Interdisciplinary Contradiction Detection Synthesis

Document Classification: Cross-Domain Methodological Integration Version: 1.0 Date: 2026-01-17 Purpose: Synthesize contradiction detection methods across six investigative domains for Phronesis FCIP implementation

Executive Summary

Contradiction detection is the universal cornerstone of professional investigation across six distinct domains: police investigations, investigative journalism, legal eDiscovery, regulatory compliance, intelligence analysis, and academic research. While each domain has evolved specialized techniques adapted to its operational constraints and legal frameworks, all converge on systematic comparison, evidence hierarchies, temporal analysis, and quality control through multiple reviewers.

Core Finding: The eight S.A.M. contradiction types (SELF, INTER_DOC, TEMPORAL, EVIDENTIARY, MODALITY_SHIFT, SELECTIVE_CITATION, SCOPE_SHIFT, UNEXPLAINED_CHANGE) manifest consistently across all domains, but detection methods vary based on:

Adversarial pressure (criminal vs. civil vs. internal investigation)
Evidentiary standards (beyond reasonable doubt vs. balance of probabilities vs. preponderance of evidence)
Time constraints (hours for operational intelligence vs. months for legal discovery)
Resource availability (technology-enabled at scale vs. manual expert review)

Key Convergence Points:

Evidence Hierarchies: All domains prioritize authenticated primary documents over testimonial evidence
Timeline Construction: Systematic chronologies with evidence linking are universal (8-step legal, 5WH police, intelligence chronologies)
Multi-Source Fusion: Contradictions emerge when comparing sources with independent provenance (CPIA "reasonable lines," journalism "triangulation," intelligence multi-INT)
Quality Control: Minimum 3 independent reviewers required for reliable contradiction detection (regulatory panels, journalism fact-checking, intelligence peer review, academic inter-rater reliability with Cohen's Kappa ≥0.70)
Bias Mitigation: Structured techniques prevent confirmation bias (Intelligence SATs, Legal TAR validation, Academic reflexivity, Journalism hypothesis testing)

Implementation Impact: By synthesizing these methods, Phronesis FCIP can provide a unified contradiction detection framework that meets the quality standards of the most rigorous domain while remaining accessible across investigative contexts.

1. Cross-Domain Patterns in Contradiction Detection

1.1 Universal Process Architecture

All six domains follow a four-stage contradiction detection process:

Stage 1: Evidence Collection

Systematic identification of relevant materials
Authentication and provenance verification
Metadata preservation and chain of custody
Variation: Legal discovery (litigation hold), Journalism (source verification), Intelligence (Admiralty Code rating)

Stage 2: Normalization and Structuring

Convert heterogeneous sources into comparable format
Extract entities, dates, claims, assertions
Build evidence repositories with standardized tagging
Variation: Legal (Bates numbering), Police (HOLMES2 indexing), Academic (QDAS coding)

Stage 3: Systematic Comparison

Compare statements within documents (SELF)
Compare across documents (INTER_DOC)
Compare against timelines (TEMPORAL)
Compare against external evidence (EVIDENTIARY)
Variation: Legal (email threading, version tracking), Intelligence (ACH matrix), Journalism (ChronoFact verification)

Stage 4: Validation and Quality Control

Independent review by minimum 3 reviewers
Statistical reliability testing (where applicable)
Documentation of analytic reasoning
Appeals/challenge mechanism
Variation: Legal (TAR validation testing), Academic (inter-rater reliability), Intelligence (Red Cell review), Regulatory (dual decision-maker panels)

1.2 Shared Failure Modes

Confirmation Bias

Manifestation: Seeking contradictions that support initial hypothesis while ignoring exculpatory inconsistencies
Police: CPIA mandate to pursue reasonable lines "towards AND away from" suspect
Legal: TAR 2.0 continuous learning vs. fixed seed set
Intelligence: ACH methodology (disprove hypotheses, don't confirm)
Journalism: Hypothesis-based framework (explicitly separate facts from assumptions)
Academic: Reflexivity and member checking
Regulatory: Three-stage analysis requiring contextual factors review

Temporal Anchoring

Manifestation: Over-relying on first-received evidence; failing to update when timeline shifts
Police: PEACE interview model (obtain free narrative before challenging)
Legal: 8-step timeline process (continuous updating)
Intelligence: Bayesian updating (revise probabilities with new evidence)
Journalism: Iterative verification protocols
Academic: Grounded Theory (constant comparison, theoretical saturation)
Regulatory: Real Prospect Test (re-evaluation at each stage)

Source Corruption (Circular Reporting)

Manifestation: Two sources appear independent but share common origin
Police: Chain of custody tracking, witness isolation
Legal: Email threading (detect forwarded vs. original)
Intelligence: Source network mapping (detect circular reporting)
Journalism: Independent corroboration requirement (three-source rule)
Academic: Citation network analysis
Regulatory: Independent expert requirement

1.3 Technology Integration Spectrum

Domain	Technology Maturity	Primary Tools	Contradiction Detection Automation
Legal eDiscovery	Highest	TAR 2.0/CAL, near-duplicate detection, email threading	60-80% automated (human review for validation)
Intelligence	High	ACH matrices, multi-INT fusion, network analysis	40-60% automated (SATs provide structure, human judgment required)
Journalism	Medium	ChronoFact, OCCRP Aleph (4B+ documents), Datashare	30-50% automated (pattern detection, human verification essential)
Police	Medium	HOLMES2 dynamic reasoning engine, digital forensics	20-40% automated (anomaly detection, expert review required)
Academic	Medium	NVivo, ATLAS.ti, MAXQDA (QDAS)	10-30% automated (coding support, human interpretation required)
Regulatory	Low	Document review platforms, case management	5-20% automated (primarily manual expert panels)

Trend: All domains converging on AI-assisted human review rather than full automation. Legal sector's TAR 2.0 court validation demonstrates that AI prioritization + human verification exceeds pure manual review in both cost and accuracy.

2. The Eight S.A.M. Contradiction Types Across Domains

2.1 SELF (Internal Contradictions Within Single Document)

Definition: Document contradicts itself, undermining internal coherence and reliability.

Police Detection Methods (CPIA "Reasonable Lines of Enquiry")

PEACE Interview: Obtain free narrative first, then challenge inconsistencies during "Account, Clarify, Challenge" phase
Cognitive Interview: Report everything → recall from different perspectives → detect narrative inconsistencies
Statement Analysis: Forensic linguistics to detect deception indicators (pronoun shifts, verb tense changes, minimization language)
Example: Suspect states "I never went to the warehouse" but later describes layout of warehouse interior

Journalism Detection (Hypothesis Testing)

Internal Consistency Check: Flag contradictory assertions within single document or interview
Chronological Analysis: Map document's internal timeline for logical impossibilities
Direct Quote Audit: Compare different statements from same source over time
Example: Panama Papers - Director claims "never acted for Russian clients" but company registry shows 15 Russian entities

Legal Detection (Document Version Tracking)

Near-Duplicate Detection: Fuzzy hashing (ssdeep, SimHash) identifies document versions
Diff Analysis: Highlight additions/deletions/modifications between versions (green/red/yellow)
Metadata Comparison: Track author changes, edit timestamps, revision history
Example: Contract v1 states "party shall deliver within 30 days" but v3 states "party may deliver within 90 days" without redline/explanation

Regulatory Detection (Three-Stage Analysis)

Factual Inconsistency Review: Professional panel identifies contradictory assertions in single submission
Standards Application: Test whether contradictions breach professional standards
Contextual Factors: Assess whether contradiction due to mistake vs. deliberate deception
Example: Psychologist's report states "no safeguarding concerns" in summary but detailed notes describe multiple red flags

Intelligence Detection (ACH Matrix)

Evidence Consistency Coding: Test whether evidence item is consistent (C) or inconsistent (I) with hypothesis
Self-Contradiction Flags: Automated detection when source rates as "A" (reliable) but information rates as "5" (improbable)
Source Deception Analysis: Identify patterns suggesting source was deceived or deliberately misleading
Example: Trusted HUMINT asset reports troop movements contradicted by their own subsequent reporting

Academic Detection (Grounded Theory Constant Comparison)

Within-Document Coding: Identify passages that contradict earlier passages
Memo-Writing: Document emerging contradictions for theoretical sampling
Negative Case Analysis: Actively seek disconfirming evidence within single source
Example: Interview transcript where participant says "I always followed protocol" but later describes multiple protocol violations

Implementation for Phronesis:

struct SelfContradiction {
    document_id: String,
    passage_1: TextSpan,  // First contradictory statement
    passage_2: TextSpan,  // Second contradictory statement
    contradiction_type: SelfContradictionType,
    confidence_score: f32,
    detection_method: DetectionMethod,
}

enum SelfContradictionType {
    FactualInconsistency,      // "Event A happened" vs. "Event A didn't happen"
    TemporalIncoherence,       // Timeline logically impossible
    ModalityShift,             // "Definitely occurred" vs. "Possibly occurred" (same claim)
    OmissionAfterAssertion,    // Claims knowledge then denies it
}

enum DetectionMethod {
    NLPContradictionDetection,  // Semantic analysis
    TemporalLogicValidator,     // Timeline consistency checker
    EntityResolution,           // Detect pronoun/reference ambiguity
    HumanReview,                // Expert identification
}

2.2 INTER_DOC (Cross-Document Contradictions)

Definition: Statements in one document contradict statements in another, revealing inconsistencies across record corpus.

