cfs-mapper

name

cfs-mapper

description

Map any content — articles, news, books (fiction and non-fiction), scriptures, legal filings, transcripts, technical documentation, academic papers — into a valid Claim Formalization Schema (CFS) v0.5.0 JSON instance. Use this skill whenever the user asks to "formalize", "extract claims from", "map to CFS", "analyze the argument structure of", "decompose the logic of", or otherwise convert textual content into structured claim graphs. Also trigger when the user uploads or pastes content and asks for "propositions", "claim extraction", "argument mapping", "logical decomposition", "epistemic analysis", or "narrative formalization". This skill handles all source kinds: journalism, technical_documentation, legal_filing, policy_paper, transcript, academic_paper, fiction, scripture, and other. Even if the user does not mention CFS by name, trigger this skill whenever structured claim extraction from text is requested.

CFS Mapper

Map any textual content into a valid Claim Formalization Schema (CFS) v0.5.0 JSON instance that passes both Zod structural parsing and programmatic referential-integrity validation.

Before you begin

1. Read references/cfs-schema-reference.md — it contains the full type inventory, enum values, and validation rules (V1–V21, L1–L10) you must satisfy.
2. Read references/mapping-guide.md — it contains decision tables and worked examples for each content type.

These two files are your ground truth. Do not guess enum values or field constraints from memory; confirm against the reference.

---

Core Workflow

Phase 1 — Classify the Source

Determine these three properties before extracting anything:

Property	How to decide
source_kind	Match the genre to one of: journalism, technical_documentation, legal_filing, policy_paper, transcript, academic_paper, fiction, scripture, other.
truth_regime	realist for empirical claims about the real world. diegetic for fiction (claims true within the story). scriptural for sacred texts with doctrinal authority. mixed when the text blends regimes (e.g., historical fiction, exegesis).
narrative_frame	Required when source_kind ∈ {fiction, scripture} or truth_regime ∈ {diegetic, scriptural}. Build the layers array from outermost narrator inward.

If source_kind is fiction or scripture, or truth_regime is diegetic or scriptural, you must populate narrative_frame — the schema enforces this (V17). Failure to do so is a parse error.

Phase 2 — Entity Extraction

Scan the content and register every agent, organization, place, or notable object as an Entity.

Rules:

• IDs are UPPER_SNAKE_CASE (/^[A-Z][A-Z0-9_]*$/).
• IDs must not collide with reserved prefixes P, R, Q, S, A followed by digits (e.g., P1 is invalid as an entity ID).
• Choose ontological_status carefully: real for real-world, fictional for in-story, disputed for contested existence, traditional for figures whose historicity is a matter of tradition rather than evidence.
• Assign roles generously — they feed lint rule L7 (revelation without authority role).
• Use affiliations, developer, parent, founder, associated_with, location to link entities. Every ID referenced here must be declared.
• For fiction/scripture: include the narrator(s) and audience(s) referenced in narrative_frame as entities.
• For journalism: include an Editorial_Voice entity for the publication whenever editorial synthesis or unsigned editorial claims appear.

Phase 3 — Proposition Extraction

This is the core analytical step. For each substantive claim in the text, create a Proposition.

Decision checklist per proposition:

1. speaker: Who is making this claim? Must be a declared EntityID.
2. claim: State the claim in your own words as a clear, atomic declarative sentence. One claim per proposition — do not bundle.
3. speech_act: Choose from assertion, directive, recommendation, definition, editorial, prediction, narration, characterization, figuration, revelation. See the mapping guide for content-type-specific heuristics.
4. verifiability: Choose from machine_verifiable, empirically_testable, expert_judgment, unfalsifiable, diegetic_fact, diegetic_testimony, traditional_doctrine.
   • Under scriptural regime: avoid machine_verifiable (lint L1).
   • Under diegetic regime: avoid machine_verifiable (lint L2).
5. fidelity: How faithfully does the source render this claim? verbatim_quote, close_paraphrase, loose_paraphrase, editorial_synthesis, authorial_construction.
6. literal: Default true. Set to false for metaphors, irony, hyperbole. If speech_act is figuration, literal must be false (V18).
7. provenance: Use when the claim passed through intermediaries. Min 2 hops. provenance[0].speaker must equal the proposition's speaker (V19). Chain runs outer→inner.
8. anchor: Optional locator within the source (page, paragraph, chapter, verse, timestamp).

