You have a fixed spend for a PyRIT engagement. How do you divide it between many objectives at a short turn budget and few objectives explored deeply, and how do you report what the money bought?

The tempting allocation is one turn budget applied to everything. It is the worst one available, because it spends exactly as much on an objective that will never move as on the one that would have broken on turn twelve — and it spends it on the wrong side of that pair. ### Terms An **objective** is one goal pursued in one conversation. The **turn budget** is PyRIT's `max_turns` on the attack: a ceiling on how many adversarial-target-scorer cycles that conversation may run before the loop gives up. Each turn bills three legs, so per-objective spend is roughly `max_turns` x 3 calls, and more once converters fan out or a second scorer is attached. The engagement's fixed spend has to be divided across objectives, and the only levers are how many objectives you attempt and how deep each is allowed to go. ### The staged allocation Run **stage one** as a short-`max_turns` sweep across the full objective set. Its purpose is not to clear anything — a short budget cannot, and the sibling failure mode is precisely that a too-short run reports refusal where a longer one would have found a breach. Its purpose is to **rank**: objectives that eventually break tend to show early signal — partial compliance, negotiation, a drift in the target's stance across two or three turns — and those are cheap to detect. Run **stage two** as long-`max_turns` runs against the top of that ranking, because multi-turn attacks land in depth, not breadth. Worked, to show why the split is affordable: 40 objectives at `max_turns` 5, one variant, one scorer is 5 x 3 = 15 calls each, 600 calls for total coverage. Promoting 8 of them to `max_turns` 30 is 90 calls each, 720 calls. Full breadth costs less than eight deep runs. ### Where that number misleads Call count is linear in turns; **spend is not**. The adversarial chat is conditioned on the whole transcript, so its input grows with turn number and its cumulative token cost over an objective grows roughly with the square of the turn count. A 30-turn objective therefore costs far more than six times a 5-turn one, even though it makes six times the calls. Any budget built from a per-turn average measured on the shallow sweep will under-forecast the deep stage badly. The second inversion is more important. The loop terminates on success or on the ceiling, so **the objectives that produce nothing are the expensive ones** and the ones that break on turn three are the cheapest. Ordinary intuition has this backwards. It means a pass with a low success rate costs close to its theoretical worst case, and a budget computed from average per-objective cost will be wrong in exactly the direction that hurts. Model the worst case — every objective at its cap — and treat any early termination as recovered budget for stage two. ### Reserves Two are non-negotiable. One for **reproduction**: a finding that does not reproduce is not reportable, and re-running a successful objective costs roughly what it cost the first time. One for **follow-up**: the useful output of stage two is usually an objective nobody wrote down at kickoff, and an engagement with no slack cannot chase it. Cap per-objective spend as well, so a single runaway conversation cannot consume the pass. ### What to report Cost per confirmed finding is the number leadership understands, and on its own it rewards the wrong behaviour: it looks best when you only run objectives you already knew would break, which is the opposite of discovery. Pair it with three others — objectives attempted, **objectives that terminated at their cap**, and the split of spend between the breadth and depth stages. The middle one is the honest part and the one most often missing. An objective that exhausted its turn budget without a hit is unfinished, not cleared; merging it into a pass rate manufactures a clean result out of an under-funded run. State it plainly, because whoever approves the spend will read a low finding count as good news about the system, when it is at least as likely to be news about the budget. That single line is what keeps the two readings distinguishable, and it is what lets the next engagement's budget be argued from evidence rather than from last time's total.