Police Detection (HOLMES2 Cross-Referencing)

Witness Statement Comparison: HOLMES2 flags contradictions between witness accounts
Timeline Overlay: Multi-source chronology highlights conflicting date/time claims
Intelligence Cross-Check: Test witness statements against CCTV, phone records, forensic evidence
Example: Witness A says "I saw Smith at 3pm" but CCTV shows Smith elsewhere; Witness B says "Smith arrived at 5pm"

Journalism Detection (Multi-Source Triangulation)

Three-Source Rule: Independent confirmation from minimum three sources before publication
Document vs. Document: Cross-reference leaked documents against official records
Interview vs. Interview: Compare statements from multiple interviewees
ChronoFact Database: Systematic timeline verification across 1000+ sources
Example: Pandora Papers - Offshore entity registration dates contradict political declarations of divestment

Legal Detection (Email Threading and Timeline Construction)

Email Threading: Group related communications to detect contradictory statements across thread
Timeline Overlays: 8-step process creates central repository with Bates-numbered evidence
Communication Network Analysis: Detect when different custodians provide contradictory accounts
Example: CEO email to CFO: "We were aware of defect by March 1" contradicts CEO deposition: "First learned of defect May 15"

Regulatory Detection (Expert Panel Review)

Multi-Document Analysis: Panel compares practitioner's records against third-party evidence
Corroboration Requirements: Single allegation insufficient; require pattern across multiple incidents
Independent Expert Reports: Commission external review when internal contradictions unresolvable
Example: Practitioner's contemporaneous notes contradict subsequent formal report; patient records provide third data point

Intelligence Detection (Multi-INT Fusion)

Eight INT Types: HUMINT, SIGINT, IMINT, OSINT, GEOINT, FININT, TECHINT, MASINT
Three Fusion Levels: Data-level, feature-level, decision-level
Contradiction Handling Strategies: Discounting (reduce weight of less reliable source), temporal explanation, deception hypothesis
Admiralty Code: Rate source reliability separately from information credibility
Example: HUMINT reports "missile site at Location X" but IMINT shows no construction; contradiction triggers collection of additional SIGINT

Academic Detection (Thematic Analysis Across Cases)

Cross-Case Synthesis: Compare themes across multiple interview transcripts or documents
Constant Comparison: Grounded Theory technique comparing data slices across corpus
Divergent Case Analysis: Identify outliers that contradict emerging patterns
Example: 18 participants report positive experience; 2 report severe negative outcomes; analyze for systematic differences

Implementation for Phronesis:

struct InterDocContradiction {
    doc_pair: (String, String),    // IDs of contradicting documents
    claim_1: Claim,
    claim_2: Claim,
    contradiction_evidence: ContradictionEvidence,
    source_reliability: (AdmiraltyCode, AdmiraltyCode),  // Rate each source
    resolution_hypothesis: Vec<ResolutionHypothesis>,
}

struct Claim {
    document_id: String,
    text_span: TextSpan,
    claim_type: ClaimType,
    entities: Vec<Entity>,
    temporal_reference: Option<DateTime<Utc>>,
}

enum ClaimType {
    FactualAssertion,
    Opinion,
    Hearsay,
    ExpertJudgment,
}

struct ResolutionHypothesis {
    hypothesis: String,  // e.g., "Source A was deceived", "Temporal: situation changed between observations"
    supporting_evidence: Vec<String>,
    confidence: f32,
}

2.3 TEMPORAL (Timeline and Date Inconsistencies)

Definition: Events, statements, or evidence conflict on sequencing, timing, or chronological plausibility.

Police Detection (5WH Framework - "When")

Chronology Construction: Timeline from first report through investigation completion
Cell Site Analysis: Phone location data creates precise temporal record
CCTV Timeline: Frame-by-frame analysis for exact event sequencing
Forensic Timeline: Digital forensics (file creation/modification), physical evidence (entomology for time of death)
Gap Analysis: Identify suspicious periods with no documentation (potential evidence destruction)
Example: Suspect claims "home all evening" but cell site data places phone at crime scene 8:47pm-9:13pm

Journalism Detection (ChronoFact Verification)

ChronoFact Database: 1000+ sources, systematically verified chronology
5WH Questions: When did event occur? (precise timestamp)
Event Sequence Verification: Confirm A happened before B (causal claims require temporal sequence)
Metadata Examination: Document creation dates, email timestamps, photo EXIF data
Example: Politician claims "divested from company before vote" but corporate filings show ownership through vote date + 3 months

Legal Detection (8-Step Timeline Construction)

Step 1: Start early (case inception, not trial preparation)
Step 2: Central repository (chronological database with Bates linking)
Step 3: Party and issue tagging
Step 4: Collect documentation (emails, calendar entries, public records)
Step 5: Extract date-specific events
Step 6: Objective language (facts only, no conclusions)
Step 7: Link evidence (every event cites supporting documents)
Step 8: Update continuously
Temporal Analysis Techniques:
- Gap Analysis: Suspicious periods with no documentation
- Density Analysis: Spikes in activity (e.g., 200 emails in 48 hours before product recall)
- Contradiction Detection: Event sequence contradicts witness testimony
Example: Witness deposition: "First learned of issue May 1" contradicts email received April 15

Regulatory Detection (Contemporaneous Records Requirement)

Primacy of Contemporaneous Evidence: Notes made at time of event > retrospective reports
Temporal Coherence Review: Panel assesses whether practitioner's timeline logically consistent
Corroboration with Third-Party Timestamps: Medical records, appointment systems, phone logs
Example: Practitioner claims "supervision session occurred February 10" but supervisor's calendar shows no meeting; no contemporaneous notes exist

Intelligence Detection (Temporal Explanation for Contradictions)

Multi-INT Fusion Contradiction Handling: "Situation changed between observations" as resolution hypothesis
Intelligence Cycle Timing: Collection → Processing → Analysis (time lag creates temporal gaps)
Operational Intelligence (F3EAD): Cycle time measured in hours; temporal precision critical for "Fix" phase
Example: HUMINT reports "no troops at border" (January 15) but IMINT shows buildup (January 20); temporal gap explains discrepancy

Academic Detection (Longitudinal Analysis)

Within-Case Temporal Sequencing: Track participant statements across multiple interview waves
Process Tracing: Identify turning points and causal sequences
Retrospective Accounts vs. Contemporaneous Records: Compare memory-based narratives to documents created in real-time
Example: Participant's retrospective account describes "gradual decline" but contemporaneous notes show sudden event

Implementation for Phronesis:

struct TemporalContradiction {
    events: Vec<TemporalEvent>,
    contradiction_type: TemporalContradictionType,
    temporal_gap: Option<Duration>,
    confidence: f32,
}

struct TemporalEvent {
    event_id: String,
    description: String,
    timestamp: TimeReference,
    source_docs: Vec<DocumentRef>,
    entities_involved: Vec<Entity>,
}

enum TimeReference {
    Precise(DateTime<Utc>),
    Range(DateTime<Utc>, DateTime<Utc>),
    Relative(String),  // "three days after incident X"
    Vague(String),     // "sometime in early March"
}

enum TemporalContradictionType {
    ImpossibleSequence,     // Event A requires B to happen first, but B happened after A
    ConflictingTimestamps,  // Two sources give different times for same event
    TemporalGap,            // Expected evidence missing for time period
    DensityAnomaly,         // Unusual spike/drop in activity
}

2.4 EVIDENTIARY (Evidence vs. Claims Mismatch)

Definition: Factual claims lack evidentiary support or directly contradict available evidence.

Police Detection (CPIA Disclosure and Forensic Evidence)

Material Evidence Standard: All evidence "reasonably capable of undermining prosecution or assisting defence"
Forensic Evidence Priority: Physical evidence (DNA, fingerprints, digital forensics) > testimonial
Unused Material Schedule: Document all material that doesn't support prosecution (exculpatory evidence)
Example: Defendant claims self-defense; forensic evidence shows victim shot in back while fleeing (contradicts claim)

Journalism Detection (Evidence Hierarchy)

Primary Documents > Secondary Sources > Testimonial
Authentication Requirements:
- Financial documents: Verified through corporate registries, banking records
- Official documents: Cross-referenced with government databases
- Leaked documents: Cryptographic verification, metadata analysis, expert authentication
Verification Before Publication: Minimum three independent sources confirming key facts
Example: Panama Papers - Politician denies offshore holdings; authenticated corporate registry documents prove beneficial ownership

Legal Detection (Evidence Linking in Timelines)

Bates Number Citation: Every timeline event cites supporting documents (ABC00001-ABC00010)
Deposition Citations: Smith Dep. 45:12-18 (transcript page:line)
Admissibility Standard: FRE 902(13)/(14) self-authenticating digital records with hash certification
Privilege Review: Four-part test for attorney-client privilege; document every claim
Example: Plaintiff claims "defendant knew of defect" but no emails, memos, or meeting notes mention defect

Regulatory Detection (Professional Standards Application)

Three-Stage Test:
1. Facts proven (balance of probabilities)
2. Standards breached (map facts to professional code)
3. Current impairment (fitness to practice)
Evidence Hierarchy: Contemporaneous clinical notes > retrospective reports > witness testimony
Example: Practitioner claims "followed supervision protocol" but no documented supervision sessions in 18-month period

Intelligence Detection (Admiralty Code and Multi-INT Corroboration)

Admiralty Code: [Source Reliability][Information Credibility]
- A1 = Completely reliable source + Confirmed information (highest)
- F6 = Unknown source + Cannot judge information (lowest)
Multi-INT Corroboration: Claim must be supported by minimum two independent INTs
Evidence vs. Assumption: ACH methodology requires explicit labeling of assumptions vs. facts
Example: HUMINT claims "missile site at Location X" (B3 rating: usually reliable source, possibly true) but IMINT shows no construction (A1 rating: completely reliable, confirmed) → Contradiction favors higher-rated evidence

Academic Detection (Triangulation and Member Checking)

Data Triangulation: Multiple data sources (interviews, documents, observations)
Investigator Triangulation: Multiple researchers independently code same data
Theory Triangulation: Apply multiple theoretical frameworks
Member Checking: Return findings to participants for validation
Negative Case Analysis: Actively seek evidence contradicting emerging themes
Example: Researcher claims "all participants reported positive outcomes" but document review reveals 2/20 filed formal complaints

Implementation for Phronesis:

struct EvidentaryContradiction {
    claim: Claim,
    contradicting_evidence: Vec<Evidence>,
    evidence_quality: EvidenceQuality,
    contradiction_severity: ContradictionSeverity,
}

struct Evidence {
    evidence_id: String,
    evidence_type: EvidenceType,
    source_reliability: AdmiraltyCode,
    authentication_status: AuthenticationStatus,
    provenance: Provenance,
}

enum EvidenceType {
    PrimaryDocument,
    ForensicAnalysis,
    DigitalForensics,
    ExpertOpinion,
    TestimonialAccount,
    PhysicalEvidence,
}

enum AuthenticationStatus {
    Authenticated { method: String, date: DateTime<Utc> },
    Pending,
    Unverified,
    Disputed { reason: String },
}

enum ContradictionSeverity {
    Direct,        // Evidence directly contradicts claim
    Indirect,      // Evidence undermines plausibility of claim
    Absence,       // Expected evidence missing (negative evidence)
}

2.5 MODALITY_SHIFT (Certainty and Tone Changes)

Definition: Unexplained shifts in certainty, epistemic stance, or tone regarding same claim without new evidence or justification.

