# Documentation reading trial This is a bounded editorial check, not an agent-performance benchmark or a release gate result. The September 2026 trial used an independent agent with no preceding implementation discussion. It started at `docs/agents.md` and `docs/readme.md`, followed their links, answered three tasks, then inspected source to check the answers. It did not execute an application. Its exact model, invocation trace, and source snapshot were not retained, so this is qualitative editorial history rather than a reproducible evaluation. ## Tasks and expected decisions | Task | Required decisions | | --- | --- | | Existing Backbone models and router; a detail request completes after another navigation | Retain the router and models; configure DataApi on the appropriate View classes; choose StateApi separately; cancel and suppress stale completion; let the Region own replacement; verify slow/fast success and failure, teardown, and browser history. | | New application needs an observable list and local selection | Start with the native data package for collection observation; plain local selection is sufficient unless subscription is needed; supply a child View; retain default DOM/rendering; do not introduce a router or isolated runtime without a requirement. | | Adopt prerendered Region content and later replace it | Wrap the existing populated element in a View, then establish Region ownership with show; resolve existing elements from actual DOM; replace through the Region and verify destruction and mount preservation. | The agent reached these decisions and checked them against the source. Its reading paths included the integration decision guide, routing, Application, StateApi, provider package readmes, prerendered DOM, Region, and lifecycle references. ## Defects the trial exposed - The Region replacement example omitted its required markup and put a dot inside `className`. Its prose also attributed replacement to the wrong operation. - Backbone setup examples selected StateApi even for a data-only task. - The native data package's first example introduced runtime isolation and state setup before establishing a need for either. These findings led to corrections in the canonical pages and a follow-up executable check for the Region example. They demonstrate why navigation, defaults, and copyable setup must be reviewed together. ## Repeat after material changes Give a reader the task and the two entry pages, without the expected decisions. Record the links they follow and where they guess. Check their answer against source and run the relevant executable example or application interaction. Correct the earliest page that led to a wrong choice, then repeat that task. Use additional tasks for immutable collection replacement, shared-state disposal, renderer selection, migration, and diagnostics as those guides change. Record exact package/source and test commands with each run. Do not derive comparative accuracy, token savings, or reliability percentages from this three-task pilot. The [agent benchmark contract](https://github.com/marionettejs/marionette/blob/13f4954c352e646c413091ffdd83f6da59404573/benchmarks/agent/README.md) governs broader paired evaluations and remains a separate body of work. ## Context7 public-index check The response and indexed commit were not archived. This is an uncaptured qualitative observation, not a reproducible retrieval benchmark. On September 8, 2026, the public `/marionettejs/marionette` index completed with 1,300 snippets. This indexed the existing public branch, before these local docs and `context7.json` changes were published; it did not validate the proposed scope. The topic query `routing Backbone router DataApi StateApi` returned migration and installation snippets that configured both APIs, and a routing snippet that implied those adapters were required by Backbone.Router. That recommendation is incorrect for a router-only task. The canonical routing, installation, provider, and upgrade examples now distinguish each adapter's capability. The proposed Context7 exclusions also keep before/after migration material out of general retrieval; those pages remain available directly for explicit upgrade tasks. Public registration is complete. Ownership verification, configured indexing, and a repeat retrieval check remain publication steps: the proof and configuration must first be available on the public default branch. Do not describe this initial index as curated or make it the required agent entry. After publication, repeat this query and the three tasks above, inspect the source links and setup, and reject results that select StateApi or DataApi solely because a router is present. No paid plan, shared API credential, or website dependency was introduced. ## Independent implementation trials A subsequent trial gave three fresh agents installed packages, the portable skill, and a task each: latest navigation, editable list updates, and widget lifetime. All three submitted implementations passed the withheld acceptance checks after completion. The [saved implementations and reports](https://github.com/marionettejs/marionette/blob/23c44c60904a7bb9bfce9797e4849e38b7786b99/benchmarks/docs/results/2026-09-08/README.md) include the exact documentation digest and the limitations of this local snapshot. The [trial procedure](https://github.com/marionettejs/marionette/blob/23c44c60904a7bb9bfce9797e4849e38b7786b99/benchmarks/docs/README.md) withholds acceptance files until an attempt ends and does not configure paid inference. The widget reader found misleading Region reuse prose, which was corrected. These trials add implementation evidence to the earlier reading check; they remain separate from the scored release benchmark and the deferred complete application. [Canonical source](/docs/markdown/docs/maintainers/documentation-trial.md) · [Source identity](/docs/manifest.json)