Research Tools & Process¶
This page describes the automated research system and investigative methodology that produced the 19 studies in this series.
Investigative Stance¶
Each study is produced by an agent that functions as an investigator, not an advocate. This distinction governs every step of the process:
- Gather evidence from all sides. If a passage is cited by historicists, examine it honestly. If a passage is cited by preterists, futurists, or idealists, examine it honestly.
- Do not assume a conclusion before examining the evidence. The conclusion emerges FROM the evidence, not the reverse.
- State what the text says, not opinions about it. The agent does not use editorial characterizations like "genuine tension," "strongest argument," or "non-intuitive reading." It states what each passage says and what each interpretive position infers from it.
- Never use language like "irrefutable," "obviously," or "clearly proves." Use "the text states," "this is consistent with."
How the Studies Were Produced¶
Each study was generated by a multi-agent pipeline, a Claude Code skill that answers Bible questions through tool-driven research. The pipeline ensures that:
- Scope comes from tools, not training knowledge. The AI does not decide which verses are relevant based on what it was trained on. Instead, tools search topical dictionaries, concordances, and semantic indexes to discover what Scripture says about the topic.
- Research and analysis are separated. The agent that gathers data is not the same agent that draws conclusions. This prevents confirmation bias.
- Every claim is traceable. Raw tool output is preserved in each study's
raw-data/folder, so every finding can be verified against its source.
The Three-Agent Pipeline¶
Phase 1: Scoping Agent
| Discovers topics, verses, Strong's numbers, related studies
| Writes PROMPT.md (the research brief)
Phase 2: Research Agent
| Reads PROMPT.md
| Retrieves all verse text, runs parallels, word studies, parsing
| Writes 01-topics.md, 02-verses.md, 04-word-studies.md
| Saves raw tool output to raw-data/
Phase 3: Analysis Agent
| Reads clean research files
| Applies the evidence classification methodology
| Writes 03-analysis.md and CONCLUSION.md
Why three agents instead of one?
- The scoping agent prevents training-knowledge bias. Scope comes from tool discovery, not from what the AI "knows" about theology.
- The research agent gets a fresh context window dedicated to data gathering. This maximizes the amount of data it can collect without running out of context.
- The analysis agent gets a fresh context window loaded with clean, organized research. This maximizes its capacity for synthesis and careful reasoning.
The Study Files¶
Each study directory contains these files, produced by the pipeline:
| File | Produced By | Contents |
|---|---|---|
PROMPT.md |
Scoping Agent | The research brief: tool-discovered topics, verses, Strong's numbers, related studies, and focus areas |
01-topics.md |
Research Agent | Nave's Topical Bible entries with all verse references for each topic |
02-verses.md |
Research Agent | Full KJV text for every verse examined, organized thematically |
04-word-studies.md |
Research Agent | Strong's concordance data: Hebrew/Greek words, definitions, translation statistics, verse occurrences |
raw-data/ |
Research Agent | Raw tool output archived by category (Strong's lookups, parsing, parallels, etc.) |
03-analysis.md |
Analysis Agent | Verse-by-verse analysis with full evidence classification applied |
CONCLUSION.md |
Analysis Agent | Evidence tables (E/N/I), tally, and final assessment |
Data Sources¶
The tools draw from these primary data sources:
| Source | Description | Size |
|---|---|---|
| KJV Bible | Complete King James Version text | 31,102 verses |
| Nave's Topical Bible | Orville J. Nave's topical dictionary | 5,319 topics |
| Strong's Concordance | James Strong's exhaustive concordance with Hebrew/Greek lexicon | Every word in the KJV mapped to original language |
| BHSA (Biblia Hebraica Stuttgartensia Amstelodamensis) | Hebrew Bible linguistic database via Text-Fabric | Full morphological parsing of every Hebrew word |
| N1904 (Nestle 1904) | Greek New Testament linguistic database via Text-Fabric | Full morphological parsing of every Greek word |
| Textus Receptus | Byzantine Greek text tradition | For textual variant comparison |
| LXX Mapping | Septuagint translation correspondences | Hebrew-to-Greek word mappings |
| Sentence embeddings | Pre-computed semantic vectors | For semantic search across all sources |
Evidence Classification Methodology¶
The core of the methodology is a three-tier evidence classification system that distinguishes between what Scripture directly states, what necessarily follows from it, and what positions claim it implies.
The Three Tiers¶
E -- Explicit. "The Bible says X." You can point to a verse that says X. A close paraphrase of the actual words of a specific verse, with no concept, framework, or interpretation added beyond what the words themselves require.
N -- Necessary Implication. "The Bible implies X." You can point to verses that, when combined, force X with no alternative. Every reader from any theological position must agree this follows -- no additional reasoning is required.
I -- Inference. "A position claims the Bible teaches X." No verse explicitly states X, and no combination of verses necessarily implies X. Something must be added beyond what the text contains.
Critical rule: Inferences cannot block explicit statements or necessary implications. If E and N items establish X, the existence of passages that could be inferred to teach not-X does not prevent X from being established.
The 4-Type Inference Taxonomy¶
Inferences are further classified on two dimensions:
| Derived from E/N | Not derived from E/N | |
|---|---|---|
| Aligns with E/N | I-A (Evidence-Extending) | I-C (Compatible External) |
| Conflicts with E/N | I-B (Competing-Evidence) | I-D (Counter-Evidence External) |
I-A (Evidence-Extending): Uses only vocabulary and concepts found in E/N statements. An inference only because it systematizes multiple E/N items into a broader claim. Strongest inference type.
I-B (Competing-Evidence): Some E/N statements support it, but other E/N statements appear to contradict it. Genuine textual tension where both sides can cite Scripture. Requires the SIS Resolution Protocol.
I-C (Compatible External): Reasoning from outside the text (theological tradition, philosophical framework, historical context) that does not contradict any E/N statement. Supplemental only.
I-D (Counter-Evidence External): External concepts that require overriding, redefining, or qualifying E/N statements to be maintained. Weakest inference type.
Evidence hierarchy: E > N > I-A > I-B (resolved by SIS) > I-C > I-D
Positional Classification¶
Evidence items are classified by position (Historicist, Anti-Historicist, or Neutral/Shared) based on the same methodology used across the series. Items are classified positionally only when one side must deny the textual observation. Factual observations that both sides must accept are classified Neutral regardless of which side cites them.