1. Identify who manages the place
We determine which agency or land manager controls the park, beach, or public area. A statewide rule may not answer a city-park question, and a beach rule may not apply to the rest of a park.
2. Use official sources
We use laws, agency rules, permit pages, official maps, park policies, and official FAQs. Forums, social posts, hobby sites, and commercial articles may point to a lead, but they do not establish the published answer.
3. Match the rule to the exact place
We check the geographic scope, land manager, designated areas, seasons, permits, tools, digging rules, and found-item requirements. A property can follow a park-system rule while still having local limits.
4. State only what the source supports
We separate detector use from permission, digging, recovery, and keeping or reporting finds. When the official sources do not give a clear answer, the page says so.
5. Show the source and date
Every record links to the official source and shows when it was last checked. Rules, maps, closures, and posted conditions can change.
6. Handle updates
When an official source changes or cannot be checked, the affected answer may be marked as updating or cautionary. You can still open the page to find the official sources. A change to the answer receives a dated public update.
Two-pass AI source audit
The first AI pass extracts evidence and checks material claims against official pages, rules, permit instructions and maps. It records the source, heading or page, check date and preserved source snapshot. Relevant PDF pages are rendered and inspected; colored map boundaries are checked visually.
A second AI pass in a separate model context challenges the evidence, jurisdiction, property scope, inherited rules, authorization, recovery, finds, campgrounds, seasons, hours, tools and wording. Disagreements are retained. Unsupported wording is narrowed or withheld and rechecked before publication.
Audit outcomes
- Source-supported: the reviewed sources support the stated claims and conditions.
- Source-supported with limitations: the practical answer is supported, with a disclosed operational or evidence limitation.
- Not confirmed by the reviewed official sources: a material question remains unanswered.
- Conflicting official sources: official sources materially disagree; the disagreement remains visible.
- Withdrawn: the previous answer is withheld because it cannot be supported.
Reconciliation never chooses wording stronger than either completed pass supports. A missing detail is not permission. A conservative editorial suggestion is labeled separately from an official requirement.
Reproducible evidence and dates
The audit archive records actual model and run identifiers, protocol and prompt versions, reviewed claims, source snapshots, image and document digests, input records and findings. Each later audit creates a new historical event. Preserved inputs make the assessment inspectable; rerunning an AI model does not promise identical prose.
A source-check date records inspection of that source. Audit completion records the completed two-pass process. Neither is a human-review date. Historical named reviews cover only the claims checked at that time. Read what historical named reviews cover.
An AI source audit is not legal advice, human review, government approval, legal verification or guaranteed permission. It does not replace permits, written permission, signs or current land-manager instructions.