CoMMA Research Miner MCP v0.1.2-test.3

Human-readable research-console test build layered over the CoMMA Research Miner v0.5.3-compatible retrieval/scoring engine.

Purpose of test.3

The first MCP experiments proved that remote CoMMA retrieval and scoring work, but also showed that a researcher should not have to operate a protocol debugger or read raw JSON for ordinary work. Test.3 keeps the low-level MCP tools while adding a human-facing research workstation inspired by the Chrome extension.

The governing design principle is:

Simple by default, precise when desired, and transparent at every step.

Major features

Simple Search

Paste one combined research block containing search terms, friendly weighted criteria, formal CoMMA rules, known controls and notes. Preview the interpretation before retrieval.

Advanced Search

A click-over explicit interface with separate boxes for:

• CoMMA search terms

• formal scoring criteria

• known corpus / controls

• retrieval mode

• page size

• max hits

• top-N results

A Simple Search preview can be opened directly in Advanced Search without retyping the research plan.

Human-readable results

Results are presented manuscript-first rather than as raw JSON:

• manuscript title and QID

• passage count

• score range

• contextual-fit badges

• scoring-family chips

• highlighted hit text

• CoMMA and Biblissima links

• known-control badges

• source-control review fields

Known controls vs external candidates

Known project corpora still score normally for calibration, but are separated from novelty-eligible external candidates. Built-in gold control:

• Q262962 / BSB Clm 464 / Philippus Callimachus Experiens

Priority queue

The interface highlights the strongest external candidates first under “What should I examine next?”

Inline human review

Each passage can be marked with:

• review state

• research outcome

• novelty class

• evidentiary confidence

• notes

Reviews are saved in the active run and included in exports / ChatGPT packets.

Rescore cached results

Adjust formal criteria and rescore the already retrieved passage corpus without repeating the CoMMA network search. A new child run is created and linked to the parent run.

Prepare for ChatGPT

One click creates a compact research packet optimized for an MCP-connected AI client. The MCP exposes:

• list_recent_runs

• get_research_packet

• prepare_for_chatgpt

The goal is eventual handoff without downloading and re-uploading files.

Exports

One-click browser exports:

• human-readable HTML research report

• CSV spreadsheet

• complete ZIP research package

The ZIP includes original input, normalized plan, parsed input, run manifest, full results, candidate/control CSVs, human report, research notebook and ChatGPT research packet.

Recent runs

The browser and MCP expose active recent runs so a researcher or agent can reopen the latest run without manually preserving a cache ID.

Deployment in the existing GitHub repository

Upload this entire folder intact at repository root:

comma_mcp_v012_test3/
  package.json
  Dockerfile
  server.mjs
  mcp.mjs
  public/
  test/
  ...

Create a NEW Render Web Service pointing to the same GitHub repository and set:

Root Directory: comma_mcp_v012_test3
Runtime: Docker

Do not point the existing working Groovy service at this folder.

Set ALLOWED_HOSTS to the new Render hostname after Render assigns it.

Example:

ALLOWED_HOSTS=comma-mcp-test3-xxxx.onrender.com

Optional:

PUBLIC_BASE_URL=https://comma-mcp-test3-xxxx.onrender.com
CACHE_TTL_MINUTES=120

Browser workflow

1. Open /.

2. Use Simple Search or Advanced Search.

3. Preview / validate.

4. Run research.

5. Review external candidates and known controls.

6. Optionally adjust criteria and Rescore cached results.

7. Record source-control judgments.

8. Click Prepare for ChatGPT.

9. Download a report, spreadsheet or complete research package as desired.

MCP endpoint

https://YOUR-HOST.onrender.com/mcp

Health endpoint

GET /health

Expected version:

{
  "ok": true,
  "service": "comma-research-miner-mcp",
  "version": "0.1.2-test.3",
  "engine": "0.5.3-compatible-test"
}

Research boundary

CoMMA retrieval finds passages. Contextual scoring ranks passages. Similarity and clustering can generate source-control candidates. None of these alone establishes textual independence, historical truth, manuscript novelty, authorship or scholarly originality.