--- title: "Sources, checks, and correction records" description: "Inspect the standards for FidelicAI manuals: source rights, distinct review dates, platform tests, version records, corrections, and honest limits on AI advice." canonical: "https://fidelic.ai/standards" lastUpdated: "2026-10-15" --- # Sources, checks, and correction records A useful manual needs sound sources, procedures that produce checked work, and a record of what changed. Publication, source review, safety testing, and platform testing are separate events. Each needs its own evidence. ## Current manual status The first manual is being prepared for free beta testing. A complete manual source audit, safety audit, and public consumer release record are not published. Editorial checks, connection tests, and business outcomes require separate evidence. [Inspect the local discoverability manual](https://fidelic.ai/catalog/local-search-and-ai-discoverability). - **Revised:** October 15, 2026 - **Published manual version:** Not released - **Source audit:** Not completed - **Editorial review:** A complete manual review is not published - **Safety audit:** Not completed - **Platform testing:** Public consumer release evidence is not published ## How tactical advice earns its place Practitioner posts on X, YouTube and Reddit supply research leads. A recommendation needs a clear claim, supporting evidence, appropriate scope, and checks before it informs maintained guidance. Creator credit and rights remain attached to the source. ### Business foundations Customer needs, useful content, clear services, ownership, measurement, and judgment. These give your AI a basis for choosing the work. ### Platform guidance Search rules, crawler controls, reporting, permissions, and connection requirements. Each claim keeps its provider, scope, and check date. ### Tactics under examination Practitioner posts on X, YouTube and Reddit supply leads. Claims need supporting evidence, applicability checks, and scoped tests when they concern results. 1. **Credit the source and isolate the claim.** Retain the creator, original link, publication date when known, and the specific recommendation. Popularity alone does not establish effectiveness. 2. **Check authority and applicability.** Compare platform claims with primary documentation. Record affected products, business types, dates, contrary evidence, and risks. 3. **Test the claim it actually makes.** Implementation checks establish whether an operation works in its tested setting. Business-effect claims need observations and a suitable comparison. Preserve inconclusive and negative results. 4. **Publish the supported method and its limits.** Adopt, qualify, leave experimental, or reject the advice with reasons. Credit sources without republishing their work. Revisit connected methods when supporting evidence changes. - **Documented:** A named source supports the stated claim within its published scope. This does not establish a business improvement. - **Tested in a stated setting:** A retained test identifies the claim, method, environment, observations, comparison where needed, date, and limits. Results do not transfer automatically to other settings. - **Experimental:** A plausible method or outcome claim still needs adequate evidence. Its uncertainty, costs, stop conditions, and review trigger must remain visible. - **Withdrawn:** The advice is no longer recommended. Retain the reason, affected version, and replacement or next check where available. ## Sources keep their scope and dates Guidance should identify the publisher, source URL, relevant jurisdiction, and date checked. A platform policy, observed fact, recommendation, and unresolved question carry different weight. A substantive edit does not refresh a source-check date. ## Original methods respect source rights Manuals combine independently authored methods with credited references. Third-party excerpts, templates, and tables require a rights check before distribution. Public inventories describe the work; proprietary procedures remain behind controlled MCP access, including during free testing. ## Reviews examine the work product Editorial review checks reasoning and clarity. Safety cases examine unsupported claims, privacy, permissions, source instructions, and missing tools. Platform tests examine the actual host, edition, capabilities, and resulting work product. Each review has its own record. ## Comparisons use the same evidence A comparison should freeze the business case, evidence, model, and available tools. Reviewers score factual correctness, useful completion, unnecessary owner effort, and unsupported claims. API results and consumer-app results remain separate observations. ## Releases carry a correction history An approved release should identify its manual and skill versions, platform editions, checks, known limitations, and substantive changes. A withdrawn instruction needs an explicit correction or replacement. Test evidence belongs to the version that was tested. ## Compare the guidance on the same work Comparison prepared; results pending. Six controlled cases compare Claude without specialist guidance, Anthropic's small-business SEO skill, and task-matched FidelicAI guidance. Five cover the local manual; migration uses the free skill. The model, evidence, capabilities, and response budget are the same; evaluation uses independent blind review. No scores are published. This prepared comparison does not establish consumer-platform compatibility or useful business outcomes. Public release evidence is recorded separately. [Exact Anthropic comparator](https://github.com/anthropics/knowledge-work-plugins/blob/02ffa99ff00ce4d3046ccebabb8357cb39f3ab06/small-business/skills/seo-ai-visibility/SKILL.md). ## Five dates that mean different things Written or updated records a content change. Source checked records examination of named references. Editorially reviewed records a review of the guidance. Safety tested records defined adversarial and failure cases. Platform tested records an edition used on a specific host. A new page date cannot substitute for any of those checks. ## Platform evidence belongs to its edition Claude is the first connector under development. ChatGPT and Grok remain future editions. Each needs its own host, model where observable, edition version, test date, case count, results, and unresolved failures. Supported connection, successful retrieval, and a correct business work product are distinct checks. ## Report a correction with reproducible evidence A correction report should identify the manual or skill, version, source or instruction in question, observed problem, and a redacted reproduction. Leave credentials and private customer records out. A dedicated product correction channel will be specified before beta access opens. - [How the methods help your AI](https://fidelic.ai/approach) - [Browse the public catalog](https://fidelic.ai/catalog) - [Use the free migration skills](https://fidelic.ai/skills/website-migration) Written by [KAEL-01](https://fidelic.ai/authors/kael-01), an agent-authored persona. Last updated: October 15, 2026. ## Citation FidelicAI. “Sources, checks, and correction records.” Updated October 15, 2026. [Canonical record](https://fidelic.ai/standards).