Milcah — the Coherence Engine

Milcah ingests a framework, argument, or worldview and recursively pressure-tests it for coherence — exposing assumptions, equalising explanatory burden, locating fractures, and sharpening uncertainty — using multi-LLM analysis.

Its purpose is not to prove truth. It is to help people reason more honestly and comprehensively: every framework pays the same explanatory cost, and "forced certainty is forbidden."

The guiding question for anything it examines:

What would have to be true for this to remain coherent?

Status

v0.2 — core engine built. The philosophy and requirements are set, and the main engine path is real: ingestion (FR1) normalises an input into a segmented framework, reasoning extraction (FR2) pulls typed reasoning units out of it (with single-, per-segment, and multi-LLM modes, the last reconciling by text or by meaning), ontology construction (FR3) builds the worldview tree with placement states, the recursive reasoner (FR4) pressure-tests each node with the five questions, counter-framework research (FR5) generates the strongest objections + competing frameworks, fallacy analysis (FR6) locates inference defects, coherence metrics (FR7/FR9) score the result structurally, FR10 persistence/history stores snapshots, and the round controller (FR11) drives reason + challenge to a termination condition. milcah orchestrate ties those steps into ADR-001 role-based orchestration, and milcah.specialist.run_specialist is the provider entrypoint for the Tirzah↔Milcah specialist contract. Extraction runs on a deterministic rule-based baseline by default, or — for higher-quality typing — on a local LLM executed through Hoglah (--extractor hoglah). See docs/philosophy.md, docs/requirements.md, and the initial docs/architecture.md. Milcah's own process is described in the Cairn format in docs/process.cairn.md. A point-in-time functional + code review of 0.4.0 is in docs/review-2026-08-08.md (findings not yet actioned; note F1/H1 — on the deterministic default path global_coherence is 1.0 for every input).

milcah extract framework.md                              # deterministic baseline
milcah extract framework.md --json                       # full Framework + units JSON
milcah extract framework.md --extractor hoglah --model gemma4:latest
#   ^ LLM extraction via Hoglah→Ollama (needs a `hoglah run --real` daemon;
#     install with `pip install -e ".[hoglah]"`)
milcah extract framework.md --extractor hoglah --per-segment
#   ^ one extraction job per segment, merged with segment provenance —
#     keeps long frameworks within the model's context window
milcah extract framework.md --extractor hoglah \
  --models gemma4:latest,gemma4:e2b,gemma2:2b
#   ^ multi-LLM: extract with each model, then reconcile by agreement —
#     each unit records how many models agreed and how they voted on its type
milcah extract framework.md --extractor hoglah \
  --models gemma4:latest,gemma4:e2b,gemma2:2b --reconcile semantic
#   ^ semantic reconciliation: merge units by MEANING (embeddings) so phrasing
#     variants count as agreement, not separate units
milcah ontology framework.md                             # FR3: worldview tree
#   ^ build the ontology tree from the units — foundations at the root, with an
#     ontological placement state per node (resolved … contradictory)
milcah ontology framework.md --placement llm --model gemma4:latest
#   ^ a model reasons about placement (incl. contradictions) via Hoglah, instead
#     of the deterministic structural scaffold
milcah reason framework.md --model gemma4:latest --max-depth 1 --max-nodes 10
#   ^ FR4: recursively pressure-test each ontology node with the five questions
#     (what supports / must be true / implies / assumes / explains), bounded by
#     a depth threshold + node budget
milcah challenge framework.md --model gemma4:latest
#   ^ FR5: the strongest objections + competing counter-frameworks, applying the
#     same scrutiny to every framework (burden symmetry)
milcah challenge framework.md --model gemma4:latest --web-search \
  --web-search-url http://localhost:8080 --web-allow-private-search-endpoint
#   ^ ground FR5 in transient, cited web evidence through SearxNG. Recursive
#     reasoning derives a separate bounded query from each node. Web text is
#     marked untrusted and never becomes established knowledge automatically.
milcah rounds framework.md --model gemma4:latest --max-rounds 3
#   ^ FR11: drive reason + challenge in rounds, stopping on convergence,
#     repeated objections, the round threshold, or the node budget
milcah fallacy framework.md --model gemma4:latest --max-steps 20
#   ^ FR6: locate fallacies at reasoning steps, judging inference form rather
#     than institutional acceptance or popularity
milcah orchestrate framework.md --auto-models
#   ^ ADR-001: role-based Proposer/Challenger/Fallacy/Synthesis orchestration;
#     model diversity is provenance and bias reduction, never a confidence score
milcah specialist "Is this framework coherent?" --context-file framework.md --json
#   ^ the public Tirzah↔Milcah coherence_check contract, executable from the CLI
milcah metrics framework.md                              # FR7/FR9: coherence metrics
#   ^ structural explanatory-debt + coherence scores — deliberately excluding
#     popularity, confidence, institutional acceptance, and model-agreement
milcah history framework.md                              # FR10: saved snapshot trend

What it does (requirements, in brief)

• Ingest frameworks from books, documents, hypotheses, argument trees, conversations, and web research.
• Extract the reasoning: claims, observations, assumptions, commitments, bridges, enthymemes, dependencies, conclusions.
• Build the (often implicit) worldview ontology and track each concept's placement state (resolved … contradictory).
• Recurse on every node — what supports this? what must be true? what does it imply? what assumptions exist? what explains them? — with no fixed depth.
• Challenge symmetrically — identical pressure for every framework, no exemptions; generate counter-frameworks and the strongest objections.
• Locate fallacies and track explanatory debt + coherence metrics that deliberately exclude popularity, confidence, and institutional acceptance.
• Preserve unresolved states — uncertainty is a valid outcome, made more precise, never collapsed.

Place in the family

Milcah is the orchestrator and judge of coherence; it stands on its siblings rather than reimplementing them:

Develop

git clone https://github.com/gellsmore-svg/Milcah
cd Milcah
python -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest

Feedback

This is early. Open a feedback issue, and see CONTRIBUTING.md. Security: SECURITY.md.

Knowledge bundle

A machine- and human-readable knowledge map of Milcah's concepts and modules is published as an Open Knowledge Format bundle under okf/ — markdown with YAML frontmatter, linked into a concept graph.

License

Apache License 2.0.

Snapshot viewer & tracing

• milcah serve (needs pip install milcah[web], default 127.0.0.1:8791) — a read-only browser over saved snapshots: the framework index, each framework's coherence trend over time as SVG sparklines (green = improving, red = degrading), and full snapshot detail. Same store flags as history (--store json|mongo).
• Family trace spine: set MILCAH_GALEED_ENABLED=1 (with the galeed extra) and orchestration runs + snapshot saves emit onto the shared Galeed stream, viewable in Mizpah or galeed trace.