Marc Manley

Musnad Middle Ground — Phase 5 Progress and the Agentic Next Step

Phase 5 began by taking the first real Musnad records through the full human-review and scholarly-verification workflow.

What We Completed

Three publication-ready canonical records are now complete:

  • mm-h-0001 — Friday ghusl / walking to Jumuʿah reward
  • mm-h-0002 — Ibn ʿAbbās on preaching with restraint
  • mm-h-0003 — renewing faith in the heart

Along the way, the pilot records exposed several useful design decisions:

  • canonical Arabic and English may be reviewer-selected rather than mechanically copied from the original Middle Ground usage;
  • narrator spelling is preserved exactly as supplied during verification;
  • narration_type was removed because it added technical classification complexity without serving the purpose of Musnad Middle Ground;
  • fallback slugs are now temporary only and must be explicitly finalized before publication readiness;
  • the immutable mm-h-* ID remains the canonical identity;
  • historical as_published evidence stays separate from later verified data;
  • verified transmission details are not back-filled into transmission_as_published when they were not actually present in the original Middle Ground source.

The updated Musnad validation and test suite passed after these changes.

Shared Review Workflow

A shared Google Drive workspace was established for moving large review artifacts between the NucBox and MacBook Pro:

My Drive/_Musnad Middle Ground/
├── review/
├── reports/
└── handoff/

From WSL:

/mnt/g/My Drive/_Musnad Middle Ground

This is now the review/handoff workspace only. The Git repository remains the authority for canonical Musnad data, scripts, schemas, and source files.

Batch Review

The historical extraction produced 48 staged candidates. After earlier adjudications, 46 remained pending.

Claude Code generated batch review artifacts, but the first versions were too verbose and exposed an important usability problem: the review process should not force the human reviewer to parse extractor internals or long forensic dossiers for every candidate.

The review interface therefore needs to compress repository complexity into a short list of human decisions such as:

LINK → mm-h-XXXX
PROMOTE
DISMISS
INVESTIGATE

Detailed source evidence should be available only when needed.

What Comes Next

The next milestone is to build an agent-ready review layer that can later be operated by the existing Middle Ground local-agent stack:

Telegram
    ↓
Hermes on NucBox
    ↓
Qwen via LM Studio
    ↓
deterministic Musnad tooling
    ↓
Google Drive review / handoff
    ↓
human approval
    ↓
validation / Git

The immediate goal is not Telegram integration. It is to build three deterministic local commands:

musnad:prepare-review
musnad:process-review
musnad:apply-review

These commands should:

  • group related candidate detections;
  • compare them against canonical Musnad records;
  • generate compact human decision surfaces;
  • read approved handoff files;
  • translate decisions into dry-run Musnad CLI actions;
  • apply only explicitly approved actions;
  • run validation and tests;
  • generate reports;
  • prepare clean Git work.

Hermes and Qwen will handle the repetitive operational work cheaply and locally. Human scholarly judgment will remain required for ambiguous hadith identity, variant relationships, citation conflicts, canonical Arabic/English selection, scholarly verification, and final publication approval.

The next phase is therefore to formalize and implement this deterministic review-layer contract so Hermes/Qwen can operate it without reinventing the workflow on every run.