Guardrails

• do not answer current questions from stale memory when fresh search is cheap
• separate:
  • sourced fact
  • user-provided evidence
  • inference
  • recommendation
• do not spin up a heavyweight research pass if the answer is already in local code or docs

Workflow

1. Start from what the user already gave you

Normalize any supplied material into:

• already-evidenced facts
• needs verification
• open questions

Do not restart the analysis from zero if the user already built part of the model.

2. Classify the ask

Choose the right lane before searching:

• quick factual answer
• comparison or decision memo
• lead/enrichment pass
• recurring monitoring candidate

3. Take the lightest useful evidence path first

• use exa-search for fast discovery
• escalate to deep-research when synthesis or multiple sources matter
• use market-research when the outcome should end in a recommendation
• hand off to lead-intelligence when the real ask is target ranking or warm-path discovery

4. Report with explicit evidence boundaries

For important claims, say whether they are:

• sourced facts
• user-supplied context
• inference
• recommendation

Freshness-sensitive answers should include concrete dates.

5. Decide whether the task should stay manual

If the user is likely to ask the same research question repeatedly, say so explicitly and recommend a monitoring or workflow layer instead of repeating the same manual search forever.

Output Format

QUESTION TYPE
- factual / comparison / enrichment / monitoring

EVIDENCE
- sourced facts
- user-provided context

INFERENCE
- what follows from the evidence

RECOMMENDATION
- answer or next move
- whether this should become a monitor

Pitfalls

• do not mix inference into sourced facts without labeling it
• do not ignore user-provided evidence
• do not use a heavy research lane for a question local repo context can answer
• do not give freshness-sensitive answers without dates

Verification

• important claims are labeled by evidence type
• freshness-sensitive outputs include dates
• the final recommendation matches the actual research mode used

santa-method

通过双评审智能体对结果进行对抗式校验，提升输出发布前的可靠性

research

针对技术、市场与竞品等主题开展多源深度调研并输出综合洞察。