Distributive Intent Lattices for Prompt-to-Specification

Pipeline Process

1 Natural Language Input

"If the button is pressed, the light turns on"

2 Parse & Extract Facets

3 Build Intent Lattice

Model all possible interpretations

4 Disambiguate

Ask clarifying questions

5 Generate Formal Spec

always (ButtonPressed → next LightOn)

Each node represents a possible interpretation
∀ = Universal, ∃ = Existential
Im = Immediate, Ev = Eventually

Input: "When a user logs in, they must accept the Terms or they cannot proceed."

Ambiguities Detected:

Clarifying Question: "Should this rule apply every time a user logs in?"

Final Spec: always (Login → (TermsAccepted ∨ ¬Proceed))

Mathematical structure capturing all possible interpretations ordered by specificity

Independent semantic dimensions like temporal scope, quantifiers, and guards

Information-theoretic questions that maximally reduce ambiguity

Unambiguous controlled English and logic-ready specifications

FDL bitset GLB (feature-flagged): join-irreducible facets encoded as bits; GLB via bitwise AND. Toggle with TAU_OPTIMIZER_USE_FDL=1.
Code: backend/unified/domain/prompt_optimizer_pnf_ilgo.py
GS/NSO clean-room synthesis: build GS IR and pretty-print to controlled English (no Tau assets). Toggle with TAU_OPTIMIZER_USE_GS=1.
Code: backend/unified/domain/gs_ir.py
Endpoint integration: Optimizer runs before LLM; constraints gate tokens; clarifications surfaced in suggestions.
Code: backend/unified/api/llm_endpoints.py
A/B comparison tool: legacy vs FDL paths on sample prompts.
Tool: tools/compare_optimizers.py (local TestClient)
Observed (local sim): FDL reduced avg latency and clarifications with identical validity across 8 prompts.
License-safety: Tau printing remains a user-installed adapter; this page and code target semantics only.

See also: the interactive translator at translator.html.