# lean-ctx footprint ablation suite (#959). # # Unlike the accuracy suite, the "context" under test is lean-ctx's OWN injected # footprint (rules block, tool schemas, wakeup briefing), not a corpus. Tasks are # chosen to be footprint-SENSITIVE so removing an element can move the pass-rate: # - route-* probe whether the advertised tool schemas enable correct tool choice, # - rule-* probes whether the rules block changes behaviour, # - control-* is footprint-INSENSITIVE (pure reasoning) and must stay NonInferior, # proving the harness does not spuriously credit the footprint. {"id":"route-callers","domain":"qa","prompt":"A user asks for every caller of a function. Which single lean-ctx tool should you call first? Answer with just the tool name.","workspace":"footprint/ws","answers":["ctx_callgraph"]} {"id":"route-symbol","domain":"qa","prompt":"Which lean-ctx tool finds the exact definition of a named symbol? Answer with just the tool name.","workspace":"footprint/ws","answers":["ctx_symbol"]} {"id":"route-read-large","domain":"qa","prompt":"To read a large source file with the fewest tokens, which lean-ctx tool do you use? Answer with just the tool name.","workspace":"footprint/ws","answers":["ctx_read"]} {"id":"rule-before-edit","domain":"qa","prompt":"Per the lean-ctx rules, what must you do to a file before editing it?","workspace":"footprint/ws","answers":["read"]} {"id":"control-arithmetic","domain":"qa","prompt":"What is 17 plus 25? Answer with the number only.","workspace":"footprint/ws","answers":["42"]}