“I found a strong pattern last week” is difficult to check without the filter, source and observation period. A screenshot preserves what was visible, but it may omit the criteria that produced it. A short research note makes the comparison easier to revisit.
The template here works with a manual table review or an MCP-assisted session. It is an original practical suggestion, not a claim that OddsTips automatically exports every field listed below.
Begin with the question and its boundary
Write one question that could be answered from the available data. For example: “Among completed fixtures from the chosen competition and source, how often did the specified opening-to-closing movement occur during this calendar window?” Identify the market and its settlement definition before entering counts.
Keep the period used to discover the question separate from any later test period. If you changed the rule after looking at the result, label that version exploratory. A note that records the change is more useful than one that presents the final rule as if it had existed all along.
Keep a small evidence card
- Question: the outcome or movement being measured.
- Source: bookmaker, dataset and available version information.
- Scope: sport, competitions, dates and timezone.
- Criteria: columns, bounds, endpoint rules and reference fixture, if any.
- Counts: eligible fixtures, complete records, returned rows and unique fixtures.
- Retrieval: when the records were obtained and whether more pages remained.
- Limitations: missing fields, uncertain timestamps and changes made during the investigation.
W3C's provenance work treats the origin and processing history of data as information that helps assess it. You do not need a complex format to apply that idea to a small odds investigation. Consistent labels and preserved source references are a useful start.
Distinguish observation from interpretation
An illustrative observation might be: “The condition appeared in 14 of 25 complete records.” An interpretation might be: “This merits testing in a later period.” Keep those sentences separate. The first can be checked against a count; the second proposes what to do next.
Do not silently expand the claim to all fixtures if incomplete records were excluded. If you have only the first result page, state its size and leave the overall total unknown unless the service supplies it.
Record an MCP session without recording secrets
Retain the research prompt, applied criteria and permitted result summary. A saved-filter name is useful context, but its actual definition matters more. Leave connection tokens, private authentication headers and unrelated account details out of a shared note.
When repeating the study, compare the new criteria and coverage against the old card before comparing the headline percentage. A changed source, larger archive or different market may explain the result. The note's purpose is to make that explanation possible, not to turn a historical finding into a prediction guarantee.
Sources and method
Original research examples and prompts, not observed match results or performance claims. Prepared with AI assistance; source statements, examples and calculations were checked.
- W3C: PROV overview and data provenance — accessed 13 October 2026