Methodology
This page is the canonical V1 method for the workbench. It defines how a source may become a corpus record, how a sentence may become an annotation, and how an annotation may support a case-level or cross-case claim.
Method reader’s guide
| Need | Start here |
|---|---|
| Understand the full study design and claim limits | Research Design |
| Review metaphor-identification rules | MIPVU Annotation Guide |
| Inspect corpus selection, provenance, and rights | Corpus and Sources and Corpus Rights Policy |
| Understand blind multi-model annotation diagnostics | Multi-Model Annotation Stress Test |
| Understand blind human double-coding and adjudication | Human Reliability and Adjudication Methodology and architecture detail |
| Recruit and onboard source-language human coders | Human Coder Recruitment and Onboarding Protocol |
| Prepare a qualified human coder for calibration | Human Coder Training Guide |
| Calibrate source-language human coders | Multilingual Human Coder Calibration Packets |
| Understand human reliability sample selection | Stratified Human Reliability Sampling Strategy |
| Check current model-stress-test execution and publication limits | Model Reliability Results and Model-Reliability Disclosure |
| Understand validation and audit controls | Validation Protocol and Research Archive |
The sequence below is the canonical pipeline. Detailed governance and review documents remain available without occupying the primary reader path.
Research Design
This project assesses the degree of evidentiary support for Richard A. Koenigsberg’s Law of Sacrifice through case-centered corpus analysis. The working question is how recurring metaphor systems in leader-centered political corpora, when compared with historically documented practices of mobilization, killing, dying, purification, and enemy destruction, support, complicate, or limit Koenigsberg’s Law, its body-politic corollary, and the construction of enemies as bringers of death.
The fuller publication-facing protocol is maintained in RESEARCH_DESIGN.md. That document defines the project aim, research questions, corpus scope, inclusion and exclusion rules, units of analysis, method sequence, validation strategy, AI-use policy, data availability policy, and expected outputs for the scholarly upgrade.
V1 uses a balanced-core plus extended-corpus design.
The balanced core is the basis for cross-case comparison. Each case should move toward a comparable core of high-priority documents that covers major phases, registers, and rhetorical settings. The initial target is 25,000-50,000 words per case, revised after source discovery and rights review.
The extended corpus preserves additional material that is analytically useful but not balanced across cases. Extended materials may support case-level close reading, source discovery, and hypothesis generation, but cross-case frequency claims should identify whether they use the balanced core or the extended corpus.
The cases are leader-centered because the project is studying how political symbols, sacred objects, enemies as bringers of death, and sacrificial obligations are organized in public and semi-public language. This does not mean that leaders are treated as isolated causes of war or genocide. It means their corpora provide an auditable entry point into the symbolic systems through which violence was authorized, interpreted, and made morally necessary.
V1 can support claims about textual patterns, metaphor clusters, registers, diachronic shifts, absences, historical alignment, and competing interpretations. V1 cannot by itself diagnose historical actors, infer private mental states, prove monocausal claims about historical violence, or establish Koenigsberg’s Law as a final explanation.
Corpus Construction
Corpus construction begins with candidate sources, not with raw text ingestion. A candidate becomes an approved corpus record only after it has sufficient metadata, a rights review, and an explicit inclusion rationale.
Minimum metadata for an approved document includes:
- stable
document_id; - title and short title;
- date and date precision;
- register, genre, phase, and authorship fields;
- source URL or bibliographic citation;
- rights status and storage decision;
- expected raw-text path;
- verification expectations such as minimum word count and anchor phrases when feasible;
- analytical priority;
- risk flags and notes.
Sources are prioritized by relevance, provenance, register value, and rights clarity. Public-domain and open-access sources are preferred. Translations are permitted when necessary, but translation risk must be recorded and later tested through sensitivity checks where feasible.
The rights gate is mandatory. No text should be copied into cases/<case>/corpus/raw/ until the rights policy permits committed storage, gitignored local storage, metadata-only reference, or an explicit unavailable status. Missing or inaccessible texts should be documented rather than silently omitted.
Acquisition is now a controlled local workflow rather than an ad hoc copy step. The project uses scripts/acquire-sources.py to report source readiness, scripts/fetch-corpus.py for supported direct-fetch sources, source-specific download skills for browser-rendered or human-assisted sources, and scripts/verify-corpus.py as the required post-acquisition integrity gate.
Register balance matters because metaphor frequency is register-sensitive. Formal speeches, legal texts, military orders, manifestos, letters, memoirs, and ceremonial addresses do not make identical rhetorical demands. Frequency claims should therefore identify the register mix behind the claim and, when possible, report register-controlled versions.
