Settle

**The problem.** A product manager asks for a feature: multi currency, team accounts, single sign on. Engineers read the code and estimate "about two days". Three days in, they hit something nobody knew was there: a column that cannot hold the new values, a query that silently gives wrong answers, a cache shared across customers. The estimate was not wrong because the engineers were careless. It was wrong because the risky parts were only discovered by building them. **The solution.** Settle moves those discoveries to the start. You give it one sentence, the feature you are about to estimate, and it runs against your real repository: 1. **Bob finds the risks.** IBM Bob reads the code the feature would touch and names the assumptions the feature silently depends on, phrased as what must be true for the feature to be as easy as it looks. 2. **Bob runs experiments.** One Bob Shell session per assumption, all at once, each in its own git worktree, writes the smallest test that proves or disproves that assumption against today's code. 3. **Settle checks the work.** Settle reruns every experiment itself instead of trusting Bob's word. Passing means proven. Failing means blocked, with the exact error or wrong value. Anything that could not run, or that edited existing code or tests, is unknown. 4. **Bob writes the plan** from what was actually found: what we now know, the landmines to fix first, the build order, and how the findings change the estimate. The result is an evidence board a whole team can read: plain words for the product manager, the failing test and the branch for the engineers. **Target users.** Engineering teams during planning and estimation: tech leads, engineers, and the product managers who need an honest size before committing to a date. **What the demo shows.** The request: "Let customers pay in their own currency (USD, EUR, JPY and IDR), and keep the top customers ranking correct." In about five minutes, Bob names three assumptions (revenue totals can hold large amounts, orders can record a currency, and the ranking compares revenue fairly across currencies), builds an experiment for each in parallel, and Settle reruns them. The evidence board shows which assumptions fail against today's code and why, with the real error output, then Bob's plan puts those landmines first. **Why it is new.** An AI can guess what might go wrong with a feature. Settle proves it, with experiments against the real code, verified independently of the agent that wrote them. Earlier planning tools predict risks from text; Settle turns each risk into a runnable test before anyone commits to an estimate.