Quantity guidance: Extract substantive claims — aim for thorough coverage of the text's argumentative or narrative structure, but do not pad with trivial restatements. A short article might yield 5–15 propositions; a book chapter 20–50; a scripture passage 10–30.

Phase 4 — Axiom Identification

Axioms are unstated foundational assumptions the text relies upon but does not argue for. They are background premises.

• Mark controversial: true when the axiom is debatable (feeds lint L8).
• Assign a domain when the axiom belongs to a specific field.
• Axioms are optional (the array can be empty), but look for them — they are epistemically important.

Phase 5 — Rule Construction

Rules express logical or argumentative relations between propositions and/or axioms.

• antecedent_ids and consequent_ids accept only P<n> or A<n> IDs.
• relation_type must match the form arrow (lint L5): entails → →, motivates → ⇝, contradicts → ⇄, weakens → ↓, strengthens → ↑, is_consistent_with → ∥, presupposes → ⊲, refines → ≻, mirrors → ⇔.
• plain is a human-readable restatement of the rule.
• Watch for lint L3 (entails from unfalsifiable), L8 (controversial axiom in rule), L10 (cross-regime entailment under mixed).

Phase 6 — Open Questions

Identify at least one genuine question the text raises, leaves unanswered, or that a critical reader would ask.

• kind: factual, conceptual, methodological, ethical.
• related: At least one ID. Can reference P, R, A, S, or Entity IDs — but not other Q IDs.

Phase 7 — Silences

Silences capture what the text doesn't say. These are critical for intellectual honesty.

• kind: absent_perspective, unasked_question, suppressed_counter, missing_context.
• related_propositions: At least one P<n> ID.
• affected_entities: Optional list of EntityIDs impacted by the silence.
• Silences are optional at the schema level, but the skill should always attempt to identify at least one unless the text is genuinely exhaustive (rare).

Phase 8 — Assembly & Self-Validation

Before emitting the final JSON:

1. ID uniqueness (V4): Every ID in each namespace must be unique.
2. Referential integrity (V8–V12): Every EntityID, PropositionID, AxiomID, RuleID, SilenceID referenced anywhere must be declared.
3. V17: narrative_frame present when required.
4. V18: figuration ⟹ literal: false.
5. V19: provenance[0].speaker == proposition.speaker.
6. V20: Provenance chain length ≥ 2 when present.
7. Lint scan: Check L1–L10 and resolve or annotate.
8. schema_version: Always "0.5.0".

Walk through these checks mentally before outputting. If you find a violation, fix it — do not emit invalid JSON.

---

Output Format

Emit a single JSON code block (```json ... ```) containing the full CfsInstance. The JSON must:

• Be valid JSON (no trailing commas, no comments).
• Parse successfully against the CFS v0.5.0 Zod schema.
• Pass the validate() referential integrity checks with zero errors.
• Minimize lint warnings (zero is ideal, but some are acceptable if annotated in note fields).

If the content is very long (full book, lengthy scripture), offer to formalize it chapter-by-chapter or section-by-section, each as a separate CFS instance. Ask the user how they want to scope it.

---

Important Reminders

• Never invent claims not present in the source. Every proposition must trace back to something the text actually says or clearly implies.
• Distinguish speaker from author. In journalism, the journalist reports what sources say — track provenance. In fiction, the narrator speaks, not the author.
• Axioms are unstated. If the text argues for it, it's a proposition, not an axiom.
• Be honest about verifiability. Don't upgrade expert_judgment to machine_verifiable or downgrade unfalsifiable to empirically_testable.
• Silences matter. The absence of counterargument, alternative perspective, or context is itself a finding.
• When in doubt, add a note. The note field exists on most objects. Use it to flag uncertainty, editorial judgment calls, or caveats.