Phase coverage matters because sacrificial language may develop under pressure. Each case should identify relevant historical phases before final corpus selection, then track whether the corpus supports diachronic analysis or only case-level close reading.
Annotation Sequence
Annotation proceeds in four layers. Each layer should preserve sentence IDs, source spans, confidence, and uncertainty notes.
First, identify metaphor-related lexical units using MIPVU. The annotator works from source-language lexical-unit worklists in cases/<case>/corpus/mipvu/. Every word-like lexical unit receives a decision. Detailed contextual meaning, basic meaning, dictionary/source citation, contrast explanation, comparison basis, confidence, and review notes are mandatory for metaphor-related or uncertain decisions. German and French are annotated in the source language, with English glosses used only as analytical aids.
Second, identify the CMT evidence from MIPVU-positive or uncertain lexical units. The annotator records the linguistic span, MIPVU IDs, source domain, target domain, mappings, entailments, linguistic form, and whether the metaphor is conventional, active, novel, extended, or co-activated with another cluster.
Third, add Koenigsbergian interpretation. This layer may classify fantasy type, magical object, violence logic, guilt distribution, obligatory frame, sacrificial economy, psychic defense, and exit condition. These fields interpret the metaphor pattern; they do not license diagnosis or claims about private mental states.
Fourth, record absence and suppression. Absence claims require an explicit search scope. A sentence, document, register, phase, or case may support an absence claim only when the analyst can state what was searched, what would have counted as presence, and why the absence is analytically meaningful rather than merely missing data.
When evidence is ambiguous, the annotation should preserve the ambiguity rather than force a cluster assignment. Low-confidence or contested annotations remain usable for exploratory review, but should be filtered out of high-confidence aggregate claims unless explicitly discussed.
Cluster Method
Starter clusters are hypotheses, not findings. They provide a controlled vocabulary for first-pass annotation and cross-case comparison, but they must be revised in response to corpus evidence.
A cluster may be retained when repeated annotations show stable source domains, target domains, entailments, and political-psychological work. A cluster may be merged when two labels repeatedly capture the same mapping. A cluster may be split when a single label hides distinct source domains, violence logics, or register effects. A cluster may be rejected when it does not appear, appears only as a weak artifact of the prompt, or fails rival-explanation review.
Case-specific clusters remain case-local until cross-case mapping justifies a shared category. Shared clusters should preserve both similarity and difference: a body-politic cluster in one case may perform healing, purifying, sacrificing, or exterminating work in another. The mapping should not flatten those differences into a generic label.
Cluster changes should be reflected in case configuration, concordance outputs, analysis artifacts, and the relevant methodology or status notes.
Analysis Method
The pipeline produces rebuildable case artifacts:
manifest -> acquisition/verification -> raw -> text -> segmented JSON -> MIPVU JSON -> annotated JSON -> concordance -> analysis
The concordance is the main evidence index. It records MIPVU lexical-unit denominators and flattens interpretive annotations across a case, indexing them by document, register, cluster, fantasy type, violence logic, absence flag, high-confidence status, and suppression status.
Case analysis recomputes aggregate distributions from the concordance while preserving manually written interpretive notes. Aggregate claims should identify their denominator, corpus scope, confidence threshold, and whether they are balanced-core or extended-corpus claims.
Confidence thresholds should be used conservatively. High-confidence claims should generally rely on annotations with strong textual evidence and limited ambiguity. Exploratory claims may use broader sets, but must be labeled as exploratory.
Diachronic analysis should be attempted only when date coverage supports it. Register analysis should be attempted whenever registers vary enough to affect metaphor density or interpretive weight. Absence analysis should distinguish between a local nonappearance, a register effect, a source limitation, and a structurally meaningful suppression.
Cross-case synthesis should proceed from case outputs into cases/x-case/, not from direct impressionistic comparison. Shared claims should cite the case-level concordance, analysis, validation, or synthesis artifacts that support them.
Claim Discipline
Every public claim should be classified by evidentiary status.
- Evidence: directly observable source text, metadata, annotation, or pipeline output.
- Interpretation: an argued reading of evidence using CMT, Koenigsbergian theory, corpus analysis, or close reading.
- Inference: a cautious claim that follows from multiple pieces of evidence but is not directly observed.
- Speculation: a possibility worth recording but not yet supported enough to function as a finding.
- Open question: a live uncertainty that requires more sources, annotation, validation, or theoretical work.
The project should prefer qualified, auditable claims over dramatic certainty. Claims about war, genocide, sacrifice, sacred objects, body-politic fantasies, and psychological mechanisms should remain proportional to the corpus evidence and should preserve rival explanations where they remain plausible.