Doubt v0.3.1 — evidence maps readers can test

[alsoleg89]

Update: v0.3.1 is now the latest release. It fixes remote GitHub CLI
discovery and has been tested end to end against the public repository:

gh skill install alsoleg89/doubt doubt

The clean test selected v0.3.1, discovered all four skill files, and installed
them with source/ref tracking. v0.3.0 had the project-level
.github/skills/doubt layout but not the separate skills/doubt publisher
layout required by remote gh skill; the patch adds it and CI now byte-checks
both layouts against the canonical skill.

The main v0.3 launch remains the reader-testable evidence-map loop:

1. Explore the Agent Skills vs MCP map
   and challenge one edge in the review issue.

2. Take the five-topic map-vs-memo study.
   It runs locally, sends nothing over the network, and downloads an anonymous
   result JSON. Condition-level analysis stays locked until ten complete unique
   sessions.

3. Install the workflow with the verified command above, or use:

   npx doubt-ai init --agent all

The structural contract benchmark is 12/12, but that establishes structural
traceability and safe rendering—not source truth, faithful AI extraction, or
reader preference. The new study exists because those claims require evidence.

Latest release and tarball

If you try one thing, take the reader study. If an edge looks wrong, comment on
issue #5 with the exact source region that changes it.

[alsoleg89]

Update: zero-install playground is live

You can now validate and render a Doubt map without npm or an agent:

https://alsoleg89.github.io/doubt/playground/

• paste, open, or drop a .doubt.json file;
• inspect fail-closed contract findings;
• receive the same canonical SHA-256 receipt as the Node CLI;
• preview the existing interactive renderer;
• download portable JSON or self-contained HTML.

The browser imports the same evidence-contract module used by the CLI, Action,
and benchmarks. User map content stays in the tab: there is no account,
telemetry, analytics, or network submission. The rendered map runs in a
sandboxed iframe without same-origin access.

The public Agent Skills vs MCP example produces receipt
0444afe44c523678f2ad8eb7267e0d7c7a392709921abf16474e68d2ef5a3991
in both paths.

If you can make the playground accept a structurally misleading map, please
open the smallest failing JSON as an evidence-contract issue.