"Why does CTF optimize for speed to flag rather than coverage?"	"A CTF is not a smaller pentest — different problem, different target. Competitions rank by flag count then time. A harness taking 40 min on a challenge a human solves in 10 is not competitive, regardless of evidence quality. CTF inverts pentest's thoroughness constraint."	c2a::s04::analysis
"Name the 6 CTF domains and their reasoning patterns."	"Web (request manipulation), Pwn (memory layout/gadget chains), Crypto (weakness detection/math), Forensics (artifact carving/parsing), Reversing (disassembly/decompile triage), Misc (stego/scripting/encoding). Each has distinct tools and context budget."	c2a::s04::recall
"Which CTF domain is the heaviest context consumer, and why?"	"Reversing — decompiler output saturates context in a handful of tool calls. Defense: aggressive summarization, retain only structural facts (function signatures, cross-references, suspicious comparisons). Forensics rarely saturates; bottleneck is tool selection."	c2a::s04::analysis
"Why does a single undifferentiated harness with all 6 tool stacks fail?"	"Three compounding reasons: (1) context pollution — Web doesn't need ROPgadget, every tool dilutes selection; (2) reasoning drift — Pwn mode is opposite of Web mode, model good at neither; (3) token waste — 6 domains of schemas paid every turn, every challenge."	c2a::s04::analysis
"What is the meta-agent-plus-specialists pattern, and what is the meta-agent's ONLY job?"	"A small classifier (meta-agent) reads challenge description + artifacts, emits a domain label, dispatches to a specialized agent with its own prompt/tool subset. Meta-agent's ONLY job is ROUTING — it does not solve challenges. Specialists solve; router routes."	c2a::s04::recall
"What happens when the meta-agent CANNOT route a challenge with confidence?"	"Goes to a TriageAgent that runs cheap probes (file type, checksec, strings frequency, HTTP response code) and re-routes with the added signal. Two-stage routing catches ambiguous cases. The router is ALLOWED to be uncertain; the agents are not."	c2a::s04::application
"Why must the router use BOTH text and file-type signals (not text alone)?"	"Text-only misroutes. A Pwn challenge described as 'analyze this binary' would route to Reversing on text alone. File-type signals (ELF with checksec showing NX disabled + win symbol) correct to Pwn. Both signals together."	c2a::s04::analysis
"Name the canonical tool stack for the Web domain and its reasoning loop."	"Tools: Burp headless, sqlmap, ffuf, curated XSS payload set (the 50 that catch 95% of reflections, not the full list). Loop: request-send-inspect-response. Escalate from error-based to blind."	c2a::s04::recall
"Name the canonical tool stack for the Pwn domain. What is the most common Pwn failure mode?"	"Tools: pwntools, GDB+pwndbg, checksec, ROPgadget, one_gadget. Loop: protections-check-then-offset-find-then-chain-build. FAILURE MODE: skipping checksec — agent reasons about ROP against a binary with stack canaries, wasting the engagement."	c2a::s04::analysis
"Name the 4 properties that produce benchmark-level CTF speed."	"(1) No approval gates — every tool call executes immediately. (2) Pre-loaded tool context in system prompt. (3) Narrow domain scope (router already restricted tools). (4) Deterministic tool routing (checklists, not rediscovery — checksec first, always). The first is most consequential AND most dangerous."	c2a::s04::recall
"State the 3-criterion test for safe HITL removal in CTF."	"Safe when ALL three hold: (a) target owned or sandboxed by operator, (b) isolated from production/third-party at network layer, (c) worst-case action is bounded (restart target, waste time). Any one fails → gate stays."	c2a::s04::application
"For a remote CTF service, which gate stays even when others are removed, and why?"	"The credential-reuse gate. The service itself is fair game, but an agent that discovers creds in the challenge and tries them against the competition's login endpoint commits the CTF equivalent of a scope violation. Credential reuse against other services is the unbounded-risk action."	c2a::s04::analysis
"Distinguish intra-challenge vs inter-challenge parallelism."	"Intra-challenge: multiple exploit hypotheses against ONE challenge, concurrently (e.g., Pwn as stack-overflow OR format-string). First flag wins; loser killed. Inter-challenge: 3–5 challenges running concurrently in isolated sessions. Both trade compute for wall-clock."	c2a::s04::analysis
"Why is shared memory forbidden across parallel solve attempts?"	"Reintroduces the context pollution that domain routing was designed to eliminate. Two PwnAgents working the same challenge would overwrite each other's intermediate state. Cure: isolated sessions, isolated tool state, isolated flag-extraction sinks."	c2a::s04::application
"How is flag detection implemented, and why as middleware rather than an agent step?"	"Regex sweep over EVERY tool's structured output, as MIDDLEWARE (not a step the agent remembers to call). Agent does not 'submit a flag' — it runs tools and middleware notices flag patterns. Same pattern as S02.3 evidence logger. Patterns: flag{...}, CTF{...}, competition prefix."	c2a::s04::recall
"Why does a verifier step exist between regex match and flag submission?"	"Regex finds flag-shaped strings in challenge descriptions, examples, and decoy output. Submitting decoys burns attempts and trips rate limits. Verifier (small model or operator) confirms the flag is plausibly the answer to THIS challenge before submission. Allowlist of flag sources."	c2a::s04::analysis
"What are the 3 CTFd submission response states, and how do you handle each?"	"correct → mark solved, kill parallel attempts, free slot. incorrect → log, agent continues. already_solved → treat as correct (parallel session or teammate got there first), do NOT re-submit (burns attempts, trips rate limits)."	c2a::s04::application
"Why rate-limit the submission client, and to what?"	"Most scoreboards throttle or temporarily ban teams that submit too fast. 2–3 seconds between submissions; per-challenge lockout after N incorrect attempts. Self-inflicted throttling ends competitions."	c2a::s04::application
"What does the CTF scorecard track, and why is it the operator's only sane view?"	"Per challenge: state (unattempted/in_progress/solved/abandoned), sessions, attempts, first_blood_at. Global: flags_captured (source), flags_submitted (CTFd response). With 5 concurrent sessions, it is the only way to know what is happening. Metrics: challenges/hr, MTTF per domain, router accuracy, flag-extraction precision."	c2a::s04::recall
"Why is the 'skip checksec/file before reasoning' anti-pattern so damaging in CTF?"	"Pwn agent builds ROP chain without reading protections = wasted engagement if binary has stack canaries. Forensics agent runs binwalk without file = wrong tool, wasted turn. CTF turns are time-budgeted. Cure: ordered checklists in prompt, first step always the cheap probe that constrains the rest."	c2a::s04::analysis