Back to the Tracker

Promised vs. Produced Tracker · methodology

How we collected this.

Our Promised vs. Produced Tracker (“Tracker”) is the central artifact of our methodology. Its purpose is to measure to what extent announced American factories actually result in output.

01 · What we record

What we record for each project.

For each individual project that is announced, we obtain the date of announcement, the promised completion date, and the scale of the project in terms of both promised capital (USD) and promised jobs (all if available).

Promise field

Date of announcement

When the project was first announced.

Promise field

Promised completion date

When the announcement said the project would be done.

Promise field

Promised capital

The scale of the project, in USD.

Promise field

Promised jobs

The scale of the project, in jobs.

Status field

Date of first industrial output

The first time the facility produces any goods related to what was promised for commercial purposes.

  • Goods related to what was promised, made for commercial purposes
  • Test or pilot runs
Status field

Current status

We also assign each project one of six current statuses:

  • As announced
  • Under construction
  • Paused
  • Producing
  • Closed
  • Canceled

What the dates let us measure

An illustrative project, drawn the way the Tracker draws each row

announcement to promised first output past the promised date, not producing producing
Notably, this allows us to consider how long a factory took to produce and how much it was ahead or behind schedule.

02 · What counts

Which factories we consider.

We consider factories that satisfy the following:

  1. 01

    In the United States

    Located in a US state.

  2. 02

    One facility, one promise

    Covers a single physical facility, not a multi-site announcement — and tracks the specific promise made about that facility, not the facility itself.

  3. 03

    Announced in 2017 or later

    Their announcement is published on or after January 1, 2017.

  4. 04

    At least $1 billion or 2,000 jobs

    Their announcement promises either at least $1 billion in capital OR 2,000 new jobs (directly in the factory).

    $1B+
    promised capital
    OR
    2,000+
    new jobs, directly in the factory
  5. 05

    In one of ten sectors

    Produces in one of the following industrial sectors:

    • Aerospace and Defense
    • Auto Assembly
    • Battery
    • Chemicals and Plastics
    • Food and Beverage
    • Machinery
    • Pharmaceuticals
    • Semiconductors
    • Solar
    • Steel

    All others are categorized as “Other.”

03 · The pipeline

Three layers: Source, Screen, Verify.

To collect the Tracker, we construct a pipeline with three layers: Source, Screen, and Verify.* Both Source and Screen are extracted agentically, while Verify — the Tracker in its final form — is a human-only gate.

run by an agent† human only the published Tracker
01

Source

Agent†

We start with Source, where we search for a new project and collect URLs that contain information on both its initial announcement and its current status.

Passes on

  • Announcement URLs
  • Current-status URLs
02

Screen

Agent†

Screen explicitly extracts the information from these links to provide provenance for our measured values on the factory’s development.

Passes on

  • Promise fields, from the announcement
  • Status fields, from the status articles
  • Blanks where nothing is attributable
03

Verify

Human only

Verify is where a human makes the final call and any remaining edits to add the new project to the Tracker, confirming that a human will have examined and verified all provenance from the links.

Passes on

  • A verified row

The Tracker

Verify’s output, in its final form.

Keeping the agents controlled

  • We run the Source and Screen agents† in a controlled environment where they always see the same prompts and files for their context.
  • We build the database up from empty tables — we found that doing otherwise only raises concerns about how much we are steering the agent.

Tools for the verifiers

  • A local interactive user interface
  • Data quality checkers

Screen binds every field to its source

Data fields in this table are bound by role

  • Announcement articles
    • Date of announcement
    • Promised completion date
    • Promised capital
    • Promised jobs
  • Status articles
    • Date of first industrial output
    • Current status
  • Nothing attributable
    • left blank
Promise fields must be attributable to the announcement articles and status fields to the status articles, with anything unattributable left blank.

* Akin to ELT medallion pipelines in data engineering.

† Claude Code, using Opus 4.8 on high reasoning.

04 · The repository

The Tracker is open to add to.

