"What is RedTeamLLM's core architectural contribution?"	The summarize-reason-act decomposition: three separate LLM calls per iteration. Summarize compresses context before reasoning. Reason plans (with correction) over the compressed state. Act executes. This keeps the loop coherent on long engagements where a monolithic loop drifts.	course2a::sdd02::recall
"Why three LLM calls instead of one monolithic reasoning call?"	A monolithic loop faces three compounding failure modes on long engagements: context overflow (truncation loses info), plan drift (objective buried in noise), memory loss (early findings truncated). Decomposition addresses all three: summarize prevents overflow, reason-on-compressed-state prevents drift, persistent memory prevents loss.	course2a::sdd02::analysis
"What is 'plan correction' in RedTeamLLM, and which Course 1 weakness does it address?"	The reason step sees the prior action's outcome and the summarized state, so it can REVISE the plan when an action fails or surprises. This explicitly handles the error-compounding weakness that Course 1 Module 1 flags as the dumb-loop's failure mode.	course2a::sdd02::analysis
"Why is persistent finding memory separate from the conversation context?"	So findings SURVIVE context compression. If findings lived only in context, summarization/truncation would lose early discoveries — exactly the findings that matter. Separate memory means a finding recorded at step 5 is still available at step 50.	course2a::sdd02::analysis
"RedTeamLLM vs CAI: what is each one's advantage, and what is the strongest architecture?"	RedTeamLLM = DEPTH (plan correction, memory, context mgmt for a single agent). CAI = BREADTH (6 domains + bug bounty + meta-harness + scope enforcement). Strongest = the HYBRID: RedTeamLLM's summarize-reason-act loop inside each of CAI's domain specialists, wrapped by CAI's CSI routing and hard-wired scope.	course2a::sdd02::analysis
"What is the 'summarize-step injection surface' — the novel risk RedTeamLLM introduces?"	Adversarial tool output enters the summarize step; if injected content SURVIVES compression, it reaches the reasoner and can manipulate the plan. Worse, a manipulated plan can write a FALSE finding into persistent memory, poisoning conclusions. A monolithic loop does not have this surface. The audit target: the summarizer's handling of untrusted content.	course2a::sdd02::analysis
"State RedTeamLLM's benchmark claim."	Outperforms PenTestGPT on approximately 60% of tested CTF challenges (arXiv:2505.06913). Gains attributed to plan correction, memory, and context handling — the three things the three-step decomposition addresses.	course2a::sdd02::recall
"State RedTeamLLM's score on the 12-module rubric and its strongest/weakest modules."	41/60. Strongest: Context Mgmt (5) and Error Handling (5) — exactly what it was designed to fix. Weakest: Subagents (1, single-agent), Tool Design (3, narrower than CAI), Sandboxing/Permission (3, less detailed than CAI). Lower total than CAI's 46 reflects the breadth-vs-depth tradeoff.	course2a::sdd02::analysis
"Name 3 things RedTeamLLM does BETTER than any other harness in the roster."	(1) Active context management (summarize before reasoning, not after overflow). (2) Plan correction (reason step revises on failure). (3) Coherence over long engagements (summarization + memory + correction vs monolithic drift).	course2a::sdd02::recall
"Name 3 things you would fix if you forked RedTeamLLM."	(1) Add CAI's CSI meta-harness for domain routing (Module 10). (2) Add CAI's hard-wired scope enforcement (Module 5). (3) Audit and harden the summarize step for adversarial content (the novel injection surface).	course2a::sdd02::application
"What is the cost tradeoff of the three-step decomposition?"	3x the LLM calls per iteration (summarize, reason, act). Summarize and reason are reasoning-heavy (tokens-reasoned, not tokens-moved). For long engagements, the coherence benefit outweighs the cost. For short tasks, a monolithic loop is fine.	course2a::sdd02::analysis
"State RedTeamLLM's Architect's Verdict in one line."	The definitive study of keeping an offensive agent coherent over long engagements via summarize-reason-act decomposition; sacrifices breadth and a detailed security model for depth on context, memory, and plan correction; build on its loop and pair with CAI's routing and scope enforcement.	course2a::sdd02::recall
"Does RedTeamLLM have a meta-harness / domain routing like CAI's CSI?"	NO — RedTeamLLM is single-agent, pentest scope, no domain routing. This is its weakest module (Subagents: 1/5). The strongest architecture is the hybrid that adds CAI's CSI routing.	course2a::sdd02::recall