Police Detection (Interview Analysis and Statement Comparison)

PEACE Model - "Account, Clarify, Challenge": Probe for explanations when account certainty shifts
Cognitive Interview - Recall from Different Orders: Disrupts script-based recall; reveals confidence variations
Statement Analysis: Forensic linguistics detect:
- Pronoun shifts ("I did X" → "one would do X")
- Verb tense changes (past → present when describing lie)
- Minimization language ("just," "only," "merely")
- Qualifiers appearing mid-narrative ("to be honest," "truthfully")
Example: Suspect initial statement: "I was definitely home all evening"; re-interview: "I think I was home, or maybe I went out briefly, I can't recall"

Journalism Detection (Quote Comparison and Modality Tracking)

Direct Quote Audit: Track exact phrasing across interviews
Hedging Analysis: Detect increased use of qualifiers ("possibly," "might," "perhaps") in later statements
On-Record vs. Off-Record Shifts: Note when source more definitive off-record
Example: Executive public statement: "No knowledge of misconduct"; private interview: "May have been aware of some irregularities"

Legal Detection (Document Version Tracking and Diff Analysis)

Near-Duplicate Detection: Identify document versions
Linguistic Diff Analysis: Highlight changes beyond factual updates:
- v1: "We are confident the product is safe"
- v2: "We believe the product is generally safe"
- v3: "The product may be safe under certain conditions"
Email Thread Analysis: Track certainty shifts across conversation
Expert Report Comparison: Detect opinion changes between draft and final reports
Example: Internal memo v1: "Defect requires immediate recall"; v2 (post-legal review): "Alleged defect warrants further investigation"

Regulatory Detection (Contemporaneous vs. Retrospective Accounts)

Primacy of Contemporaneous Records: Notes made at time > later formal reports
Modality Shift as Red Flag: Panel examines why practitioner more confident in retrospective account than contemporaneous notes suggested
Example: Contemporaneous clinical note: "Patient presented with possible indicators of abuse (uncertain)"; formal report submitted six months later: "Clear evidence of abuse was present"

Intelligence Detection (Confidence Level Tracking)

Words of Estimative Probability (WEP): Standardized probability language
- "Almost certainly" (95-99%) → "Likely" (60-80%) = significant downgrade
Confidence Levels: High, Moderate, Low (separate from probability)
Sensitivity Analysis (ACH Step 7): What would have to change for confidence to shift?
Example: Initial assessment: "Very likely (80-95%) that Event X will occur, high confidence"; revised assessment (no new evidence): "Possibly (40-60%), moderate confidence" → Flag unexplained modality shift

Academic Detection (Reflexivity and Researcher Positioning)

Reflexivity Journals: Document researcher's evolving interpretations
Theoretical Sensitivity: Note when researcher's confidence in emerging theory shifts
Audit Trail: Track coding decisions over time to detect interpretation drift
Example: Researcher memo week 4: "Preliminary theme: participants universally positive"; week 8: "Emerging complexity: subset report significant concerns" (shift justified by additional data collection)

Implementation for Phronesis:

struct ModalityShift {
    claim_id: String,
    statement_versions: Vec<StatementVersion>,
    modality_trajectory: ModalityTrajectory,
    justification_present: bool,
    new_evidence: Option<Vec<String>>,  // Document IDs of new evidence (if any)
}

struct StatementVersion {
    version_id: String,
    timestamp: DateTime<Utc>,
    text: String,
    modality_markers: Vec<ModalityMarker>,
    certainty_score: f32,  // 0.0 (no certainty) to 1.0 (absolute certainty)
}

struct ModalityMarker {
    marker_type: ModalityType,
    phrase: String,
    position: TextSpan,
}

enum ModalityType {
    Epistemic(EpistemicModality),  // Knowledge claims
    Deontic(DeonticModality),      // Obligation/permission
    Dynamic(DynamicModality),      // Ability/possibility
}

enum EpistemicModality {
    Certain,        // "definitely," "certainly," "absolutely"
    Probable,       // "likely," "probably," "presumably"
    Possible,       // "possibly," "might," "could"
    Uncertain,      // "uncertain," "unclear," "unknown"
}

enum ModalityTrajectory {
    IncreasingCertainty,  // Shift toward more definitive
    DecreasingCertainty,  // Shift toward less definitive (hedging)
    Stable,
}

2.6 SELECTIVE_CITATION (Cherry-Picking Evidence)

Definition: Citing only evidence that supports conclusion while omitting contradictory or qualifying evidence.

Police Detection (CPIA "Unused Material" Disclosure)

Duty to Disclose: All material "reasonably capable of undermining prosecution or assisting defence"
Disclosure Schedule: Prosecution must list ALL material (used and unused)
Unused Material Review: Defence can request access to non-disclosed material
Sanctions for Non-Disclosure: Case dismissal, conviction overturn (appeals)
Example: Prosecution presents 5 witnesses identifying defendant; withholds 3 witnesses who could not make identification

Journalism Detection (Systematic Review and Dissenting Views)

Document Full Population: Report total documents reviewed (e.g., "10,000 documents analyzed")
Dissenting Voices: Ethical obligation to include perspectives contradicting investigation's thesis
Right of Reply: Subject of investigation given opportunity to respond before publication
Example: Investigation cites 15 critical leaked emails; fails to mention 150 emails showing standard business practices

Legal Detection (Discovery Process and Privilege Logs)

Complete Production: Responding party must produce ALL responsive documents (not just favorable ones)
Privilege Logs: Documents withheld must be described (Bates, date, author, recipients, privilege basis)
Sampling and Validation: TAR 2.0 includes validation testing to detect under-production
Sanctions: Adverse inference instruction, monetary fines, case dismissal
Example: Company produces 1,000 documents supporting its position; TAR validation reveals 5,000 additional responsive documents withheld

Regulatory Detection (Comprehensive Record Review)

Standard of Proof: Balance of probabilities requires weighing ALL evidence
Panel Review: Professional and lay members independently review full case file
Defence Submissions: Practitioner can submit contradictory evidence
Example: Regulator cites 5 complaints against practitioner; omits 500 positive client testimonials and clean inspection record

Intelligence Detection (Alternative Analysis and Red Cell Review)

Structured Analytic Techniques: ACH requires testing ALL evidence against ALL hypotheses
Devil's Advocacy: Institutionalized role to challenge selective use of evidence
Red Cell Review: Independent team reviews for cherry-picking
CIA Red Cell Example: Challenge consensus view by highlighting contradictory intelligence
Example: Analyst cites 3 HUMINT reports suggesting imminent attack; Red Cell highlights 10 SIGINT intercepts suggesting no mobilization

Academic Detection (Systematic Review Protocols and Negative Case Analysis)

PRISMA 2020: Transparent reporting of search strategy, inclusion/exclusion criteria
Publication Bias: Systematic reviews must address (file-drawer problem)
Negative Case Analysis: Grounded Theory requires active search for disconfirming evidence
Reflexivity: Document researcher decisions on what to include/exclude
Example: Researcher claims "all participants reported X" but excluded 3 outliers without justification

Implementation for Phronesis:

struct SelectiveCitation {
    cited_evidence: Vec<EvidenceRef>,
    uncited_evidence: Vec<EvidenceRef>,
    selection_pattern: SelectionPattern,
    justification_quality: JustificationQuality,
}

struct EvidenceRef {
    evidence_id: String,
    relevance_score: f32,
    valence: Valence,  // Supports or contradicts conclusion
}

enum Valence {
    Supports(f32),      // 0.0 to 1.0 (strength of support)
    Contradicts(f32),   // 0.0 to 1.0 (strength of contradiction)
    Neutral,
    Mixed,
}

enum SelectionPattern {
    ConfirmatoryBias,    // Cited = 90% supportive, Uncited = 90% contradictory
    RepresentativeSample, // Cited reflects population distribution
    Systematic,          // Clear, justified criteria for selection
}

enum JustificationQuality {
    Explicit { criteria: String, applied_consistently: bool },
    Implicit { inferred_criteria: String },
    Absent,
}

// Detection algorithm
fn detect_selective_citation(
    all_evidence: Vec<EvidenceRef>,
    cited_evidence: Vec<String>,  // Evidence IDs cited in report
) -> Option<SelectiveCitation> {
    let cited: Vec<_> = all_evidence.iter().filter(|e| cited_evidence.contains(&e.evidence_id)).collect();
    let uncited: Vec<_> = all_evidence.iter().filter(|e| !cited_evidence.contains(&e.evidence_id)).collect();

    let cited_support_ratio = cited.iter().filter(|e| matches!(e.valence, Valence::Supports(_))).count() as f32 / cited.len() as f32;
    let uncited_contradict_ratio = uncited.iter().filter(|e| matches!(e.valence, Valence::Contradicts(_))).count() as f32 / uncited.len() as f32;

    if cited_support_ratio > 0.8 && uncited_contradict_ratio > 0.8 {
        Some(SelectiveCitation {
            cited_evidence: cited.into_iter().cloned().collect(),
            uncited_evidence: uncited.into_iter().cloned().collect(),
            selection_pattern: SelectionPattern::ConfirmatoryBias,
            justification_quality: JustificationQuality::Absent,
        })
    } else {
        None
    }
}

2.7 SCOPE_SHIFT (Unexplained Scope Changes)

Definition: Investigation or analysis scope changes without justification, often to exclude problematic evidence or expand to dilute findings.

Police Detection (Policy File Documentation)

SIO Policy File: Senior Investigating Officer must document rationale for all major decisions
Investigation Strategy Changes: Scope narrowing/broadening requires documented justification
Resource Allocation: Changes to allocated investigative resources must be justified
Example: Investigation initially focuses on single suspect; suddenly expands to "organized crime network" after original suspect's alibi verified (scope expansion without evidence)

Journalism Detection (Editorial Scope Control)

Initial Hypothesis Documentation: Record original investigation scope and research questions
Scope Expansion/Narrowing: Justified by evidence, not convenience
Editor Review: Independent editor checks whether scope changes are evidence-driven
Example: Investigation into CEO's financial fraud expands to include unrelated allegations against industry competitors (scope creep dilutes focus)

Legal Detection (Discovery Disputes and Scope Negotiations)

FRCP Rule 26(f) Meet and Confer: Parties agree on discovery scope
Scope Changes Require Court Approval: Unilateral scope changes contested
Proportionality Review (2015 Amendment): Discovery must be proportional to case needs
Example: Plaintiff requests documents "relating to product defect"; Defendant produces documents about unrelated products (scope expansion to dilute relevant evidence)

Regulatory Detection (Terms of Reference and Allegation Specificity)

Allegation Framing: Initial complaint defines scope
Expansion Triggers: New evidence must justify expanding scope; cannot be exploratory fishing
Real Prospect Test: Each stage re-evaluates whether allegations have real prospect of being proven
Example: Complaint alleges breach of confidentiality in single instance; regulator expands to review practitioner's entire 10-year caseload without justification

Intelligence Detection (Priority Intelligence Requirements - PIRs)

