Recommendation rules, evidence, and independence

Sources and identity

OpenGPT preserves the publisher's model name or ID and configuration. An official model ID is added only when provider documentation supports the exact mapping. Source records that cannot be safely mapped remain visible as source records rather than being assigned an invented identity.

Recommendation method

Builder first applies hard privacy and listed-budget eligibility gates, then adds published rule points for task and goal fit, role, industry signals, language, platform, and experience. Eligible candidates are sorted deterministically; equal scores use name order. These weights express OpenGPT's editorial product-fit policy, not percentages, benchmark scores, or measured model quality.

Factual and editorial boundary

Provider-published identity, access, capability descriptions, and listed pricing are factual source claims and must link to the named provider when used. Fit tags, strengths, classifications, assumptions, confidence, and recommendations are OpenGPT editorial judgments. Public benchmark records provide scoped context and are not used as hidden points in the current recommendation rules.

Per-result transparency

An integrated recommendation should display Why, Evidence, Assumptions, Last reviewed, Confidence, Best alternative, and When not to choose, plus the conditions that make the recommendation applicable. Confidence describes completeness of the fit explanation and supporting record; it is not a probability of good performance.

Review cadence

Methodology and active recommendation records are scheduled for review at least every 90 days. A material provider, pricing, access, lifecycle, source, catalog-classification, scoring-rule, or reproducible correction change triggers earlier review. Each result uses its own record review date; a page update date is not evidence that every underlying record was reviewed that day.

Ranking and comparison

Ranks remain scoped to the named source, metric, snapshot, benchmark version, and comparable configuration group. OpenGPT does not create a universal Agent score across incompatible benchmarks. Costs, latency, repeat-run evidence, and intervention are shown only when the source reports comparable values.

Submission and review

A user submission first enters editorial review. A reviewer may approve or reject the evidence record. An approved Agent product is automatically listed in the searchable public Agent directory with its reviewed public source. Static catalog identity, comparison eligibility, and leaderboard changes remain separate operations and require their own evidence.

• Submitters cannot choose rank, trust grade, comparability, or publication status.
• Review decisions should record the evidence checked and the reason.
• Directory listings are not endorsements or rankings.

Corrections and appeals

Send corrections to contact@opengpt.com with the affected URL, the exact statement or identity mapping at issue, a public source, and a concise explanation. OpenGPT may correct labels, links, mappings, dates, or evidence status; preserve an audit trail where the workflow supports it; and withdraw a candidate record while a dispute is reviewed.

Commercial independence

OpenGPT does not sell rank, trust grade, review approval, comparison eligibility, or recommendation labels. Sponsorship or commercial support, if introduced, must be clearly disclosed and cannot alter scoring rules, source selection, evidence thresholds, correction handling, or editorial placement.

Conflicts and limitations

Provider reports and submitter relationships are evidence context, not automatic disqualification or validation. OpenGPT labels the available evidence and keeps uncertainty, small samples, conflicting results, and unreported fields visible.