# Aegis Moderation Decision Pipeline

This pipeline is **mandatory** and must be executed in full for every item. No shortcuts are permitted.

## Phase 1: Ingestion & Normalization
- Capture exact original text (preserve edits history if available)
- Strip or note zero-width and control characters
- Identify primary language and any code-switching
- Parse metadata (author tenure, verification, prior strikes, visibility, media descriptions)

## Phase 2: Contextual & Intent Analysis
- Reconstruct full conversational or thread context
- Evaluate author intent using multiple plausible frames (sincere, satirical, journalistic, trolling, hypothetical)
- Assess target (individual, group, public figure, no specific target)
- Note power dynamics and historical targeting patterns

## Phase 3: Violation Hypothesis Generation

Produce the top 2–3 candidate violations from the taxonomy. For each hypothesis record:
- Supporting verbatim evidence
- Counter-evidence
- Most applicable policy clause

## Phase 4: Policy Application & Scoring
- Select the single best-matching specific policy provision
- Assign severity level using the matrix
- Apply reach, recidivism, and vulnerability multipliers
- Generate initial recommended action

## Phase 5: Counterfactual & Confidence Check
- Explicitly test the strongest "allow" interpretation
- Assign final confidence (0–100)
- If confidence < 80 on non-critical items → recommend human review
- For all CRITICAL items → escalate regardless of confidence

## Phase 6: Decision Finalization & Documentation

Select action from the approved menu:
- ALLOW
- LABEL / WARNING STICKER
- REDUCE_VISIBILITY
- REMOVE_CONTENT
- TEMPORARY_SUSPENSION (specify duration)
- PERMANENT_ACCOUNT_TERMINATION
- LEGAL_REFERRAL (specify authority)

Produce full audit record containing every phase output.

## Phase 7: User Communication & Logging
- Generate appropriate user-facing notice (when policy requires)
- Create immutable internal log entry
- Flag novel patterns or policy gaps for human policy team review