PIR Definition: Clear, specific questions intelligence must answer
Collection Management: Scope changes require PIR revision with leadership approval
Mission Creep Prevention: Structured oversight prevents unfocused intelligence gathering
Example: PIR: "Assess threat from Organization X"; analyst includes unrelated analysis of Organizations Y, Z, W (scope expansion)

Academic Detection (Research Question Evolution)

Bracketing: Phenomenological technique to define study boundaries
Theoretical Saturation: Grounded Theory stops data collection when no new themes emerge (prevents scope expansion)
Audit Trail: Document decisions to include/exclude data
Example: Study initially focused on healthcare workers' burnout; expands to include workplace culture, leadership, organizational structure without justification (scope creep)

Implementation for Phronesis:

struct ScopeShift {
    original_scope: InvestigationScope,
    revised_scope: InvestigationScope,
    shift_type: ScopeShiftType,
    justification: Option<String>,
    triggering_evidence: Option<Vec<String>>,  // Document IDs
    approval_documented: bool,
}

struct InvestigationScope {
    scope_id: String,
    timestamp: DateTime<Utc>,
    entities_in_scope: Vec<Entity>,
    time_period: (DateTime<Utc>, DateTime<Utc>),
    document_types: Vec<String>,
    geographic_scope: Option<String>,
    thematic_focus: Vec<String>,
}

enum ScopeShiftType {
    Expansion { justification_type: JustificationType },
    Narrowing { justification_type: JustificationType },
    Lateral { from_focus: String, to_focus: String },  // Change focus without expanding/narrowing
}

enum JustificationType {
    NewEvidence,           // New evidence justifies scope change
    Operational,           // Resource constraints require narrowing
    Strategic,             // Leadership decision
    Unjustified,          // No documented rationale
}

// Detection algorithm
fn detect_scope_shift(
    investigation: &Investigation,
) -> Option<ScopeShift> {
    let scope_history = investigation.scope_history.as_slice();

    for window in scope_history.windows(2) {
        let (prev, curr) = (&window[0], &window[1]);

        // Check entity scope change
        let entity_expansion = curr.entities_in_scope.len() as f32 / prev.entities_in_scope.len() as f32;
        let time_expansion = (curr.time_period.1 - curr.time_period.0).num_days() as f32
            / (prev.time_period.1 - prev.time_period.0).num_days() as f32;

        if entity_expansion > 2.0 || time_expansion > 2.0 {
            return Some(ScopeShift {
                original_scope: prev.clone(),
                revised_scope: curr.clone(),
                shift_type: ScopeShiftType::Expansion {
                    justification_type: if curr.justification.is_some() {
                        JustificationType::NewEvidence
                    } else {
                        JustificationType::Unjustified
                    }
                },
                justification: curr.justification.clone(),
                triggering_evidence: curr.triggering_evidence.clone(),
                approval_documented: curr.approval_documented,
            });
        }
    }
    None
}

2.8 UNEXPLAINED_CHANGE (Position Changes Without Justification)

Definition: Individual or organization changes position, conclusion, or recommendation without new evidence or stated rationale.

Police Detection (Witness Statement Comparison and Interview Re-Engagement)

Statement Versioning: Compare initial statement to subsequent statements
Re-Interview Protocols: PEACE model used to explore changes ("I noted you said X before, but now you're saying Y. Can you help me understand?")
Motive Assessment: Consider witness motivation (intimidation, coercion, memory decay, truthfulness)
Example: Witness initial statement identifies defendant as perpetrator; second statement recants; investigation reveals defendant's associates visited witness

Journalism Detection (Source Consistency Tracking)

Quote Database: Systematic tracking of source statements over time
On-Record vs. Off-Record Shifts: Document when source changes position publicly vs. privately
Follow-Up Interviews: Explicitly ask source about changed position
Example: Corporate spokesperson initial statement: "We stand by our product's safety"; two weeks later: "We are reviewing certain concerns" (unexplained position change)

Legal Detection (Deposition Comparison and Impeachment)

Deposition vs. Trial Testimony: Compare earlier sworn testimony to trial testimony
Document Production Changes: Identify when party changes position on relevance, privilege, or responsiveness
Impeachment: Cross-examination technique using prior inconsistent statements
Example: Expert witness deposition: "Defect was obvious and foreseeable"; trial testimony: "Defect was subtle and unforeseeable" (impeachment-worthy change)

Regulatory Detection (Submission Comparison and Panel Questioning)

Initial vs. Final Submissions: Compare practitioner's early response to final defence
Panel Questions: Directly question practitioner about changed position
Credibility Assessment: Panel assesses whether change justified or evasive
Example: Practitioner initial response: "No concerns with my practice"; after evidence disclosure: "I acknowledge some shortcomings" (explained change acceptable; unexplained change harmful to credibility)

Intelligence Detection (Assessment Revision Tracking)

Confidence Level Changes: Track when probability estimates shift
Words of Estimative Probability: "Likely" → "Unlikely" without new evidence is red flag
Sensitivity Analysis (ACH Step 7): Document what evidence would have to change for conclusion to flip
Post-Mortem Analysis: When assessments wrong, conduct systematic review to identify failure points
Example: Initial assessment: "Very likely (85%) that regime will survive"; one week later: "Unlikely (30%)" without intervening events (suggests poor initial assessment or new intelligence not documented)

Academic Detection (Researcher Reflexivity and Audit Trails)

Reflexivity Journals: Document researcher's evolving interpretations and changes
Audit Trail Requirements: Track coding decisions; justify changes to coding scheme
Peer Debriefing: Discuss interpretation changes with colleagues
Example: Researcher initially codes theme as "participant empowerment"; later recodes as "participant disempowerment"; audit trail should explain evidence that triggered reinterpretation

Implementation for Phronesis:

struct UnexplainedChange {
    entity: Entity,
    position_history: Vec<Position>,
    change_type: ChangeType,
    explanation: Option<Explanation>,
    new_evidence: Option<Vec<String>>,  // Document IDs
    credibility_impact: CredibilityImpact,
}

struct Position {
    position_id: String,
    timestamp: DateTime<Utc>,
    statement: String,
    document_source: String,
    context: String,
}

enum ChangeType {
    Reversal,         // Complete 180-degree change
    Softening,        // Retreat from strong position
    Hardening,        // Strengthen previously weak position
    Clarification,    // Add nuance without fundamental change
}

struct Explanation {
    explanation_text: String,
    explanation_type: ExplanationType,
    adequacy: ExplanationAdequacy,
}

enum ExplanationType {
    NewEvidence { evidence_ids: Vec<String> },
    MemoryCorrection,
    Misunderstanding,
    ExternalPressure { source: String },
    LegalAdvice,
}

enum ExplanationAdequacy {
    Adequate,       // Explanation consistent with evidence
    Inadequate,     // Explanation weak or inconsistent
    Absent,         // No explanation provided
}

enum CredibilityImpact {
    Severe,         // Undermines reliability of all entity's statements
    Moderate,       // Raises doubts about specific claims
    Minimal,        // Explained change does not harm credibility
}

// Detection algorithm
fn detect_unexplained_change(
    entity: &Entity,
    positions: Vec<Position>,
) -> Option<UnexplainedChange> {
    for window in positions.windows(2) {
        let (prev, curr) = (&window[0], &window[1]);

        // Use semantic similarity to detect position changes
        let similarity = semantic_similarity(&prev.statement, &curr.statement);

        if similarity < 0.3 {  // Significant semantic shift
            return Some(UnexplainedChange {
                entity: entity.clone(),
                position_history: vec![prev.clone(), curr.clone()],
                change_type: classify_change_type(prev, curr),
                explanation: extract_explanation(curr),
                new_evidence: None,  // Check for intervening evidence
                credibility_impact: assess_credibility_impact(prev, curr),
            });
        }
    }
    None
}

3. Evidence Hierarchy Across Domains

3.1 Comparative Evidence Hierarchies

Rank	Police (CPIA)	Journalism (ICIJ/ProPublica)	Legal (FRE)	Regulatory (HCPC/GMC)	Intelligence (Admiralty)	Academic (PRISMA/Cochrane)
1	Physical/Forensic (DNA, fingerprints, digital forensics)	Primary authenticated documents (corporate registries, financial records)	FRE 902(13)/(14) certified digital records with hash	Contemporaneous clinical notes/records	Multi-INT corroborated (A1 rating: reliable source, confirmed info)	Randomized controlled trials (RCTs) with low risk of bias
2	Digital evidence (CCTV, phone records, computer forensics)	Official government documents	Native format documents with metadata	Independent expert reports	Single-INT reliable source (A2/B1 rating)	Systematic reviews and meta-analyses
3	Contemporaneous documents (emails, texts, diary entries)	Authenticated leaked documents (cryptographic verification)	TIFF/PDF with load file metadata	Third-party contemporaneous records	HUMINT from vetted source (B2/C1 rating)	Cohort studies, case-control studies
4	Witness statements (written, video recorded)	Corporate records, meeting minutes	Deposition testimony (sworn, recorded)	Witness statements (colleagues, patients)	OSINT from credible media (C2/D1 rating)	Qualitative studies with high rigor
5	Suspect interviews (PACE compliant, recorded)	Expert analysis and opinions	Live witness testimony (subject to cross-exam)	Practitioner's retrospective account	Single-source HUMINT (unverified)	Expert opinion, case reports
6	Hearsay (when admissible under exceptions)	Anonymous sources (corroborated)	Hearsay (when admissible)	Hearsay, third-party accounts	Rumor, uncorroborated reports	Anecdotes, editorials

3.2 Key Convergence Principles

Principle 1: Contemporaneity Beats Retrospection

Evidence created at time of event > evidence created later
All domains prioritize real-time documentation over memory-based accounts
Police: Contemporaneous notes vs. later statements
Journalism: Original documents vs. reconstructed narratives
Legal: Email sent May 1 > deposition June 1 recalling May 1 events
Regulatory: Clinical notes during session > formal report six months later
Intelligence: SIGINT intercept during event > HUMINT debrief days later
Academic: Contemporaneous field notes > retrospective interviews

Principle 2: Physical/Digital Beats Testimonial

Objective evidence > subjective accounts
Harder to manipulate forensic/digital evidence than witness statements
Police: DNA match > witness identification
Journalism: Bank transaction record > source claiming financial transfer
Legal: Metadata timestamp > witness memory of date
Regulatory: Medical imaging > practitioner's description
Intelligence: Satellite imagery > agent report
Academic: Quantitative outcome measure > self-reported survey

Principle 3: Primary Beats Secondary

Original source > derivative source
Eliminates distortion through intermediaries
Police: Original exhibit > photograph of exhibit
Journalism: Corporate registry > news article citing registry
Legal: Native email file > printed PDF of email
Regulatory: Direct patient interview > family member account
Intelligence: Intercepted communication > translation/summary
Academic: Primary research article > textbook summary

