Ilm al-Kalām Archive · System Documentation

Methodology

DeepSeek extraction engine + Claude logic engine: how 107+ Islamic theological propositions were extracted from primary sources, structured as formal argument chains, and validated against the seven-layer SCRA architecture.

Computational theology is not AI-generated speculation — it is AI-assisted extraction of arguments that already exist in primary Islamic sources. The human researcher defines the question; the DeepSeek engine finds the sources; Claude structures the logical form; the result is validated against the seven-layer argument chain. No proposition in this archive was invented by the AI. Every proposition was extracted from named primary sources with explicit citation.

I. The Two-Engine System

DeepSeek Extraction Engine
deepseek-chat · Knowledge Mode · Crawl Mode

Extracts raw theological knowledge from primary Islamic sources. The engine has extensive training on classical Islamic texts — Al-Kāfī, Biḥār al-Anwār, Tafsīr al-Mīzān, Fuṣūṣ al-Ḥikam, Nihāyat al-Ḥikma, and hundreds of others. When given a structured extraction prompt, it returns JSON-formatted propositions with source citations, premise arrays, and conclusions.

Two modes: Knowledge (internal training) · Crawl (live source verification) · Budget: ~$0.02 per 100 propositions
Claude Logic Engine
Anthropic Claude · Argument Chain · Validation

Structures extracted propositions into the seven-layer argument chain. Identifies logical dependencies between propositions. Validates internal consistency. Generates cross-school comparative analysis. Maps each proposition to its SCRA layer. Identifies gap propositions needed to complete logical sequences. Writes the final proposition content for the archive.

107 propositions mapped to master-chain.json · 6 critical chain nodes · All 7 layers populated

II. The Extraction Process

Step 1 Topic Definition
Process
  • The researcher defines a theological topic area (e.g., "walāya", "tawassul", "Khawarij")
  • The SCRA seven-layer architecture is consulted to identify which layer(s) this topic belongs to
  • The primary sources most likely to contain relevant propositions are identified (Al-Kāfī, Biḥār al-Anwār, specific Tafsīr works, etc.)
Output

A structured extraction prompt sent to DeepSeek: "Extract all propositions from [source set] on [topic]. Format as JSON with fields: proposition_id, source, premises (array of axioms with text and source), conclusion, scra_layer, school, certainty_grade."

Step 2 DeepSeek Extraction
Process
  • DeepSeek processes the extraction prompt against its training on Islamic primary sources
  • Returns structured JSON: each proposition has a unique ID, named source, array of premises (each with axiom text and its own source), a conclusion, and metadata (SCRA layer, school, certainty grade)
  • The system runs in Knowledge mode (fast, uses training data) or Crawl mode (slower, verifies against live sources)
Output

Raw JSON proposition objects — typically 6–14 propositions per topic extraction. The raw output includes the primary source citations that allow independent verification. DeepSeek's training on Arabic Islamic texts is extensive; citations can be cross-checked against standard hadith databases.

Step 3 Claude Validation
Process
  • Claude receives the raw extracted propositions and the SCRA seven-layer architecture
  • Claude validates: Are the premises actually present in the cited source? Does the conclusion follow from the premises? Does the proposition fit its assigned SCRA layer? Are there cross-school comparisons to be added?
  • Claude identifies logical gaps: propositions that are needed to complete argument chains but were not extracted
Output

Validated propositions with corrections, additions, and cross-references. Gap propositions identified for follow-up extraction. The argument chain map (master-chain.json) updated with new proposition nodes and their layer assignments.

Step 4 Chain Mapping
Process
  • Each validated proposition is mapped to its position in the master-chain.json: which SCRA layer, which critical chain node, what upstream propositions does it depend on, what downstream propositions does it enable
  • Cross-layer dependencies are identified: a Layer V walāya proposition may depend on Layer I ontological propositions and enable Layer VII institutional propositions
  • The complete chain is verified: does each layer have adequate proposition coverage? Are there weak links?
Output

A complete seven-layer argument chain with 107+ propositions, 6 critical chain nodes, and all cross-layer dependencies mapped. The master-chain.json file serves as the machine-readable version of the argument chain displayed on the Argument Chain page.

III. Proposition Structure

Every proposition in the archive has the same structure — no exceptions. If a proposition cannot be structured in this form, it is not a proposition; it is an assertion.

EXAMPLE Imami Layer V
Source: [Primary Source Name] · [Author/Narrator] · [Additional Cross-References]
Premises (2–4 per proposition)
  • Premise 1: A factual or textual claim from the cited source — what the source actually says, with Quranic verse or hadith reference where applicable
  • Premise 2: A second claim that logically supports the conclusion — may be from the same source or a related source
  • Premise 3: Optional additional claim or qualification — cross-school comparison, boundary condition, or contextualizing fact
Conclusion

The conclusion that follows from the premises above — stated precisely, without adding claims not present in the premises. The conclusion may then serve as a premise in a downstream proposition. The argument chain is built by these logical dependencies.

IV. Certainty Grades

Grade A
Mutawātir / Established Consensus

Established by mutawātir (mass-transmitted) hadith or Quranic text with clear interpretation consensus. Cross-school agreement. No serious scholarly dispute.

Grade B
Āḥād Ṣaḥīḥ / Strong Scholarly Support

Strong hadith chains (ṣaḥīḥ or ḥasan). Majority scholarly position within the relevant school. Minor scholarly disputes exist but the majority position is clear.

Grade C
Inferred / Analytical

Derived by analysis from established sources. The proposition is the logical conclusion of the SCRA's argument chain applied to primary materials. Clearly labeled as analytical, not textually explicit.

V. Budget and Scale

System Economics

The computational theology system is budget-efficient: the entire 107-proposition database was extracted for approximately $0.031 from a $2.00 DeepSeek API budget — 98.4% of the budget remains. At this rate, 10,000 propositions would cost approximately $3.00. The bottleneck is not API cost but validation quality: Claude's validation pass ensures the extracted propositions are logically coherent and accurately cited. The system scales: additional topic areas, cross-school comparisons, and deeper extraction of existing topics can be added at minimal cost. The Kalām Archive is designed to grow.

VI. What This System Does Not Do