Inferencing Loops

Leon Como
5 days ago
2 min read

Open Drop: Generative Inferencing Protocol Layer

Protocol-compressed decisions, LLM-decompressed pathways.

1) What I’m sharing (and why)

I’m sharing a model-agnostic protocol layer for using LLMs as a compression/decompression engine for decisions.

Most GenAI usage optimizes for outputs. This approach optimizes for decision pathways: portable, auditable, versioned, and grounded in temporal reality.

Core claim: If we compress decisions into protocol-shaped artifacts, LLMs can reliably decompress them into context-specific pathways—without devolving into performative content.

2) The Decision Cycle Reference Architecture (open)

This is the cycle I’m releasing publicly as a reference, not a turnkey playbook:

Prompt → COP → PIT → FNT → Decision/UVT → COPOP → (Nudge) → Prompt

Stage contracts (anti-performativity gates)

Each stage has a pass/fail contract. If it fails, you loop back intentionally.

A) Prompt (intent + bounds)

Input: a real decision need

Must include: purpose, scope, constraints, success criteria Fail if: it cannot change a decision, reduce uncertainty, or trigger an experiment

B) COP (Chain of Prompts)

Goal: controlled divergence → convergence

Must produce: 2–3 viable options plus explicit constraints and unknowns

Fail if: it produces “more content” without narrowing decisions

C) PIT (Protocolized Insight Token)

Definition: a decision-preserving claim with explicit metadata

Must include: assumptions, scope, confidence, evidence/provenance, validity window

Fail if: it’s a slogan, timeless claim, or lacks invalidation conditions

D) FNT scoring (Fidelity, Novelty, Translation) — applied to PIT

Purpose: quality gates, not vanity ratings

Actions triggered:

Low F → add evidence or narrow scope
Low T → rewrite into executable steps/tests
Low N → acceptable if it improves reliability; flag if exploration is needed

E) Decision / UVT (Unique Value Token)

Decision output: choice + rationale + test/measurement plan

UVT requirement: attach an NPO trace (what changed outside the model)

Fail if: no real-world effect, adoption, test, or measurable change is recorded

F) COPOP (Chain-of-Prompts Organized Prompting)

Output: reusable prompt-program packaging

Must include: retrieval hooks, versioning notes, trigger conditions, rollback/failover

Fail if: it cannot be reused by another person/context without the author present

G) Nudge (event-driven re-entry)

Triggers: new data, drift, failure, time decay, context change, stakeholder shift

Fail if: iteration is “because we feel like it” rather than triggered

3) What I’m NOT sharing (reserved leverage)

To be clear about boundaries:

I am not open-sourcing my proprietary modeling system (CTF)
I am not releasing my tuned orchestration heuristics (“the runner”)
I am not releasing curated UVT corpora, retrieval ranking logic, or distribution strategy