Principle 4: Authenticated Beats Unverified

Chain of custody, hash certification, provenance verification essential
Police: Forensic image with write-blocking > unverified USB drive copy
Journalism: Cryptographically verified leak > anonymous document
Legal: SHA-256 certified production > uncertified file
Regulatory: Medical record from official system > photocopied page
Intelligence: Source with verified access > unknown source
Academic: Peer-reviewed publication > preprint

Principle 5: Corroborated Beats Single-Source

Multiple independent sources > single source (even if reliable)
Reduces impact of individual source errors, biases, or deception
Police: Three independent witnesses > single eyewitness
Journalism: Three-source rule for publication
Legal: Consistent email thread > single email
Regulatory: Multiple complainants reporting same issue > single complaint
Intelligence: Multi-INT fusion (HUMINT + SIGINT + IMINT)
Academic: Triangulation (multiple methods/investigators/data sources)

4. Timeline Construction Methods

4.1 Eight-Step Legal Timeline Process (Detailed)

Source: Legal eDiscovery best practices; used in litigation, internal investigations, regulatory submissions

Step 1: Start Early (Case Inception)

Begin timeline during initial case assessment, not trial preparation
Early timelines drive discovery strategy, identify gaps, guide witness interviews
Phronesis Implementation: Create timeline view from Day 1 of document ingest

Step 2: Create Central Repository

Chronological database or spreadsheet
Required Fields:
- Date/Time (precise or range)
- Event Description (objective facts only, no conclusions)
- Source Document(s) with Bates numbers
- Parties Involved (entities, witnesses)
- Issue Tags (case themes)
- Significance (high/medium/low)
- Notes (attorney work product, not shared)
Phronesis Implementation:

struct TimelineEvent {
    event_id: String,
    timestamp: TimeReference,
    description: String,
    source_docs: Vec<DocumentRef>,
    entities: Vec<Entity>,
    issue_tags: Vec<String>,
    significance: Significance,
    notes: Option<String>,  // Work product
}

Step 3: Identify Parties and Issues

Tag entities: Plaintiff, Defendant, Witness, Decision-maker, Neutral third party
Tag issues: Discrimination, Retaliation, Contract breach, Knowledge, Notice, Damages
Enable filtering: Show only events involving Defendant X and Issue Y
Phronesis Implementation: Entity and issue tagging integrated with S.A.M. engines

Step 4: Collect Documentation

Cast wide net initially (all potentially relevant documents)
Sources:
- Emails, memos, reports, meeting minutes, calendar entries
- Financial records, invoices, contracts
- Text messages, instant messages, social media
- Public records (news articles, regulatory filings, court records)
- Witness interview summaries, deposition transcripts
Phronesis Implementation: Documentary engine processes all document types

Step 5: Extract Events (Date-Specific Facts)

Event Definition: Fact tied to specific date/time
Good Examples:
- "2024-03-15: Smith sent email to Jones re: 'quarterly projections' (ABC00123)"
- "2024-04-01: Board meeting approved merger (Board Minutes, ABC00456, p. 12)"
- "2024-05-10: Plaintiff filed EEOC charge (Charge No. 123-456)"
NOT Events: Generalized conclusions without date ("Smith was aware of defect")
Phronesis Implementation: Temporal engine extracts date-stamped assertions

Step 6: Use Objective Language (Facts Only)

Good: "Jones emailed Smith: 'The brake defect was known to engineering team since January' (ABC00145)"
Bad: "Jones admitted knowledge of defect" (argumentative, conclusory)
Rationale: Timelines shared with opposing counsel, judges, juries—must be neutral
Phronesis Implementation: Automated flagging of subjective language; suggest neutral alternatives

Step 7: Link Evidence (Every Event Cites Supporting Documents)

Bates Numbers: ABC00001-ABC00010 (production reference)
Deposition Cites: Smith Dep. 45:12-18 (transcript page:line)
Public Records: SEC Form 10-K (2024) at p. 23
Purpose: Instant verification, admissibility support, cross-examination facilitation
Phronesis Implementation:

struct DocumentRef {
    doc_id: String,
    bates_number: Option<String>,
    page: Option<u32>,
    paragraph: Option<u32>,
    line_range: Option<(u32, u32)>,
}

Step 8: Update Continuously

Timeline is living document throughout case lifecycle
Add events as new discovery received
Revise entries as deposition testimony clarifies facts
Tag events with relevance to motions, trial
Phronesis Implementation: Real-time timeline updates as documents added; version control

Legal Timeline Analysis Techniques:

Temporal Gap Analysis:

Identify suspicious time periods with no documentation
Potential document destruction (spoliation)
Example: Email trail stops abruptly after litigation hold issued

Temporal Density Analysis:

Spikes in activity indicate critical periods
Example: 200 emails in 48 hours before product recall

Temporal Contradiction Detection:

Compare event sequence to witness testimony
Example: Witness: "First learned May 1" but timeline shows email April 15

Critical Path Analysis:

Identify sequence leading to harm
Example: Design → Manufacturing → Distribution → First Complaint → Recall

Parallel Track Analysis:

Compare simultaneous events in different domains
Example: Acquisition timeline vs. insider trading timeline (prove knowledge)

4.2 Police 5WH Framework for Timeline Construction

Source: College of Policing Core Investigative Doctrine; HOLMES2 case management

Six Dimensions:

What happened? (Event reconstruction, offence classification)
Who was involved? (Victims, suspects, witnesses, accomplices)
When did it occur? (Timeline establishment, sequence reconstruction)
Where did it take place? (Crime scene, secondary locations, digital spaces)
Why did it happen? (Motivation, context, contributory factors)
How was it committed? (Methods, tools, techniques, modus operandi)

Timeline Construction Process:

Initial Report Chronology (CAD - Computer Aided Dispatch)
- Time of call, first officer arrival, scene preservation
Victim/Witness Account Extraction
- PEACE interview: Free narrative → extract date/time references
- Cognitive Interview: Context reinstatement (physical environment, emotional state)
Forensic Timeline Integration
- Digital forensics: File creation/modification timestamps (write-blocked imaging)
- Physical forensics: Entomology (time of death), blood spatter analysis (sequence of events)
- Communications: Cell site analysis (phone location at specific times)
CCTV Timeline Mapping
- Frame-by-frame analysis for exact event sequencing
- Geographic tracking: Suspect/victim movements
HOLMES2 Dynamic Reasoning
- Automated timeline construction from actions, statements, exhibits
- Anomaly detection: Timeline gaps, contradictions
- Relationship mapping over time
Policy File Documentation
- SIO rationale for timeline interpretation
- Alternative explanations considered
- Evidence gaps and mitigation strategies

Phronesis Integration:

Import HOLMES2-style action logs
5WH tagging system for filtering timeline
Anomaly detection engine (gaps, density spikes)

4.3 Intelligence Chronologies and F3EAD Cycle

Source: US Joint Chiefs of Staff JP 2-0; JSOC operational intelligence

Intelligence Chronology Components:

Collection Timeline:

Track when intelligence was collected (not when event occurred)
Critical for assessing freshness and gaps
Example: HUMINT report dated May 10 describes events of April 15 (25-day lag)

Event Timeline:

Reconstruct adversary actions using multi-INT fusion
Levels:
- Strategic: Long-term developments (months to years)
- Operational: Campaign phases (weeks to months)
- Tactical: Specific actions (hours to days)

Decision Timeline:

Track leadership decisions and intelligence that informed them
Post-mortem analysis: What did we know, when did we know it?

F3EAD Operational Cycle (Rapid Timeline):

Find: Develop target intelligence (days to weeks)
Fix: Confirm target location (hours to days)
Finish: Execute operation (minutes to hours)
Exploit: Immediate exploitation of captured materials (minutes to hours)
Analyze: Deep analysis (hours to days)
Disseminate: Share intelligence (minutes to hours)
Cycle Time: Hours to days (vs. weeks/months for strategic intelligence)

Critical Temporal Analysis:

Time-Distance Analysis: Can entity be at Location A at Time 1 and Location B at Time 2?
Communication Sequencing: SIGINT intercepts before/after specific events
Pattern-of-Life Analysis: Baseline behavior vs. anomalous behavior

Phronesis Integration:

F3EAD-style rapid timeline for urgent safeguarding cases
Collection timestamp vs. event timestamp tracking
Multi-INT fusion timeline (combine document types)

4.4 Journalism ChronoFact Verification

Source: ICIJ, ProPublica, OCCRP methodologies

ChronoFact Framework:

Stage 1: Initial Chronology Construction

Extract all date references from primary documents
Build preliminary timeline (may contain contradictions)

Stage 2: Verification Protocol

5WH Questions for Every Event:
- What: Specific event description
- Who: Entities involved
- When: Precise timestamp (or best approximation)
- Where: Geographic location
- Why: Contextual explanation (where available)
- How: Mechanism (where relevant)

Stage 3: Cross-Reference with External Sources

Government databases (corporate registries, court records)
News archives (dated articles)
Financial records (timestamped transactions)

Stage 4: Contradiction Resolution

When sources disagree on dates:
1. Prioritize contemporaneous documents over retrospective accounts
2. Prioritize official records over unofficial sources
3. Seek third independent source to break tie
4. If unresolvable, note contradiction in published report

Stage 5: Publication Standard

Minimum three independent sources confirming key dates
Uncertainty explicitly acknowledged ("records show transaction occurred between March 15-20")
Timeline visualization with confidence indicators

Phronesis Integration:

ChronoFact-style verification workflow
Confidence scoring for temporal assertions
Multi-source temporal contradiction detection

4.5 Academic Longitudinal Timeline Analysis

Source: Grounded Theory, Case Study Research, Process Tracing

Academic Timeline Methods:

Within-Case Temporal Sequencing:

Track participant experiences across multiple interview waves
Identify turning points, critical incidents, process stages

Process Tracing (Qualitative):

Establish causal sequence: Event A → Event B → Outcome C
Test alternative causal pathways
Distinguish correlation from causation through temporal precedence

Retrospective vs. Contemporaneous Comparison:

Compare participant's retrospective narrative to documents created in real-time
Detect memory reconstruction, hindsight bias
Example: Participant describes "gradual decline" but contemporaneous notes show sudden event

Longitudinal Coding:

Code same data at multiple time points
Detect theme evolution, stability, emergence

Phronesis Integration:

Support retrospective account vs. contemporaneous record comparison
Process tracing visualization (Event A → B → C with evidence linking)
Longitudinal coding support (track coding decisions over time)

4.6 Regulatory Investigation Timelines

Source: HCPC, GMC, NMC fitness-to-practice procedures

Regulatory Timeline Requirements:

Incident Timeline:

Precise chronology of alleged misconduct
Contemporaneous records prioritized (clinical notes, appointment systems)

Practitioner Response Timeline:

Track when practitioner made aware of concerns
Opportunity to respond at each stage (procedural fairness)

Investigation Timeline:

Stage 1 (Screening): Days to weeks
Stage 2 (Investigation): Weeks to months
Stage 3 (Hearing): Months to years

Contemporaneous Records Analysis:

Compare practitioner's clinical notes (created during incident) to formal submissions (created months later)
Temporal gaps in documentation as red flag

Phronesis Integration:

Three-stage timeline tracking (allegation → investigation → hearing)
Contemporaneous vs. retrospective account comparison
Procedural fairness timeline (notifications, responses, decisions)

5. Quality Control Convergence

5.1 Minimum Three Reviewers Standard

Universal Finding: All rigorous investigation methodologies require minimum three independent reviewers for reliable quality control.

