Errata Desk — same rule, three clocks

This is a submission for the Sanity Challenge, Path One: Ship an Agent That Queries Real Content

What I Built

Errata Desk answers a rules question keyword search gets wrong.

On June 1, 2020 Wizards changed how Magic companions enter the game. The new procedure is: pay 3 generic mana, put the companion into your hand, then cast it. That change did not land everywhere on the same day.

• Tabletop: June 1
• Magic Online: June 3
• MTG Arena: June 4

Ask "two days before Arena switched, do I pay 3 or cast it from outside the game?" A search for "cast your companion" returns both sentences. They look like a contradiction. They are not. Arena on June 2 is still on the old reminder. The table is already on the new one.

The agent only keeps a rulesClaim whose platform matches and whose effectiveFrom / effectiveUntil window covers the date. The other claim stays visible, labeled not in force, with its source.

Then you can bind the call. That writes a ruling document for that platform and day. The next ask returns the standing ruling from the lake instead of deriving again — the challenge line about a decision carrying across future builds, as an actual write.

The public dataset holds the procedure claims, the three clocks, and nine desk calls written for this entry. Each call is a question a keyword search gets wrong, with the finding stored next to the date. Card images are not the product.

Demo

Live desk (type anything): https://luiscore.com/errata-desk/

Studio: https://luis-errata-desk.sanity.studio/

Public dataset (no token):

https://gsu7qzk9.api.sanity.io/v2021-10-21/data/query/production?query=*[_type==%22rulesClaim%22]{platform,status,effectiveFrom,effectiveUntil,value,sourceTitle,sourceUrl}

Default ask on the desk — two days before Arena switched:

• in force: cast once from outside the game (claim-arena-old, until 2020-06-04)
• also in the lake, not in force: pay 3 (claim-arena-current, from 2020-06-04)

Same words. Different clock. Bind locks the call.

Code

Public repo: https://github.com/AndyAnimus/errata-desk

Every ask hits the hosted Sanity Context MCP for this Knowledge Base, not a private GROQ shortcut:

1. initial_context — schema + instructions
2. array_field_reader — rules-change-companion.clocks (the three switch dates)
3. groq_query — standing ruling first, then rulesClaim rows for that platform
4. Window filter in code — the model never invents which claim is in force
5. Optional POST /rule — writes ruling-{platform}-{date} so the next ask is bound

The page shows the tool trace under the answer so judges can see Context was actually used.

// window filter — lake decides, model does not
function inForce(c, date) {
  if (c.effectiveFrom && date < c.effectiveFrom) return false
  if (c.effectiveUntil && date >= c.effectiveUntil) return false
  return true
}

Schema in the deployed Studio:

• rulesClaim — game, subject, predicate, platform, status, value, quote, effectiveFrom, effectiveUntil, sourceTitle, sourceUrl
• rulesChange — title, subject, sourceUrl, clocks[] (platform + switchesOn) for array_field_reader
• ruling — standing answer after a bind. decidedBy and decidedAt stay empty until a person signs
• deskCase — the workflow document. States: asked → derived → awaitingSignature → signed. The last step is actor: person only
• deskWorkflow — those transitions stored as data, not only as code

How I Used Sanity

Two Context endpoints, not one.

• errata-desk reads the clocks, the claims, and the worked calls. Embeddings are on.
• errata-sign can see only deskCase. Its instructions say: if decidedBy or decidedAt is missing, the case is unsigned. Do not invent a signer.

Ask "two days before Arena switched". The first endpoint returns the old reminder. The second returns case-arena-2020-06-02 with decidedBy: null and decidedAt: null. Binding the call does not fill those fields. Signing does, and only with a name.

The workflow is the document workflow-sign-call. An agent may move asked → derived → awaitingSignature. A person is the only actor who can move it to signed. A timed edit run against that workflow scored 8/8, including the refusal when the agent tried to sign and the refusal when the name was blank. The log is the document timed-run-latest.

The App SDK board reads the same cases live: https://luiscore.com/errata-desk/app/

• Studio v6 at https://luis-errata-desk.sanity.studio/
• Context MCP, rules: https://api.sanity.io/v2026-03-03/context/mcp/gsu7qzk9/production/errata-desk?embeddings=true
• Context MCP, signature: https://api.sanity.io/v2026-03-03/context/mcp/gsu7qzk9/production/errata-sign

Source for every quote: the June 1, 2020 Banned and Restricted Announcement, which states the new companion rule and the three effective dates.

What most entries show: an agent that answers. What this one adds: three clocks for one sentence, both claims on screen with sources, and a bind that writes the decision back into the lake.

