## Status: follow-up experiment — hold from merge

This PR evaluates a different idea: Jev makes small opportunity and representation decisions around a generator, then separately checks the candidate’s source support and explanatory value. It does **not** add visual generation to the hosted Planner. It is based on [PR #408](https://github.com/githubnext/chopin/pull/408), so bringing it to main would also bring the #408 evaluator code. It is not an alternative to #408.

## What the three passages showed

- **Rust async RFC:** Jev said a visual could help (0.72), then chose no suitable type (0.74). Direct generation produced a flowchart spec with one label two characters over the renderer limit. A separately logged label shortening made it render; that is an inspection artifact, not a successful baseline output. - **ESLint autofix RFC:** Jev chose a table (0.79), but the constrained generator returned prose, explaining that prose better preserved the proposal status and unresolved behavior. Direct generation produced a table that omitted the concrete base-versus-test override values. Jev’s separate checks gave that table 0.91 for source support and 0.54 for added value. - **Deno prose control:** Jev said no visual opportunity (0.37); direct generation returned prose.

Four primary model generations and seven Jev calls were used. Two separate checks of the repaired Rust diagram and two labelled synthetic critic controls brought the Jev total to eleven. One false claim was flagged at 0.05 probability of support; the other scored 0.50 and is inconclusive. These three development cases do not establish classifier accuracy or justify routing decisions.

## What this PR contains

A bounded local evaluator, frozen-input hash and offset checks, pinned prompts, stage records, renderer validation, and fixture tests. Public passage text, raw model results, and provisional human-review annotations remain in ignored local data. No production routing changes.

## Flow and inspected render



## Before reconsidering a merge

Expose exact renderer limits to generation, make the public evaluation inputs reproducible, and have a human review the actual source/output pairs. All GitHub validation, browser, and container checks passed; they do not validate the three-case conclusion.