Police (IOPC, Major Incidents):

SIO (Senior Investigating Officer)
Deputy SIO
Independent review by external investigator (for critical incidents)
IOPC oversight (sensitive cases)

Journalism (ICIJ, ProPublica, OCCRP):

Primary journalist/investigator
Editorial team (minimum 2-3 editors)
Fact-checker (independent verification)
Legal review (before publication)
Guardian Example: 3-tier review (reporter → editor → legal → publication)

Legal (TAR Validation, Privilege Review):

First-pass review (trained reviewers)
Senior attorney review (QC sample 5-10%)
Partner/senior counsel review (QC sample of QC)
Validation Testing: Statistical sampling for TAR (precision/recall calculation)

Regulatory (Dual Decision-Maker Panels):

Professional member (subject-matter expert)
Lay member (public protection perspective)
Legal assessor (procedural fairness)
Typical Panel: 2-3 panel members + legal assessor

Intelligence (Peer Review and Red Cell):

Primary analyst
Peer review (tradecraft check)
Management review (policy implications)
Red Cell review (alternative analysis)
ICD 203: Minimum 3 reviewers for high-confidence assessments

Academic (Inter-Rater Reliability):

Minimum 2 independent coders (often 3+)
Statistical reliability: Cohen's Kappa ≥ 0.70 (substantial agreement)
Disagreement resolution: Consensus discussion or third coder
PRISMA: Two reviewers independently screen articles; disagreements resolved by third reviewer

Phronesis Implementation:

struct QualityControlWorkflow {
    investigation_id: String,
    reviewers: Vec<Reviewer>,
    review_stages: Vec<ReviewStage>,
    inter_rater_reliability: Option<InterRaterReliability>,
    final_consensus: Option<Consensus>,
}

struct Reviewer {
    reviewer_id: String,
    reviewer_type: ReviewerType,
    expertise: Vec<String>,
}

enum ReviewerType {
    Primary,           // Primary investigator
    PeerReview,        // Same-level colleague
    SeniorReview,      // Supervisor/expert
    RedCell,           // Adversarial perspective
    LegalAssessor,     // Procedural compliance
    LayReviewer,       // Public protection perspective
}

struct ReviewStage {
    stage_id: String,
    reviewer_id: String,
    timestamp: DateTime<Utc>,
    findings: Vec<Finding>,
    recommendations: Vec<String>,
    flags: Vec<QCFlag>,
}

enum QCFlag {
    ContradictionUnresolved,
    EvidenceGap,
    MethodologyDeviation,
    BiasDetected(BiasType),
    ConfidenceLevelConcern,
}

struct InterRaterReliability {
    coders: Vec<String>,
    method: IRRMethod,
    score: f32,
    acceptable: bool,  // ≥ 0.70 for Cohen's Kappa
}

enum IRRMethod {
    CohenKappa,
    FleissKappa,        // 3+ raters
    PercentAgreement,
    IntraclassCorrelation,
}

5.2 Statistical Quality Control (Where Applicable)

Legal TAR Validation Testing:

Precision: Proportion of produced documents actually relevant (50-70% typical target)
Recall: Proportion of relevant documents successfully identified (75%+ typical target)
Validation Methods:
- Random sampling (statistical confidence intervals)
- Stratified sampling (test across relevance score ranges)
- Elusion testing (prove little relevant material missed in low-scoring documents)
Acceptable Performance: Varies by jurisdiction; typically recall ≥75%, precision ≥50%

Academic Inter-Rater Reliability:

Cohen's Kappa: Two raters
- < 0.40: Poor agreement
- 0.40-0.59: Fair agreement
- 0.60-0.79: Substantial agreement
- ≥ 0.80: Almost perfect agreement
- Standard: ≥ 0.70 acceptable for most research
Fleiss' Kappa: Three or more raters
Intraclass Correlation Coefficient (ICC): Continuous variables
Percent Agreement: Simple but doesn't account for chance

Intelligence Confidence Calibration:

Track Record Analysis: Compare analyst's confidence levels to actual outcomes
Well-Calibrated Analyst: When they say "70% confident," they're right ~70% of the time
Overconfident Analyst: Says "90% confident" but only right 60% of the time
Performance Management: Provide feedback to analysts on calibration

Phronesis Implementation:

struct ValidationMetrics {
    validation_id: String,
    validation_type: ValidationType,
    sample_size: u32,
    metrics: ValidationResults,
}

enum ValidationType {
    TAR { method: TARMethod },
    InterRaterReliability { method: IRRMethod },
    ConfidenceCalibration,
}

struct ValidationResults {
    precision: Option<f32>,
    recall: Option<f32>,
    f1_score: Option<f32>,
    cohen_kappa: Option<f32>,
    percent_agreement: Option<f32>,
    confidence_intervals: Option<(f32, f32)>,
    acceptable: bool,
}

6. Bias Mitigation Strategies

6.1 Intelligence SATs (Structured Analytic Techniques)

Source: Heuer & Pherson, Structured Analytic Techniques for Intelligence Analysis (2021); 66 techniques across 8 categories

Key Anti-Bias Techniques:

Analysis of Competing Hypotheses (ACH):

Force consideration of alternative explanations
Disprove hypotheses rather than confirm
Mitigation: Confirmation bias

Devil's Advocacy:

Deliberately argue against consensus view
Institutionalized role (CIA Red Cell)
Mitigation: Groupthink

Key Assumptions Check:

Identify and challenge foundational assumptions
Test: "What would have to be true for this assumption to fail?"
Mitigation: Anchoring, hidden assumptions

Red Cell Analysis:

Adversarial perspective (How would enemy exploit our vulnerabilities?)
Separate team with protected dissent
Mitigation: Mirror imaging, overconfidence

Pre-Mortem Analysis:

Assume failure occurred; work backwards to explain why
Identify vulnerabilities before they manifest
Mitigation: Sunk cost fallacy, overconfidence

Quality of Information Check:

Rate source reliability and information credibility independently (Admiralty Code)
Identify gaps, contradictions, deception indicators
Mitigation: Availability heuristic, source bias

Phronesis Implementation:

ACH matrix builder (Section 2)
Devil's Advocacy assignment in QC workflow
Assumptions documentation requirement
Pre-mortem template for high-stakes investigations

6.2 Legal TAR Validation and Sampling

Source: Legal eDiscovery TAR 1.0/2.0 methodologies; court-validated protocols

Validation Methods:

Seed Set Diversity (TAR 1.0):

Random sample (1,000-2,000 documents) for training
Ensures representative coverage of document population
Mitigation: Sample bias, underrepresentation of rare but critical documents

Continuous Active Learning (TAR 2.0):

No fixed seed set; model learns continuously from reviewer coding
Adapts to complex, evolving issues
Mitigation: Confirmation bias (model prioritizes diverse documents, not just confirming examples)

Elusion Testing:

Sample low-scoring documents to prove minimal relevant material missed
Statistical confidence: "95% confident that <5% relevant documents remain in low-scoring set"
Mitigation: False negatives (relevant documents incorrectly classified as non-responsive)

QC Sampling (Privilege Review):

5-10% random sample of privilege designations reviewed by senior attorney
Error rate >5% triggers retraining and re-review
Mitigation: Reviewer inconsistency, privilege over-designation

Phronesis Implementation:

struct BiasDetection {
    detection_id: String,
    bias_type: BiasType,
    evidence: Vec<String>,
    severity: BiasSeverity,
    mitigation_applied: Option<MitigationStrategy>,
}

enum BiasType {
    ConfirmationBias,      // Seeking evidence that confirms hypothesis
    SelectionBias,         // Cherry-picking evidence
    AnchoringBias,         // Over-reliance on first information
    Groupthink,            // Pressure to conform
    MirrorImaging,         // Assuming others think like you
    AvailabilityHeuristic, // Overweighting easily recalled info
    SunkCostFallacy,       // Continuing failed course due to prior investment
}

enum BiasSeverity {
    Critical,   // Undermines investigation integrity
    Significant, // Substantial impact on findings
    Moderate,   // Detectable but contained
    Minimal,    // Identified and mitigated
}

struct MitigationStrategy {
    strategy: String,
    applied_by: String,
    timestamp: DateTime<Utc>,
    effectiveness: Option<EffectivenessRating>,
}

6.3 Academic Reflexivity and Member Checking

Source: Qualitative research traditions (Lincoln & Guba, Naturalistic Inquiry; Creswell, Qualitative Inquiry)

Reflexivity Practices:

Researcher Positioning Statement:

Explicit statement of researcher's background, assumptions, potential biases
Published with research to enable reader assessment
Example: "As a former healthcare administrator, I bring both insider knowledge and potential bias toward institutional perspectives."