Though we keep our methodology tighter for this work, we publish our Git repository to allow fair use and (agent-agnostic) additions to the Tracker.

github.com/industriousaf/
promised-vs-produced-tracker branch · mvp Open the repository →

Wider criteria in the repository

PromisedOn the TrackerIn the repository
Capital$1B$100M
Jobs2,000200

With that, we loosen our current criteria from $1B in promised capital to $100M, and from 2,000 promised jobs to 200, because as currently defined, the criteria cap the pool at 162 projects.

We also update the Tracker quarterly alongside our previous project, the Industrial Dollar, for public output on the Tracker page.

05 · Prior art

What already exists.

Existing work falls into three groups of trackers: those that record what was promised, those that follow what happened after the promise, and those that do some of both but within a single sector. The Promised vs. Produced Tracker brings these together by following each announced U.S. facility from promise to first output, across sectors.

01 “Promised” trackersThose that record what was promised. 5 trackers
  • IndustrialSageWeekly tally of announced U.S. manufacturing commitments of $50M+ since 2025; records announced dollars only, with no facility status or output date.
  • Avison YoungMonthly tally of announced manufacturing investments over $100M since January 21, 2025, classified by NAICS subsector; keeps withdrawn projects but does not record whether projects reach production.
  • Reshoring InitiativeAnnual aggregate counts of jobs announced through reshoring and FDI; assumes a two-year lag from announcement to hire rather than observing timing project by project.
  • fDi Markets (Financial Times)Commercial database of announced cross-border greenfield projects; covers foreign investors only and does not track realization.
  • Good Jobs FirstNational database of company-specific subsidy awards, with a Megadeals category for awards of $50M or more; organized by incentive, not by production milestone.
02 “Produced” trackersThose that follow what happened after the promise. 3 trackers
  • Engineered-VisionMap of $1B+ U.S. manufacturing megaprojects (59 projects, updated April 2026) with narrative status notes; has no dated first-output field, jobs-based floor, or downloadable data.
  • Dezernat ZukunftQuarterly tracker of large private investment announcements in Germany, including subsidies, revised when projects change; does not record construction or production milestones.
  • E2 (Environmental Entrepreneurs)Monthly tracking of clean-energy project announcements, cancellations, closures, and downsizes since August 2022; covers the cancellation side but not whether projects reach production.
03 Sector-specific trackersThose that do some of both, but within a single sector. 4 trackers
  • Rhodium Group & MIT CEEPRQuarterly, facility-level data on investment in clean-energy manufacturing and deployment since 2018; limited to clean technologies and measures capital spent, not first output.
  • SEMIPaid, line-by-line tracking of global fab spending, construction, and production capacity; covers semiconductors only and is not public.
  • Semiconductor Industry Association (SIA)Map of U.S. semiconductor sites flagged as existing or announced, with investment and expected jobs, largely representing SIA member companies; has no timing or realization dimension.
  • Kiel Institute for the World EconomyTracks European military orders, including earliest and latest expected delivery dates; it applies a promise-to-delivery design to procurement rather than factories.

06 · Limits

Where this method falls short.

Nonetheless, this methodology does carry some limitations.

Smaller

Some follow from our inclusion criteria

  • We only examine the United States.
  • We only rely on publicly available sources.
  • We cover a subset of industrial sectors.
Smaller

Others come from how we record our projects

  • One status per project. We assign every project a single current status, but our fixed set of categories cannot capture the full nuance of a company’s situation.
  • Jobs without kinds. Likewise, when announcements promise jobs, they seldom specify what kinds of roles are involved, so we cannot say what kind of employment a promise represents.
Larger

The caveats we must always face while working with black-box AI models

  • Reproducibility is not guaranteed. Even though we run the agents in a controlled environment, reproducibility is not guaranteed due to stochasticity and our inability to know whether the model is being changed behind the API.
  • Little to benchmark against. Finally, because this methodology is novel, there is little existing work measuring the same thing that we can benchmark our results against.