privileged cell) — too small to confirm or refute Benchmark 007's larger Qwen-only finding.
• TOKEN_EFFICIENCY: neither model shows a positive return on the extra context tokens mutual sharing costs (~4x growth, no solve-rate gain for either).
• STABILITY: Gemma was more reliable in both runs — 90/90 comparable rounds and zero interrupted trials each time, versus Qwen's 89/90 (1 interruption, re-run) and 83/90 (3 interruptions, original).

ATTACK EFFECTIVENESS: 0.0-7.1% (self-play only, 1/6 scenarios)
DEFENDER ENTROPY:     36-92 bits (context-sensitive)
CONTEXT SENSITIVITY:  HIGH (collapses under mutual sharing)
STRUCTURED OUTPUT:    89/90 rounds (1 trial interrupted)
LATENCY:              mixed, scenario-dependent

ATTACK EFFECTIVENESS: 0.0-6.7% (self-play only, 1/6 scenarios)
DEFENDER ENTROPY:     38-57 bits (flat, low)
CONTEXT SENSITIVITY:  LOW (unchanged across scenarios)
STRUCTURED OUTPUT:    90/90 rounds (0 trials interrupted)
LATENCY:              mixed, scenario-dependent

> BENCHMARK_007 // PRIVILEGED_INFORMATION & ORACLE_CONTROL

Question: Why did earlier benchmarks hit a 0% solve floor? Was it capability ceiling or strict information boundaries?

Result: Under current calibrated protocol, normal_control produced measurable solves (13.3%). Attacker privilege modestly changed solve rate (16.7%), while defender privilege worsened attacker success (6.7%). Oracle controls hit 100% proving deterministic pipeline validity.

> BENCHMARK_004-006 // CO_ADAPTATION, REPLICATION_DRIFT & KNOWLEDGE_RETENTION

Question: How does bounded information sharing, replication, and accumulated knowledge affect agent behavior?

Result: Self-learning and mutual sharing drastically increased token consumption but actually reduced defender entropy generation. Over long-term retention, agents slowly adapted but with immense inefficiency.

> BENCHMARK_003 // CALIBRATION

Question: How can we identify survival of structurally weak passwords?

Result: Integrated CalibrationPolicy and a short exhaustive strategy. Flagged objectively weak targets that survive due to missing attacker coverage.

> BENCHMARK_002 // QWEN3:4B OLLAMA

Question: How does a local Qwen3 4B model perform against the arena?

Result: Zero attacker solves. Exposed that survival rate alone is a misleading metric for AI capabilities.

> BENCHMARK_001 // BASELINE

Question: Is the arena deterministic?

Result: Rule-based agents establish a reproducible synthetic baseline.