Skip to content

Research and Evidence Principles

Every research artifact produced for this company — market research, competitive analysis, customer feedback summaries, technical investigation notes — must clearly distinguish the kind of claim it is making. This applies whether the artifact was produced by a human or an AI agent.

Required evidence categories

Every substantive claim in a research artifact should be labeled as one of the following:

Category Meaning Example
Verified fact Directly confirmed, sourced, and checkable "Etsy's seller fee is 6.5% as of the cited fee schedule page."
Observed evidence Something directly seen/measured, not inferred "Three test users completed the onboarding flow without assistance during a moderated session on [date]."
Customer statement Something a customer said, attributed and quoted or accurately paraphrased "One beta user said the pantry-expiry alert was 'the main reason I kept using it.'"
Hypothesis An untested idea proposed for validation "We hypothesize that a $9/month price point converts better than $15/month for this segment."
Assumption A working belief adopted without direct evidence, because a decision has to be made anyway "We are assuming most target customers already use a smartphone reminder app."
Inference A conclusion drawn by reasoning from other evidence, not directly observed "Given the fee schedule and typical order size, expected marketplace take-rate is approximately X%."
Recommendation An action proposed based on the above "Recommend testing both price points before committing to one."

A research artifact that mixes these categories without labeling them is not acceptable output — the reader should never have to guess whether a number is a measurement or a guess.

Required metadata on research artifacts

Where practical, research artifacts should record:

  • Source — where the information came from (a specific page, dataset, conversation, or person).
  • Access or publication date — when the source was retrieved or when it was published.
  • Scope — what the research does and does not cover (e.g., "US market only," "based on 5 interviews, not a statistically representative sample").
  • Limitations — known weaknesses in the method or data.
  • Confidence level, where meaningful — e.g., high/medium/low, tied to the strength of the underlying evidence category.

Explicitly prohibited: fabrication

The following must never be fabricated, estimated-and-presented-as-measured, or invented by an AI agent or a human under time pressure:

  • Market size figures
  • Search volume figures
  • Marketplace demand data (e.g., Etsy search/demand signals)
  • Revenue estimates presented as fact rather than clearly labeled projections
  • Competitor pricing, presented without a real, checkable source
  • Customer quotes that were not actually said by a real customer
  • Test results that were not actually run
  • Product reviews that were not actually written by a real reviewer

If a number or quote is needed but no real source exists, the correct move is to say so explicitly ("no verified data available; the following is a rough assumption for planning purposes only") — not to produce a plausible-looking fabrication. This is a hard rule, not a style preference; see docs/ai/prohibited-behaviors.md for how this is enforced specifically for AI agents.

Relationship to the assumptions register

Assumptions that materially affect a decision — not just research-artifact-local guesses — should be surfaced to docs/governance/assumptions-register.md so they remain visible and can be revisited as real evidence becomes available.


Report an issue about this page