EM-0037 — Human-readable cases and agent research kit

Source content

EM-0037 — Human-readable cases and agent research kit

Objective

Polish the human projection of Cases 002–004 without changing accepted evidence meaning, and make

the same bounded research method recoverable by a cold-start coding agent through static files,

local REST, read-only MCP, and CLI.

Implementation

1. Add projection-only editorial framing for false independence, missing comparison class, and

scope inflation: familiar claim, why it matters, finding, practical reading, and a three-step

evidence map before technical ledgers.

2. Normalize every Case 002 human ledger member to a typed identity, show count-card ledgers first,

and place additional audit groups behind one native disclosure.

3. Collapse repeated human-display source spans by exact accepted span ID and digest while

retaining all occurrences and relations in JSON, Markdown, calculations, and machine ledgers.

4. Publish one public research protocol, per-case briefs, a draft proposal template, fail-closed

validation, local CLI commands, and read-only MCP/REST discovery.

5. State the queue boundary precisely: EM-0037 does not submit or store bundles; EM-0038 separately

governs a future authenticated coordination queue with zero automatic evidential credit.

Preservation boundary

All accepted dossier, review-receipt, and selection-manifest inputs remain byte-identical. Case 001

is unchanged except for shared discovery of its research brief. No accepted count, evaluation,

relation, uncertainty, license, calculation, or review conclusion changes.

Validation

• Recompute every case count from accepted typed members and reject missing human ledger identity.
• Prove repeated-span collapse affects HTML display only.
• Exercise cold-start static/REST/MCP/CLI parity and draft/ready validation behavior.
• Reject traversal, local/private URLs, secret-shaped data, oversized bundles, unsupported fields,

duplicate identities, missing source/span closure, and admission-like top-level status.

• Run Ruff, focused tests, full deterministic make check, desktop/mobile no-JavaScript browser QA,

and fresh-clone exact-head independent review before protected merge.

Deployment boundary

This task does not deploy Pages or a hosted API/MCP, activate credentials, accept a proposal, or

implement the EM-0038 queue. Live-route and provider read-back require a separately authorized

accepted deployment.

Author validation

• Focused Case 001–004, interface, and agent-kit suite: 84 tests passed.
• New research-kit suite: 27 tests passed, including adversarial bundles and static/REST/MCP/CLI

parity.

• Full make check PYTHON=.venv/bin/python: 134 tests passed; public audit clean; independent

deterministic rebuild identical; tracked source state unchanged.

• Browser QA at 1280×720 and 390×844: one H1, zero scripts, no horizontal page overflow, compact

case hierarchy, collapsed technical disclosures, and visible 3px focus outline. Mouse/click

activation of native disclosures passed. The browser driver focused a native summary but did not

dispatch Enter, so keystroke activation remains an explicit post-merge/live-readback item rather

than an inferred pass.