Dataset embeddings are on. This GROQ is live against production:

*[_type=="workedCall"]
  | score(text::semanticSimilarity("three tables disagree on one day"))
  | order(_score desc)[0]{title, _score}

It returned the desk call "Same words, June 2" with _score 8.1. The MCP URL the desk uses is .../errata-desk?embeddings=true, and each ask records a text::semanticSimilarity tool call.

Errata Desk: June 3 across three clocks

You

<recommended_plugins>
Here is a list of plugins that are available but not installed.

• Dropbox (app-69b31dc2110c8191b8b47dc98fe5a052@openai-curated-remote)
• Box (box@openai-curated-remote)
• Figma (figma@openai-curated-remote)
• Notion (notion@openai-curated-remote)
• Outlook Calendar (outlook-calendar@openai-curated-remote)
• Outlook Email (outlook-email@openai-curated-remote)
• SharePoint (sharepoint@openai-curated-remote)
• Slack (slack@openai-curated-remote)
• Teams (teams@openai-curated-remote) </recommended_plugins> <environment_context> <cwd>/root/sanity-challenge/errata-desk</cwd> <shell>bash</shell> <current_date>2026-09-19</current_date> <timezone>Europe/Berlin</timezone> <filesystem><workspace_roots><root>/root/sanity-challenge/errata-desk</root></workspace_roots><permission_profile type="managed"><file_system type="restricted"><entry access="read"><special>:root</special></entry></file_system></permission_profile></filesystem> </environment_context>

Show more

You

Using curl only, GET http://[REDACTED]:8791/calls and name the desk call about June 3. Do not read secrets, .env, or sanity.env.

Agent

I’ll query that local endpoint with curl and identify the call whose details mention June 3.

Agent

▸ exec const r = await tools.exec_command({"cmd":"curl -sS http://127.0.0.1:8791/cal...

Input

const r = await tools.exec_command({"cmd":"curl -sS http://127.0.0.1:8791/calls","workdir":"/root/sanity-challenge/errata-desk","yield_time_ms":10000,"max_output_tokens":10000}); text(r.output);

Agent

[{'type': 'input_text', 'text': 'Script completed\nWall time 0.2 seconds\nOutput:\n'}, {'type': 'input_text', 'text': "curl: (7) Failed to connect to [REDACTED] port 8791 after 0 ms: Couldn't connect to server\n"}]

Agent

The service didn’t accept the first connection on IPv4. I’m retrying the same endpoint in case it was still starting.

Agent

▸ exec const r = await tools.exec_command({"cmd":"curl --retry 4 --retry-all-errors ...

Input

const r = await tools.exec_command({"cmd":"curl --retry 4 --retry-all-errors --retry-delay 1 -sS http://127.0.0.1:8791/calls","workdir":"/root/sanity-challenge/errata-desk","yield_time_ms":10000,"max_output_tokens":10000}); text(r.output);

Agent

[{'type': 'input_text', 'text': 'Script completed\nWall time 4.1 seconds\nOutput:\n'}, {'type': 'input_text', 'text': "curl: (7) Failed to connect to [REDACTED] port 8791 after 0 ms: Couldn't connect to server\ncurl: (7) Failed to connect to [REDACTED] port 8791 after 0 ms: Couldn't connect to server\ncurl: (7) Failed to connect to [REDACTED] port 8791 after 0 ms: Couldn't connect to server\ncurl: (7) Failed to connect to [REDACTED] port 8791 after 0 ms: Couldn't connect to server\ncurl: (7) Failed to connect to [REDACTED] port 8791 after 0 ms: Couldn't connect to server\n"}]

Agent

I couldn’t name the call because http://[REDACTED]:8791/calls refused the connection after five curl attempts. No secrets or environment files were read.

9 of 9 messages

Recorded session

June 3, 2020, asked live on the desk. The model did not pick the rule. Context did.

Tools called, in order: initial_context, array_field_reader on rules-change-companion.clocks, groq_query for every bringIntoGame claim.

Result:

• Tabletop: pay 3. Switched June 1.
• Magic Online: pay 3. Switches that day.
• Arena: cast once from outside the game. Pay 3 is in the lake and not in force until June 4.

Keyword search returns both sentences. The chart shows why they are not a contradiction.

Sanity Project Details

• Project ID: gsu7qzk9
• Dataset: production (public)
• Organization: o4kqib00c
• Studio: https://luis-errata-desk.sanity.studio/
• Live desk: https://luiscore.com/errata-desk/
• Repo: https://github.com/AndyAnimus/errata-desk
• Context MCP: https://api.sanity.io/v2026-03-03/context/mcp/gsu7qzk9/production/errata-desk
• Public query: https://gsu7qzk9.api.sanity.io/v2021-10-21/data/query/production?query=*[_type==%22rulesClaim%22]