---
title: Croms™ — The unit of preventable AI workflow waste
description: A Crom is one unit of preventable AI workflow waste. Croms measure inefficiency across four dimensions — cost, latency, failure risk, and structural inefficiency — giving teams a single deterministic score for any workflow.
canonical: https://cromus.ai/croms
source_html: https://cromus.ai/croms
---

# Croms™

> **One Crom = one unit of preventable AI workflow waste.**

Croms are Cromus's proprietary measurement unit. They quantify how much of a workflow's cost, time, and failure risk is *preventable* — meaning it could be removed without changing the workflow's outcome.

---

## The four dimensions of waste

A workflow's total Croms score is the sum of waste across four orthogonal dimensions:

| Dimension | What it measures | Example |
|-----------|------------------|---------|
| **Cost waste** | Tokens spent on tasks a cheaper model could do equally well | Using GPT-4 for a classification a 3B-param model handles |
| **Latency waste** | Time lost to serialised steps that could run in parallel | Three independent API calls chained instead of fanned out |
| **Failure risk** | Cost of expected retries, fallbacks, and error correction | A flaky JSON-mode prompt with no schema validation |
| **Structural inefficiency** | Tokens spent on redundant context, duplicated prompts, dead branches | Re-pasting a system prompt into every sub-agent call |

---

## How Croms are computed

Croms are computed **deterministically** — no LLM call, no fuzzy reasoning. Given a compiled SKILL.md and a verified model registry, the score is a pure function of:

- Token estimates per step (input + output)
- Per-model unit cost from the verified registry (62 models)
- Step dependency graph (serial vs. parallelisable)
- Declared retry / fallback policy from ETHOS.md
- Workflow structure (redundancy detection, context reuse opportunities)

The same workflow always produces the same Croms score. No randomness, no model-dependent variance.

---

## Why a new unit

Existing metrics measure different things:

- **Cost in dollars** is observable but conflates necessary spend with waste.
- **Tokens** are observable but don't account for opportunity cost across model tiers.
- **Latency** misses cost-waste and risk-waste entirely.
- **Failure rate** misses everything that succeeded but cost too much.

Croms is the single composite metric that lets a team say *"this workflow has 340 Croms of preventable waste"* and act on it directly — by downgrading models, parallelising steps, deduplicating prompts, or adding schema-based retries.

---

## Where Croms appear in the product

- **5-step demo** — Croms score is part of the optimisation output
- **Skill Studio** — every loaded SKILL.md gets a Croms score
- **MCP server** — `score_skill` tool returns Croms breakdown
- **Cost simulator** — Croms delta shown for each cost-mode comparison
- **Ops Efficiency Score** — derived from Croms via a deterministic formula

The same Croms score appears across web, MCP, and in-app surfaces — single source of truth, no drift.

---

## Related concepts

- [Pre-Execution Governance](/pre-execution-governance.md)
- [Workflow Intelligence as a Service](/workflow-intelligence-as-a-service.md)
- [Baseline Cost per Workflow](/baseline-cost-per-workflow)
- [Total Cost of Workflow Ownership (TCWO)](/total-cost-of-workflow-ownership)