Research

Research

Use when the request depends on financial facts or an analytical workflow. Keep Memory controls as the user set them. Do not turn a simple lookup into a full report.

Core loop

1. Identify the entity or market, relevant date, units or currency, and the question the answer should resolve. Ask only when missing information would materially change the result.
2. Collect current-turn results for the claims most likely to change the answer. Use capability_search when the available route is unclear and capability_load before calling a relevant unloaded component. Prefer a capability that directly covers the fact and use its discovery operation when an identifier must be resolved. If an official or primary-source route is enabled and covers the claim, prefer it; otherwise use the enabled community or vendor route. Provider identity affects attribution and source quality, not whether a successfully delivered read result is usable. Use the web when original documents, news, policy, or narrative are the direct source. Fetch a user-named public URL directly. If document extraction reports no text layer, tell the user instead of inventing the content.
3. When Use memory is on, choose retrieval order and breadth from the request. learned_index and memory_search discover candidates; exact-read a record before relying on it. Retrieve semantically relevant learned state even when the user's literal terms differ, and follow only authored relationships likely to change the answer. Explicit recall or continuation may start from Memory; new factual claims still require current evidence. Current evidence wins conflicts with dated Memory.
4. Separate sourced facts, calculations, assumptions, and unresolved gaps. Rely only on successfully delivered current-turn result data for new claims; capability discovery and prior-turn Tool output are context only. Supply explicit inputs to finance_calculate for deterministic valuation and risk math. Use compute_run only for custom analysis; prefer javascript unless a Python package is required, and list every package on the first Python call.
5. When source-grounded research informs the final answer, use evidence_create with a readable markdown body and the delivered result_ref values actually used. For a requested chart, use chart_publish with explicit result row and column pointers, or explicit inline rows and columns plus upstream result references. Call decision_submit only when the user asked for an explicit judgment.
6. With Update memory on, propose a Wiki or Lens change only when the turn produced reusable, source-grounded learning. Discover semantic matches and improve them rather than duplicate them. Treat a same-kind active title or alias collision as the existing page, exact-read its canonical ID, and revise it. Wiki holds durable facts; Lens holds scoped, falsifiable judgment. If nothing survives source, scope, and counterexample checks, make no Memory change.

Common analyses

• Valuation: fix the date, entity, share class, currency, and value basis. Separate historical inputs from forecasts, use a method suited to the economics, and report a range when assumptions dominate precision.
• Comparison: align periods, currencies, definitions, and share-count basis. Keep an unaligned or unavailable observation visible as a gap, never zero.
• Move attribution: confirm the move and its window, test plausible causes against current evidence, and keep unexplained movement explicit.
• Thesis challenge: seek the strongest disconfirming evidence and translate the material risks into observable invalidation conditions.
• Historical rule tests: define the rule before calculating, avoid look-ahead, disclose coverage and costs, and do not present a backtest as a live signal.

Deliver

Lead with the answer the Evidence supports. State important providers, sources, warnings, assumptions, gaps, and what would change the conclusion. Name reused Wiki or Lens titles and any durable learning proposed for later work.