# Em 0037

- Object ID: `em:task:sha256:b86991732f8403ee058a6ff06d85b34810eecc16207eb26c9220d595363ec610`
- Kind: `task`
- Repository path: [`tasks/contracts/EM-0037.json`](https://github.com/yoheinakajima/epistemedia/blob/f92846570180dfa4511263f8ba98ecd18f7772c9/tasks/contracts/EM-0037.json)
- Content digest: `06608a8e52a365cb3a87d16196acb2ed998aa359d88ede76d7c1fc1b66f39a1e`

**Also filed under:** [Agent Operations](https://epistemedia.org/topics/agent-operations/), [Research Program](https://epistemedia.org/topics/research-program/)

## Source content

{
  "id": "EM-0037",
  "title": "Polish Cases 002-004 and publish the agent research kit",
  "status": "ready",
  "change_class": "ordinary-implementation",
  "objective": "Make Cases 002, 003, and 004 as legible and narratively useful as Case 001 while preserving their accepted evidence meaning, and publish a cold-start agent research kit that lets an unfamiliar coding agent discover a case, inspect the protocol, run a comparable bounded investigation, validate a portable proposal bundle, and understand that submission is queued for review rather than admitted as knowledge.",
  "depends_on": [
    "EM-0036"
  ],
  "authority": {
    "allowed_paths": [
      "README.md",
      "src/**",
      "tests/**",
      "docs/**",
      "ops/**",
      "runs/**"
    ],
    "forbidden_paths": [
      "constitution/**",
      "policies/**",
      "schemas/**",
      "governance/events/**",
      ".github/**",
      "tasks/contracts/**",
      "research/**",
      "catalog/dossiers/**"
    ]
  },
  "required_evaluation": [
    "byte and semantic preservation of all four accepted dossier, review-receipt, and manifest inputs",
    "human-first narrative hierarchy and progressive-disclosure review for Cases 002, 003, and 004 without hiding unresolved, no-credit, license, dependence, calculation, or review limitations",
    "deterministic collapse of repeated human-display source spans by exact accepted identity while retaining every occurrence and relation in Markdown, JSON, ledgers, and machine traces",
    "complete displayed-count reconciliation with typed member labels and no unknown object type, key, or ID on human ledger summaries",
    "cold-start agent discovery from homepage, llms.txt, Markdown, JSON, local REST, MCP, and CLI with no inherited repository or conversational context",
    "portable research-proposal bundle generation and fail-closed structural validation with explicit source, edition, span, warrant, dependence, uncertainty, negative-result, runtime, and license fields",
    "agent instructions that distinguish exploration, research execution, proposal preparation, queue submission, independent review, and accepted Git history",
    "prompt-injection, path-traversal, private-context, secret-shaped-data, oversized-input, unsupported-format, duplicate-identity, and admission-language adversarial tests",
    "no-JavaScript fallback, native keyboard accessibility, desktop and mobile visual review, link integrity, content-type, and canonical-URL checks",
    "whole-release identity parity, public disclosure noninterference audit, full deterministic make check, exact path-scope comparison, and clean tracked state",
    "fresh-clone independent exact-head review before protected merge",
    "provider and route read-back after a separately authorized accepted Pages deployment"
  ],
  "acceptance": [
    "Cases 002, 003, and 004 open with a concise claim, why-it-matters sentence, bounded finding, practical reading, and visual evidence map before technical ledgers",
    "long ledgers and repeated spans are summarized into readable groups by default while exact accepted objects, occurrences, relations, calculations, and receipts remain reachable without JavaScript",
    "no human-facing count member displays an unresolved placeholder for object type, key, or ID; any genuinely missing identity is labeled as an evidence gap rather than rendered as unknown metadata chrome",
    "every case offers an agent-ready action that leads to one stable public research protocol and a case-seeded portable brief",
    "a cold-start coding agent can recover the research question, comparison target, source and span requirements, dependence rules, negative-result policy, output format, validation command, and submission boundary from public files alone",
    "the local CLI and read-only MCP surface can prepare and validate a deterministic proposal bundle but cannot submit, admit, merge, publish, or mutate accepted state",
    "public copy identifies the intended MCP review queue as a separate future authority surface and accurately reports whether hosted submission is available",
    "Case 001 remains semantically and visually unchanged except for shared discovery links that do not alter its accepted evidence projection",
    "make check passes without changing accepted source state"
  ],
  "limitations": [
    "This task changes application presentation and publishes a non-admitting research kit; it does not change accepted dossier evidence, policy semantics, review conclusions, or counts.",
    "The proposal bundle is untrusted intake and carries zero evidential or independence credit until its sources, spans, derivations, and lineage are independently reviewed.",
    "The public API and MCP remain read-only in this task. A hosted write-capable submission queue requires EM-0038 governance, least-privilege credentials, provider deployment, and external read-back.",
    "No research provider call, model run, automatic admission, pull-request creation, merge, publication, deployment, DNS change, credential use, account provisioning, or spend is authorized by this task."
  ]
}