Syl Pulse

Estimate a causal effect from a dataset and answer a business question.

Provide exactly one of csv (CSV text) or data (list of row objects). Returns the full long-form Pulse report (the same answer an email reply contains) as the primary text content and as report in the structured result, followed by chart PNGs as image content blocks (unless include_charts is false). Structured fields: the adjusted effect size (ate/effect_size) with its confidence interval (ci_low, ci_high, ci_range), per-channel effects (channels, keyed by outcome), confidence, and caveats. A test_design (how long a spend-cut test would take to confirm the effect, and whether the data already suffices) is included only when the question asks about sample size, power, or how long a test would take, or when include_test_design is true; otherwise it is null. With include_guardian_verdict it also returns guardian_verdict {verdict: pass|caution|hold, reason}, computed locally. All numbers are deterministic statistics, never LLM-generated.

Args: question: The business question, e.g. "Does ad spend drive sales?". csv: Dataset as CSV text. data: Dataset as a list of row records (alternative to csv). cause: Treatment column (inferred from the question if omitted). effect: Outcome column (inferred if omitted). confounders: Columns to adjust for, in addition to those discovered. outcomes: Several outcome columns, for by-channel effects. time_col: Time column (used for the test-design cadence). include_guardian_verdict: Also return a pass/caution/hold verdict. include_charts: Attach chart images (DAG, effect, trend) as image content. include_test_design: Force the spend-cut test design / sample-size section on (true) or off (false). Omit (null) to include it only when the question asks about sample size, statistical power, or how long a test would take. estimator: Optional estimation method: ols_adjusted, aipw, ipw, psm or dml_forest (or "rule" for the deterministic default). Omit to follow config.estimator (default "rule"; "llm" lets the model choose from that list, falling back to the rule choice when no LLM key). config: Optional per-step engine config, e.g. {"summary": "rule"}. Keys: cleaning, discovery, estimator, summary; each "rule" (deterministic, no LLM call) or "llm" (LLM-assisted; falls back to rule and says so in engine_steps when no key is configured or the call fails). Defaults: cleaning=llm, discovery=llm, estimator=rule, summary=llm. All "rule" makes zero LLM calls. Numbers are deterministic in every mode.