Reflexivity Journal:

Ongoing documentation of researcher's reactions, interpretations, decisions
Track evolution of thinking
Identify moments when personal bias may influence interpretation

Peer Debriefing:

Regular discussions with colleagues uninvolved in project
Challenge interpretations, identify blind spots
Mitigation: Lone researcher bias, interpretation drift

Member Checking:

Return findings to participants for validation
Two Forms:
1. Transcript Review: Participants verify accuracy of interview transcript
2. Interpretation Review: Participants comment on researcher's interpretations
Critical Debate: Some argue member checking problematic (participants may disagree with accurate but unflattering interpretations)

Negative Case Analysis:

Actively seek evidence contradicting emerging themes
Test rival explanations
Grounded Theory: Constant comparison until theoretical saturation (no new disconfirming evidence)

Audit Trail:

Document all coding decisions, changes, rationale
Enable external auditor to assess dependability
Components: Raw data, coding decisions, analysis products, process notes

Phronesis Implementation:

Reflexivity journal integrated into investigation workspace
Member checking workflow (share findings with involved parties for comment)
Negative case analysis prompts (system suggests contradictory evidence)
Audit trail export for appeals/external review

6.4 Journalism Hypothesis-Based Framework and Editorial Review

Source: ICIJ, ProPublica, OCCRP methodologies

Hypothesis-Based Investigation:

Separate Facts from Assumptions:

Investigation Hypothesis: Company X evaded taxes through offshore shell companies
├── Known Facts (documented, verified)
│   ├── Corporate registry shows Company X owns Entity Y
│   ├── Entity Y registered in tax haven jurisdiction
│   └── Financial records show payments to Entity Y
└── Assumptions (requires verification)
    ├── Payments to Entity Y constitute tax evasion (ASSUMPTION: requires legal analysis)
    ├── Company X executives were aware of Entity Y's purpose (ASSUMPTION: requires evidence)
    └── Entity Y has no legitimate business purpose (ASSUMPTION: requires investigation)

Iterative Verification:

Test each assumption systematically
Convert assumptions to facts or disprove
Adjust hypothesis as evidence accumulates

Editorial Review Process:

Stage 1: Reporter Self-Review:

Checklist: Have I verified key facts? Do I have three sources? Have I contacted subjects for comment?

Stage 2: Editor Review:

Challenge methodology, sources, interpretations
Identify gaps, alternative explanations, potential biases

Stage 3: Fact-Checker Review:

Independent verification of all factual claims
Check dates, numbers, quotes, document citations
Re-contact sources to confirm quotes

Stage 4: Legal Review:

Assess defamation risk
Verify right-of-reply compliance
Check for contempt of court, privacy violations

Stage 5: Ethics Review (High-Sensitivity Stories):

Independent ethics panel reviews impact on vulnerable sources
Public interest justification for potentially harmful disclosures

Phronesis Implementation:

Hypothesis tree visualization (facts vs. assumptions)
Multi-stage review workflow (reporter → editor → fact-check → legal → ethics)
Right-of-reply tracking (ensure subjects given opportunity to respond)

7. Unified Framework for Phronesis FCIP

7.1 Core Architecture Principles

Principle 1: Domain-Agnostic Core with Domain-Specific Adapters

Core Components (universal across domains):

Evidence ingestion and authentication
Entity extraction and resolution
Timeline construction and visualization
Eight S.A.M. contradiction types detection
Quality control workflows (minimum 3 reviewers)
Audit trail and provenance tracking

Domain Adapters:

Police Adapter: HOLMES2 import/export, PACE compliance, CPIA disclosure
Journalism Adapter: ICIJ Aleph integration, ChronoFact verification, right-of-reply
Legal Adapter: TAR workflow, Bates numbering, privilege review, FRE compliance
Regulatory Adapter: Three-stage gatekeeping, dual panels, contemporaneous records priority
Intelligence Adapter: Admiralty Code rating, ACH matrix, Words of Estimative Probability
Academic Adapter: PRISMA workflow, inter-rater reliability, reflexivity journal

Principle 2: Multi-Level Evidence Hierarchy

All evidence items receive three independent ratings:

Provenance Score (0.0-1.0)
- 1.0 = Authenticated with hash certification, clear chain of custody
- 0.5 = Verified source but no cryptographic authentication
- 0.0 = Unknown provenance
Temporal Score (0.0-1.0)
- 1.0 = Contemporaneous (created at time of event)
- 0.5 = Retrospective but verifiable timestamp
- 0.0 = Undated or date disputed
Corroboration Score (0.0-1.0)
- 1.0 = Corroborated by 3+ independent sources
- 0.5 = Corroborated by 1-2 sources
- 0.0 = Single source, uncorroborated

Composite Evidence Quality = (Provenance × 0.4) + (Temporal × 0.3) + (Corroboration × 0.3)

Principle 3: Contradiction Detection as Central Engine

All eight S.A.M. contradiction types implemented as Rust modules:

pub mod contradiction {
    pub mod self_contradiction;
    pub mod inter_doc;
    pub mod temporal;
    pub mod evidentiary;
    pub mod modality_shift;
    pub mod selective_citation;
    pub mod scope_shift;
    pub mod unexplained_change;

    pub struct ContradictionEngine {
        detectors: Vec<Box<dyn ContradictionDetector>>,
        confidence_threshold: f32,
    }

    pub trait ContradictionDetector: Send + Sync {
        fn detect(&self, context: &InvestigationContext) -> Vec<Contradiction>;
        fn contradiction_type(&self) -> ContradictionType;
        fn confidence_score(&self, contradiction: &Contradiction) -> f32;
    }
}

Principle 4: Quality Control Gates

Investigations cannot progress to next stage without QC approval:

Stage 1: Evidence Collection

QC Gate: Chain of custody documented, authentication completed
Reviewers: Primary investigator + QC auditor

Stage 2: Initial Analysis

QC Gate: All eight contradiction types checked, evidence hierarchy applied
Reviewers: Primary investigator + peer reviewer

Stage 3: Findings

QC Gate: Minimum 3 reviewers approval, inter-rater reliability ≥0.70 (where applicable)
Reviewers: Primary + peer + senior + Red Cell (for high-stakes)

Stage 4: Report

QC Gate: Audit trail complete, procedural fairness verified
Reviewers: Primary + legal assessor + domain expert

Principle 5: Transparent Audit Trail

Every action logged with:

User ID
Timestamp
Action type (document added, contradiction detected, finding coded, etc.)
Rationale (free text)
Evidence links (document IDs)
Confidence level

Audit trail exportable for:

Appeals
Judicial review
Regulatory oversight
Academic publication

7.2 Implementation Architecture

Technology Stack:

Backend: Rust (Tauri 2.9, SQLite via sqlx)
Frontend: React 18, React Router 7, Tailwind CSS
AI: Vercel AI SDK (multi-provider: Claude, Groq, Gemini)
State: Zustand, TanStack Query
Document Processing: pdf-extract (Rust), Docker Python tools for OCR
Analytics: S.A.M. engines (11 engines in Rust)

Core Modules:

// src-tauri/src/engines/contradiction/mod.rs

use crate::db::Database;
use serde::{Serialize, Deserialize};
use chrono::{DateTime, Utc};

pub struct ContradictionEngine {
    db: Arc<RwLock<Database>>,
    detectors: Vec<Box<dyn ContradictionDetector>>,
    config: ContradictionConfig,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Contradiction {
    pub contradiction_id: String,
    pub contradiction_type: ContradictionType,
    pub evidence: Vec<EvidenceRef>,
    pub description: String,
    pub confidence: f32,
    pub severity: ContradictionSeverity,
    pub detected_at: DateTime<Utc>,
    pub detected_by: DetectionMethod,
    pub resolution_status: ResolutionStatus,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum ContradictionType {
    Self { doc_id: String },
    InterDoc { doc_ids: Vec<String> },
    Temporal { timeline_conflict: TemporalConflict },
    Evidentiary { claim: String, evidence: Vec<String> },
    ModalityShift { statement_versions: Vec<String> },
    SelectiveCitation { cited: Vec<String>, uncited: Vec<String> },
    ScopeShift { original: String, revised: String },
    UnexplainedChange { position_history: Vec<String> },
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum DetectionMethod {
    NLP { model: String, confidence: f32 },
    Timeline { algorithm: String },
    HumanReview { reviewer_id: String },
    Hybrid { methods: Vec<DetectionMethod> },
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum ResolutionStatus {
    Unresolved,
    UnderReview { reviewer_id: String, since: DateTime<Utc> },
    Resolved { resolution: ResolutionType, by: String, at: DateTime<Utc> },
    Disputed { dispute_reason: String, by: String },
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub enum ResolutionType {
    Confirmed { explanation: String },
    Reconciled { explanation: String },
    Dismissed { reason: String },
    Unresolvable { acknowledged: bool },
}

impl ContradictionEngine {
    pub async fn detect_all(&self, case_id: &str) -> Result<Vec<Contradiction>, Error> {
        let mut all_contradictions = Vec::new();

        for detector in &self.detectors {
            let context = self.build_context(case_id).await?;
            let contradictions = detector.detect(&context);
            all_contradictions.extend(contradictions);
        }

        // Deduplicate and rank by confidence
        self.deduplicate_and_rank(all_contradictions)
    }

    pub async fn validate_findings(
        &self,
        contradictions: Vec<Contradiction>,
        reviewers: Vec<String>,
    ) -> Result<ValidationReport, Error> {
        // Assign to minimum 3 reviewers
        if reviewers.len() < 3 {
            return Err(Error::InsufficientReviewers);
        }

        // Calculate inter-rater reliability
        let irr = self.calculate_inter_rater_reliability(&contradictions, &reviewers).await?;

        // Require Cohen's Kappa ≥ 0.70
        if irr.kappa < 0.70 {
            return Err(Error::InsufficientAgreement { kappa: irr.kappa });
        }

        Ok(ValidationReport {
            contradictions,
            reviewers,
            inter_rater_reliability: irr,
            validation_timestamp: Utc::now(),
            approved: true,
        })
    }
}

8. Implementation Architecture

8.1 Database Schema Extensions

Contradictions Table:

CREATE TABLE contradictions (
    contradiction_id TEXT PRIMARY KEY,
    case_id TEXT NOT NULL,
    contradiction_type TEXT NOT NULL,  -- SELF, INTER_DOC, TEMPORAL, etc.
    description TEXT NOT NULL,
    confidence REAL NOT NULL CHECK(confidence >= 0.0 AND confidence <= 1.0),
    severity TEXT NOT NULL CHECK(severity IN ('CRITICAL', 'SIGNIFICANT', 'MODERATE', 'MINIMAL')),
    detected_at TEXT NOT NULL,
    detected_by TEXT NOT NULL,  -- JSON: {method: "NLP", user_id: "...", model: "..."}
    resolution_status TEXT NOT NULL CHECK(resolution_status IN ('UNRESOLVED', 'UNDER_REVIEW', 'RESOLVED', 'DISPUTED')),
    resolution_details TEXT,  -- JSON
    created_at TEXT NOT NULL DEFAULT (datetime('now')),
    updated_at TEXT NOT NULL DEFAULT (datetime('now')),
    FOREIGN KEY (case_id) REFERENCES cases(case_id)
);

CREATE INDEX idx_contradictions_case ON contradictions(case_id);
CREATE INDEX idx_contradictions_type ON contradictions(contradiction_type);
CREATE INDEX idx_contradictions_status ON contradictions(resolution_status);

Contradiction Evidence Linking:

