/nai-option-study — verified architecture options

You drop a project request and your EA-repository exports into 08 Option Studies/inputs/<slug>/ and type /nai-option-study <slug>. It reviews the request against your architecture context, proposes a handful of options with trade-offs, checks each one against documented real-world use, and hands you a study whose every recommendation is either evidenced or explicitly flagged as thin. You make the trade-off call.

What makes this one different from the other three is the shape of the middle. The others fan out once and synthesize. This one loops. After a one-time context review, the orchestrator runs a bounded generate–verify–gate cycle: option-recommender proposes options, option-verifier researches each for documented use and verdicts it PASS, THIN, or FAIL, and the orchestrator gates — a FAIL sends the options back for another round, a THIN never does, and the bound is absolute at three rounds. The critique travels between rounds as a file on disk, not a conversation, so each agent stays in its own context.

%%{init: {'theme':'neutral', 'flowchart':{'htmlLabels':false,'nodeSpacing':45,'rankSpacing':55}}}%%
flowchart TD
    H["Human types /nai-option-study"] --> C

    subgraph L1 ["1 - Orchestrator (.claude/commands)"]
        C["/nai-option-study<br/>reviews once, then loops<br/>max 3 rounds, gates, synthesizes"]
    end

    C -->|"Wave 1: once"| R["ea-context-reviewer"]
    R -.->|"context-review.md"| C

    subgraph L2 ["2 - Bounded loop, max 3 rounds"]
        P["option-recommender<br/>no web tools"]
        V["option-verifier"]
        G{"gate: FAIL with rounds left?"}
    end

    C -->|"each round"| P
    P -.->|"options.md"| V
    V -.->|"verification.md: PASS / THIN / FAIL"| G
    G -->|"yes: revise"| P
    G ==>|"no: stop"| OUT
    C ==> OUT["option-study.md + option-study.html<br/>you decide the trade-off"]

    R --> T1
    P --> T1
    subgraph L3 ["3 - Tools (the only route out)"]
        T1["Read / Write<br/>round files on disk"]
        T2["WebSearch / WebFetch"]
    end
    V --> T2

    classDef orch fill:#dcefe9,stroke:#0f6f63,stroke-width:2px
    classDef tool fill:#f4f1ea,stroke:#8a7a5a
    classDef human fill:#fdf3e0,stroke:#8a5a00,stroke-width:2px
    class C orch
    class T1,T2 tool
    class H,OUT human

The boundary is the same idea as the newsletter writer, sharpened. option-recommender has no web tools — it cannot cite. Evidence enters the study through exactly one door: option-verifier, the only agent that can reach the web. So a recommendation that survives is one an independent, differently-abled agent found real support for — and one that does not is surfaced as such, never quietly dropped. Absence of evidence is a finding, not a gap to paper over.