Catalog/Classes/Fermi
CLASS · 02 / 06v0.4Live · v0.4

Fermi

Precise enough to act on, uncertain enough to be honest.

Named afterEnrico Fermi (1901 — 1954) — physicist famous for back-of-envelope estimation under thin data.

9.4% mean absolute error90% intervals captured actuals 89% of the time
01 — The science

Why we named it Fermi.

Fermi was famous for answering questions no one thought answerable — how many piano tuners in Chicago? — with nothing but logic and back-of-envelope estimation.

The Fermi agent does the same thing with operational data: sparse signals, incomplete records, partial logs. It doesn't wait for clean data. It estimates with principled uncertainty, flags what it can't know, and gives you a number you can act on.

Every Fermi answer comes with bounds. You see the estimate, the confidence, and the assumptions — not because we bolted it on, but because order-of-magnitude reasoning is the contract.

  • 01Factor decomposition → errors cancel: decompose any target into independently estimable factors; independent errors tend to partially offset each other.
  • 02Monte Carlo propagation → output is a distribution, not a number: sample each factor's uncertainty through the factor tree (FERMIAC discipline).
  • 03Calibration as a contract: stated 90% intervals must capture actuals ~90% of the time, tracked over the fleet's history.

Read the design research →

Fermi class

How many piano tuners are in Chicago?

— ENRICO FERMI · SEMINAR, UNIVERSITY OF CHICAGO
Fermi class
1 / 4Why is it called Fermi?

How many piano tuners are in Chicago? Bounded uncertainty: precise enough to act on, uncertain enough to be honest.

Enrico Fermi · University of Chicago Seminar
tap to continue →
Field Evidence
ACADEMIC BASIS
"Approaching Human-Level Forecasting with Language Models"Retrieval + decomposition + ensembling system achieves Brier .179–.240 vs crowd .149–.247; raw LLMs without the scaffold score at or worse than random (.25) — the system forecasts, not the model.
Halawi, Zhang, Yueh-Han & Steinhardt · NeurIPS 2024
IN PRODUCTION
Amazon DeepAR — probabilistic forecastingMonte Carlo sample paths → calibrated quantile estimates; shipped in Amazon Forecast / SageMaker and Amazon labor-planning systems.
Salinas et al. · arXiv:1704.04110; Amazon, 2022
BENCHMARK
o3 Brier 0.1352 vs crowd baseline 0.149Surpasses crowd on ForecastBench; AIA Forecaster matches superforecaster parity (0.0753 vs 0.0740). Calibration contract framed as an enforced gate — LLM ECE remains several times worse than superforecaster ECE.
ForecastBench · arXiv:2507.04562, 2025; KalshiBench · arXiv:2512.16030
02 — Agents in this class

Prototype agents.

Every class ships with reference agents calibrated to operational use cases. Fork them, deploy them, or use them as a template.

Demand Signal Forecaster

MAPE 8.7% · −39% vs baseline

Procurement Spend Estimator

96% anomalies caught · 2.1% FP

Capacity Headroom Estimator

overprovisioning −18%

Inventory Risk Quantifier

11-day median stockout warning

Pipeline Revenue Sizer

±12% across 3 quarters
12-WEEK BETA · 9 DESIGN PARTNERS · 47,000 SHADOWED RUNS
03 — Qualification gate

The ALOFT pipeline, applied.

Every agent in this class passes the same five-stage gate. Below: the criteria specific to Fermi agents at each stage.

ALOFT
01 · Curation
02 · Staging
03 · Deploy
04 · Operate
05 · Generalise
A→L→O→F→T
01
Curation
Sparse-signal scope; calibration target set
  • Sparse-signal use case
  • Bounded uncertainty acceptable
  • Calibration target defined
02
Staging
Order-of-magnitude eval; bounds validated
  • Order-of-magnitude eval
  • Bounds tested on holdouts
  • Calibration plot validated
03
Deployment
Interval published; shadow-runs ≥7d
  • Confidence interval published
  • Estimate shadow-runs ≥ 7d
  • Drift baseline frozen
04
Operation
MAPE <12%; calibration drift <5%
  • Estimate vs actual MAPE < 12%
  • Calibration drift < 5%
  • Cost guardrail enforced
05
Generalisation
Estimator reused; signal source qualified
  • Estimator reused across class
  • New signal source qualified
  • Pattern memo published

Ready to deploy a Fermi agent?

Open the Workbench →Back to catalog
All classes
FeynmanFermi· currentMaxwellRamaMarcusWheeler