LLM Training and MORK Teaching Pack (MTP) Implementation Status

Unit: llm-training-mtp State: Accepted - fully implemented and validated Sketch: llm-training.md Governing ADRs: ADR-A25, ADR-A44

What was built

The MORK Teaching Pack (MTP) is a versioned, hash-pinned, machine-generated curriculum that teaches language models to emit valid MCN (MORK Compact Notation) without requiring them to memorize Mork.ttl in full. The implementation consists of:

L0 Kernel Generator (tools/mork/src/mtp/facts.py, doctrine.py, render.py)

• Automatically extracts 258 terms and 895 axioms from Mork.ttl
• Derives box families, uncertainty codes, inverse pairs, and generative mandatory parts
• Anchors curated doctrine to ontology IRIs with logical fingerprinting
• Enforces axiom ownership: every GCI, disjointness, and completeness axiom must be claimed
• Generates l0.txt kernel with 7 invariants, decision ladder, 14 hot codes, and uncertainty escape

L2/L3 Lenses and Cassettes Pipeline

• S1–S13 pipeline implements the full feed from ontology through cassette fidelity, minimisation, mutation, cassette selection, and routing
• Mutation-derived ground truth: applies mutation operators to verified examples, records what lint catches, identifies silent mistakes
• 9 lenses (L-EGR, L-GEN, L-IND, L-INT, L-REP, L-TAX, L-TBX, L-UNC, L-?) with 300–800 tokens each
• 30+ cassettes (C-LOAN, C-UNCERTAIN, etc.) with 150–400 tokens and isomorphic MCN verification
• Diagnostics and repair-loop table derived from mutation matrix
• Budget enforcement: hard gates on token counts, partition coverage, cassette fidelity

Data and Contracts

• Curated inputs: doctrine.yaml (anchored doctrine), partition.yaml (term→lens assignment), lenses/L-*.yaml, cassettes/C-*.yaml, routing.yaml
• Generated artifacts: pins.lock.json (logical fingerprints), L0/L2/L3 rendered text and JSON, coverage report, mutation matrix
• MCN codebook (mcn_codebook.py) with stable reverse lookups

Verification and CI

• MCN decoder with 337 round-trip tests including loan_mapping.ttl and UncertainMappings.ttl isomorphism proofs
• Structural tests: test_mtp.py with 9 passing tests
• CI gate: build:mtp generates output, check:mtp validates pins, cassette fidelity, mutation verdicts, and stale output
• Deterministic hashing: same inputs always produce byte-identical pins and manifest

Evidence

• Build: mise run build:mtp produces 20 output files under ontology/mork/mtp/out/ from curated inputs
• Check: mise run check:mtp validates logical consistency, pin drift detection, cassette isomorphism, and mutation fidelity
• Tests: python -m pytest tools/mork/src -q yields 346 passed, 8 skipped in 2.29s
• Facts report: python -m mtp.cli facts-report confirms 258 terms and 895 axioms extracted
• Repository integration: Follows ADR-A77 topology; uses mise for orchestration, Python project under tools/mork/ with conventional locks

Design sketches implemented

1. llm-training.md — the conceptual thesis and L0-L5 tier model
2. MTP-L0 Generator Sketch — facts.py, doctrine pinning, axiom ownership gates
3. MTP L2 Generator Sketch.md — S1-S13 pipeline, mutation-derived ground truth, cassette fidelity

Acceptance

All three design sketches have been fully implemented and validated. The MTP generator is production-ready for the LLM training backend integration work (Phase 7), which remains deferred pending backend interface specification.

Deferred (out of scope)

• Model-quality evaluation (evalset.py, eval-report CI, eval leakage/thin gates) — requires LLM backend invocation interface
• Backend selection and sampling policy — owned by separate backend-interface effort
• MCN encoder (mcnio.encode()) — only needed for future lint repair automation

Next steps

• Phase 7: LLM backend integration, model-quality metrics, eval harness
• Maintain pin determinism across future Mork.ttl edits
• Expand cassette library and lens coverage as production deployments report new patterns