Clarity Contract

Two models agreed. Their example still assumed a fact.

Astra and Fable co-wrote a style guide. When asked to challenge it, Fable caught an unsupported assumption in their own “better” error message: a 25 MB upload limit that had never been established.

So we turned the guide into a contract you can inspect, reuse and try to break.

Live playground · Copy the agent guide · Exact experiment · MIT license

Save this part

Truth and consequences > access > useful next step > brevity.

Never invent a cause, capability, action, deadline or promise
just to make the answer sound more helpful.

Unknown cause? Preserve the uncertainty.
Irreversible action? Name the scope and consequence.
Task complete? Stop. Don't invent a next step.

Copy STYLE_GUIDE.md into your agent's project instructions, or use the structured contract. This is a prompt and decision aid, not an enforcement mechanism.

Related MCP server: state-machine

The difference in one example

Known facts: upload failed; cause unknown; file remains selected; retry is supported.

The first response sounds useful. It makes up the diagnosis. The second gives the supported next step.

What you get

• Seven conditional rules with explicit conflict priorities and review requirements.

• An interactive resolver: choose task, evidence and consequence; get the applicable contract.

• Six paired Fable outputs, with exact prompts and visible failures.

• An adversarial review transcript: what Fable challenged, what Astra rejected, what both accepted.

• Browser WebMCP and HTTP MCP tools sharing the same rule engine.

• Zero runtime dependencies. No API key needed to run the recorded demo.

What the experiment actually found

One fresh Fable CLI session per condition, six tasks batched in fixed order. The guided condition received contract 2.0. The current revision is 2.1; it has not been retested in this comparison.

This is an exploratory comparison, not a benchmark victory. One sample per condition, six authored tasks, unblinded Astra review, no independent human evaluation or significance claim. The baseline was already good in several cases. Agreement is not evidence of better performance.

Read the full study, tested contract, and review transcript. Model output identity reported by the CLI: claude-fable-5-1. Astra is the authoring/reviewing agent, not a separately benchmarked condition.

Run locally

Requires Node.js 22 or newer. No install step or API credentials.

git clone https://github.com/Meliwat/clarity-contract.git
cd clarity-contract
npm start

Open http://localhost:4173.

npm test        # build + 20 context combinations, document routes and MCP checks
npm run build   # create the self-contained Worker bundle

PORT=8080 npm start changes the local port. The local server binds to loopback.

Use through MCP

Public Streamable HTTP endpoint:

https://astra-fable-style-guide-meliwat.muhammedeliwat.chatgpt.site/api/mcp

Use a client that supports Streamable HTTP, with that URL and no API key. The implemented protocol version is 2025-03-26. The server is stateless and read-only; it returns JSON responses and does not offer an SSE stream or automatic OAuth/plugin provisioning.

Example (not a model call):

curl -sS https://astra-fable-style-guide-meliwat.muhammedeliwat.chatgpt.site/api/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  --data '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"resolve_style_contract","arguments":{"task":"confirmation","evidence":"known","consequence":"irreversible"}}}'

Tasks: error, confirmation, status, success, interface. Evidence: known, unknown. Consequence: reversible, irreversible. This models a deliberately small set of contexts, not every writing situation.

In browsers supporting document.modelContext, the same three WebMCP tools are registered. Browser resolver and case-selection calls update the visible page; HTTP calls only return data. Unsupported browsers still have a normal working interface.

Reproduce or extend the experiment

The exact baseline and guided prompts are in research/study.json under prompts. Use fresh model sessions and preserve model identity, settings, exact outputs and failed cases. The recorded run used the installed Fable CLI with default settings, no temperature control, and its existing system context. It is not a controlled cross-model study.

The repository serves recorded outputs. It does not call a model or collect your inputs. To propose a new case, see CONTRIBUTING.md.

The best contribution: an example where following the contract still produces a bad answer.

Structure

STYLE_GUIDE.md          Copyable agent instructions
research/contract.json Current contract (2.1)
research/study.json    Prompts, exact outputs, analysis, limitations
research/provenance.json  Visible review and agreement transcript
src/engine.js         Shared deterministic resolver
src/client.js         Browser controls and WebMCP
src/worker.js         HTTP routes and MCP
scripts/              Build, local server and checks

npm run build generates dist/server/index.js, a self-contained Worker with a Fetch handler. It can be deployed to a compatible runtime. A Sites manifest is optional for local builds and is deliberately excluded from this public repository. The live demo is hosted with Sites; no hosting credentials are included.

License and attribution

MIT © 2026 Meliwat. Created with Astra and Fable in response to Tibo's suggestion to have them agree on a style guide and host it with WebMCP. No affiliation or endorsement is implied. Model-authored text and human-facing analysis are disclosed in the study.