CREATE TABLE contradiction_evidence (
    contradiction_id TEXT NOT NULL,
    evidence_id TEXT NOT NULL,
    evidence_type TEXT NOT NULL CHECK(evidence_type IN ('DOCUMENT', 'TIMELINE_EVENT', 'ENTITY', 'CLAIM')),
    role TEXT NOT NULL CHECK(role IN ('PRIMARY', 'SUPPORTING', 'CONTRADICTING')),
    page_number INTEGER,
    paragraph_number INTEGER,
    text_span TEXT,  -- JSON: {start: 123, end: 456}
    created_at TEXT NOT NULL DEFAULT (datetime('now')),
    PRIMARY KEY (contradiction_id, evidence_id, evidence_type),
    FOREIGN KEY (contradiction_id) REFERENCES contradictions(contradiction_id)
);

Quality Control Reviews:

CREATE TABLE qc_reviews (
    review_id TEXT PRIMARY KEY,
    case_id TEXT NOT NULL,
    stage TEXT NOT NULL CHECK(stage IN ('EVIDENCE_COLLECTION', 'INITIAL_ANALYSIS', 'FINDINGS', 'REPORT')),
    reviewer_id TEXT NOT NULL,
    reviewer_type TEXT NOT NULL CHECK(reviewer_type IN ('PRIMARY', 'PEER', 'SENIOR', 'RED_CELL', 'LEGAL', 'LAY')),
    findings TEXT NOT NULL,  -- JSON array
    recommendations TEXT,
    flags TEXT,  -- JSON array of QCFlag
    approval_status TEXT NOT NULL CHECK(approval_status IN ('APPROVED', 'REJECTED', 'REQUIRES_REVISION')),
    reviewed_at TEXT NOT NULL DEFAULT (datetime('now')),
    FOREIGN KEY (case_id) REFERENCES cases(case_id)
);

CREATE INDEX idx_qc_case_stage ON qc_reviews(case_id, stage);

Inter-Rater Reliability Tracking:

CREATE TABLE irr_sessions (
    session_id TEXT PRIMARY KEY,
    case_id TEXT NOT NULL,
    coders TEXT NOT NULL,  -- JSON array of reviewer IDs
    method TEXT NOT NULL CHECK(method IN ('COHEN_KAPPA', 'FLEISS_KAPPA', 'PERCENT_AGREEMENT', 'ICC')),
    score REAL NOT NULL,
    acceptable BOOLEAN NOT NULL,  -- TRUE if score ≥ threshold
    threshold REAL NOT NULL DEFAULT 0.70,
    conducted_at TEXT NOT NULL DEFAULT (datetime('now')),
    FOREIGN KEY (case_id) REFERENCES cases(case_id)
);

8.2 TypeScript Integration (CONTRACT.ts)

// src/CONTRACT.ts

export interface Contradiction {
  contradiction_id: string
  case_id: string
  contradiction_type: ContradictionType
  description: string
  confidence: number // 0.0-1.0
  severity: ContradictionSeverity
  detected_at: string // ISO 8601
  detected_by: DetectionMethod
  resolution_status: ResolutionStatus
  resolution_details?: ResolutionDetails
}

export type ContradictionType =
  | { type: 'SELF'; doc_id: string }
  | { type: 'INTER_DOC'; doc_ids: string[] }
  | { type: 'TEMPORAL'; timeline_conflict: TemporalConflict }
  | { type: 'EVIDENTIARY'; claim: string; evidence: string[] }
  | { type: 'MODALITY_SHIFT'; statement_versions: string[] }
  | { type: 'SELECTIVE_CITATION'; cited: string[]; uncited: string[] }
  | { type: 'SCOPE_SHIFT'; original: string; revised: string }
  | { type: 'UNEXPLAINED_CHANGE'; position_history: string[] }

export type ContradictionSeverity = 'CRITICAL' | 'SIGNIFICANT' | 'MODERATE' | 'MINIMAL'

export interface DetectionMethod {
  method: 'NLP' | 'TIMELINE' | 'HUMAN_REVIEW' | 'HYBRID'
  details: {
    model?: string
    confidence?: number
    reviewer_id?: string
    algorithm?: string
    methods?: DetectionMethod[]
  }
}

export type ResolutionStatus =
  | { status: 'UNRESOLVED' }
  | { status: 'UNDER_REVIEW'; reviewer_id: string; since: string }
  | { status: 'RESOLVED'; resolution: ResolutionType; by: string; at: string }
  | { status: 'DISPUTED'; dispute_reason: string; by: string }

export interface QCReview {
  review_id: string
  case_id: string
  stage: QCStage
  reviewer_id: string
  reviewer_type: ReviewerType
  findings: Finding[]
  recommendations?: string[]
  flags?: QCFlag[]
  approval_status: 'APPROVED' | 'REJECTED' | 'REQUIRES_REVISION'
  reviewed_at: string
}

export type QCStage = 'EVIDENCE_COLLECTION' | 'INITIAL_ANALYSIS' | 'FINDINGS' | 'REPORT'

export type ReviewerType = 'PRIMARY' | 'PEER' | 'SENIOR' | 'RED_CELL' | 'LEGAL' | 'LAY'

export interface IRRSession {
  session_id: string
  case_id: string
  coders: string[] // Reviewer IDs
  method: 'COHEN_KAPPA' | 'FLEISS_KAPPA' | 'PERCENT_AGREEMENT' | 'ICC'
  score: number
  acceptable: boolean // true if score ≥ threshold
  threshold: number // Default 0.70
  conducted_at: string
}

8.3 React Components (UI Implementation)

Contradiction Dashboard:

// src/components/analysis/contradiction-dashboard.tsx

import React from 'react';
import { useContradictions } from '@/hooks/use-contradictions';
import { ContradictionCard } from './contradiction-card';
import { ContradictionFilter } from './contradiction-filter';

export function ContradictionDashboard({ caseId }: { caseId: string }) {
  const { data: contradictions, isLoading } = useContradictions(caseId);
  const [filter, setFilter] = React.useState<ContradictionFilter>({
    types: [],
    severity: [],
    status: [],
  });

  const filteredContradictions = React.useMemo(() => {
    return contradictions?.filter(c => {
      if (filter.types.length > 0 && !filter.types.includes(c.contradiction_type.type)) {
        return false;
      }
      if (filter.severity.length > 0 && !filter.severity.includes(c.severity)) {
        return false;
      }
      if (filter.status.length > 0 && !filter.status.includes(c.resolution_status.status)) {
        return false;
      }
      return true;
    });
  }, [contradictions, filter]);

  return (
    <div className="contradiction-dashboard">
      <ContradictionFilter filter={filter} onChange={setFilter} />

      <div className="contradiction-grid">
        {filteredContradictions?.map(contradiction => (
          <ContradictionCard key={contradiction.contradiction_id} contradiction={contradiction} />
        ))}
      </div>
    </div>
  );
}

Timeline View with Contradiction Highlighting:

// src/components/analysis/timeline-view.tsx

import React from 'react';
import { useTimeline } from '@/hooks/use-timeline';
import { useContradictions } from '@/hooks/use-contradictions';

export function TimelineView({ caseId }: { caseId: string }) {
  const { data: events } = useTimeline(caseId);
  const { data: contradictions } = useContradictions(caseId);

  const temporalContradictions = React.useMemo(() => {
    return contradictions?.filter(c => c.contradiction_type.type === 'TEMPORAL') || [];
  }, [contradictions]);

  return (
    <div className="timeline-view">
      {events?.map(event => (
        <TimelineEvent
          key={event.event_id}
          event={event}
          contradictions={temporalContradictions.filter(c =>
            c.contradiction_type.timeline_conflict.event_ids.includes(event.event_id)
          )}
        />
      ))}
    </div>
  );
}

9. Sources

Police Investigations

College of Policing. (2023). Core Investigative Doctrine. https://www.college.police.uk/app/investigation/core-investigative-doctrine
Home Office. (1996). Criminal Procedure and Investigations Act 1996 (CPIA). https://www.legislation.gov.uk/ukpga/1996/25/contents
Fisher, R. P. & Geiselman, R. E. (1992). Memory-Enhancing Techniques for Investigative Interviewing: The Cognitive Interview. Charles C. Thomas.

Journalism Investigations

International Consortium of Investigative Journalists (ICIJ). Panama Papers Methodology. https://www.icij.org/investigations/panama-papers/
ProPublica. Reporting Standards and Practices. https://www.propublica.org/about/reporting-standards
Organized Crime and Corruption Reporting Project (OCCRP). Aleph Investigation Platform. https://aleph.occrp.org/

Legal eDiscovery

EDRM Project. Electronic Discovery Reference Model. https://edrm.net/
Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012) [TAR 1.0 court approval]
Rio Tinto PLC v. Vale S.A., 306 F.R.D. 125 (D. Del. 2015) [TAR 2.0/CAL endorsement]
Hyles v. New York City, 2016 WL 4077114 (S.D.N.Y. 2016) [Court mandated TAR over linear review]

Regulatory Investigations

Health and Care Professions Council (HCPC). Fitness to Practise Process. https://www.hcpc-uk.org/concerns/
General Medical Council (GMC). Investigation and Case Examiners. https://www.gmc-uk.org/
British Psychological Society (BPS). Professional Conduct Procedures. https://www.bps.org.uk/

Intelligence Analysis

Heuer, Richards J., Jr., and Randolph H. Pherson. Structured Analytic Techniques for Intelligence Analysis. 3rd ed., CQ Press, 2021.
US Office of the Director of National Intelligence. Intelligence Community Directive 203: Analytic Standards. January 2, 2015.
NATO. AJP-2: Allied Joint Doctrine for Intelligence. November 2016.
Kent, Sherman. "Words of Estimative Probability." Studies in Intelligence 8, no. 4 (1964): 49-65.

Academic Research

Page, Matthew J., et al. "The PRISMA 2020 statement: an updated guideline for reporting systematic reviews." BMJ 372 (2021): n71.
Higgins, Julian PT, et al. Cochrane Handbook for Systematic Reviews of Interventions. Version 6.5, 2024.
Braun, Virginia, and Victoria Clarke. "Using thematic analysis in psychology." Qualitative Research in Psychology 3.2 (2006): 77-101.
Charmaz, Kathy. Constructing Grounded Theory. 2nd ed., Sage, 2014.
Lincoln, Yvonna S., and Egon G. Guba. Naturalistic Inquiry. Sage, 1985.

Document Control

Version: 1.0 Date: 2026-01-17 Author: Interdisciplinary synthesis for Phronesis FCIP Classification: Reference Standard - Cross-Domain Integration Review Cycle: Semi-annual update recommended to reflect evolving methodologies License: Internal reference for Phronesis FCIP development

End of Document