Compiler

Layer 0: pre-generation rewrite engine for governed inference.

Prompt Compiler analyzes and rewrites prompts before the model generates a single token, preventing structural failures that post-generation scoring cannot fix.

Without it

Without pre-generation structural integrity, the model receives prompts that reward hallucination. Evidence gets buried by recency bias, authority signals are lost in unstructured context, and post-generation scoring catches failures too late to prevent them.

Position in the platform

Oracle Foundry (Foundation): Supplies authoritative chunks with SIRE metadata and tier signals
SIRE Crosswalk (Post-Foundry): Routes concepts across oracles and flags authority boundary risks before generation
Prompt Compiler (L0): Rewrites prompt structure before inference to preserve authority and grounding
Claim Ledger (L1-L4): Scores confidence, entailment, stability, and representation uncertainty
Process Control System (Cross-layer): Tracks quality drift and special-cause variation over time
Forensics Lab (L5): Converts failures into compiler and oracle remediation loops

CFPO ordering model

The compiler places content by function and relevance, matching known attention behavior instead of appending retrieval chunks as an undifferentiated block.

Content: Domain framing and system context (Early context)
Format: Reference structures and definitions (Mid context)
Policy: Grounding and authority constraints (Late context)
Output: Response contract and output schema (Final context)

Five rewrite passes

Context placement optimization: Positions chunks by CFPO zone and attention profile so evidence is visible where it matters.
Instruction reinforcement: Duplicates critical grounding constraints near generation to increase compliance reliability.
Authority zone isolation: Separates evidence, instructions, and user input with model-validated delimiters.
Mechanism competition detection: Detects likely parametric-vs-retrieval conflicts and strengthens retrieval authority signals.
Format-pressure resolution: Resolves schema-completeness pressure when required fields lack supporting evidence.

Crayon Contract

Only output claims that can be tied to provided source text
Numbers, entities, dates, and citations require explicit support
Conflicting sources are reported, not reconciled by guesswork
Uncertain tokens are omitted instead of emitted