\documentclass[11pt]{article}

% ---------- packages ----------
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
\usepackage[margin=1in]{geometry}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{multirow}
\usepackage{array}
\usepackage{caption}
\usepackage{subcaption}
\usepackage{float}
\usepackage{setspace}
\usepackage{natbib}
\usepackage{xcolor}
\usepackage{titlesec}
\usepackage[hidelinks,colorlinks=true,linkcolor=blue!50!black,citecolor=blue!40!black,urlcolor=blue!50!black]{hyperref}

\graphicspath{{../figures/}}
\captionsetup{font=small,labelfont=bf}
\onehalfspacing
\titleformat{\section}{\large\bfseries}{\thesection}{0.6em}{}
\titleformat{\subsection}{\normalsize\bfseries}{\thesubsection}{0.6em}{}

\newcommand{\R}{$R^2$}
\newcommand{\fig}[1]{Figure~\ref{#1}}
\newcommand{\tab}[1]{Table~\ref{#1}}

% ---------- title ----------
\title{\bfseries The Corporate Project Economy:\\[0.3em]
\large Large-Scale Extraction and Econometric Analysis of 19{,}000 Strategic
Initiatives from Two Decades of SEC Filings (2010--2026)}

\author{%
  \begin{tabular}{c}
    {\large Manel Kammoun}\\[-0.05cm]
    {\small\href{mailto:manel.kammoun@uqo.ca}{manel.kammoun@uqo.ca}}\\[0.35cm]
    {\large Djerrycharli Tandjambianda}\\[-0.05cm]
    {\small\href{mailto:djerrycharli.tandjambianda@uqo.ca}{djerrycharli.tandjambianda@uqo.ca}}\\[0.35cm]
    {\large Simon-Pierre Boucher}\\[-0.05cm]
    {\small\href{mailto:simon-pierre.boucher@uqo.ca}{simon-pierre.boucher@uqo.ca}}
  \end{tabular}\\[0.5cm]
  \small D\'epartement des sciences administratives\\
  \small Universit\'e du Qu\'ebec en Outaouais\\[0.3cm]
  \small Gatineau -- Pavillon Alexandre-Tach\'e\\
  \small 283, boulevard Alexandre-Tach\'e\\
  \small Gatineau, Qu\'ebec, Canada J9A 1L8%
}

\date{\today}

\begin{document}
\maketitle

\begin{abstract}
\noindent
Corporate strategy is enacted through \emph{projects} --- plant builds, data
centers, acquisitions, research programs, digital migrations --- yet these
initiatives are scattered, unstructured, and buried inside the narrative
sections of regulatory filings. We introduce SPID (the SEC Project Intelligence
Database), a pipeline that converts the raw text of corporate filings into a
structured, project-centric panel, and we use it to build the largest
machine-extracted census of corporate projects we are aware of: \textbf{19{,}227
resolved projects} drawn from \textbf{39{,}930 mentions} across
\textbf{153{,}894} filing sections (10-K, 10-Q and 8-K) filed by \textbf{500
large U.S.\ public companies} between 2010 and 2026. A large-language-model
extractor (\texttt{gpt-4o-mini}, JSON mode) maps each section to a controlled
21-category project taxonomy with quantitative attributes (announced capital,
status, locations, technologies, partners), which are then resolved into
firm-level project portfolios with timelines, a 32{,}226-node knowledge graph,
and 384-dimensional semantic embeddings. Using this panel we estimate a battery
of models: fixed-effects Poisson and negative-binomial count models of
project-announcement intensity; hedonic log-capital regressions with firm and
year effects; a logistic technology-diffusion model with an estimated inflection
in \textbf{2023}; a difference-in-differences event study of COVID-19 and
corporate digitalization; Cox proportional-hazards models of project lifecycles;
community detection over a technology co-occurrence network; unsupervised
clustering of project embeddings; and a gradient-boosted classifier of
``mega-projects'' that attains a cross-validated AUC of \textbf{0.74} from
text-derived features alone. We document strong persistence in project activity,
a sharp post-2020 acceleration of announcements, an energy-transition technology
cluster at the center of the corporate-innovation network, and project
lifecycles that range from a median of four years for M\&A integration to twelve
years for digital and AI programs. Beyond the substantive findings, the paper
demonstrates that LLM-based information extraction can turn the unstructured
regulatory record into research-grade panel data for empirical
finance and the economics of innovation.

\vspace{0.5em}
\noindent\textbf{Keywords:} corporate investment, text-as-data, large language
models, information extraction, innovation, SEC filings, knowledge graphs,
panel econometrics.\\
\noindent\textbf{JEL:} G31, O30, C55, D22, M15.
\end{abstract}

\clearpage
\tableofcontents
\clearpage

% =====================================================================
\section{Introduction}
% =====================================================================
The modern public corporation is, in operational terms, a portfolio of
\emph{projects}. New products are launched, plants and data centers are built,
firms are acquired and integrated, enterprise systems are migrated to the cloud,
research pipelines are advanced, and entire business lines are reoriented around
new technologies such as artificial intelligence. These initiatives are the
mechanism through which strategy becomes capital expenditure and, eventually,
realized returns. Yet for the empirical researcher they are remarkably hard to
observe. Financial statements report aggregate capital expenditure, not the
projects that compose it; press releases are selective and unstructured; and the
richest single source --- the narrative discussion inside regulatory filings ---
is written in prose, not in tables.

This paper has two goals. The first is methodological: to show that recent
advances in large language models (LLMs) make it possible to convert the
unstructured regulatory record into structured, research-grade panel data at
scale and at low cost. The second is substantive: to use the resulting data to
characterize the \emph{corporate project economy} of large U.S.\ firms over the
2010--2026 period and to estimate a series of econometric models that speak to
classic questions in corporate investment and the economics of innovation.

We construct SPID, the SEC Project Intelligence Database, on top of a 500-company
corpus of filings downloaded from EDGAR. The pipeline (i) sections raw 10-K HTML
into its project-dense items (Business, Risk Factors, Properties, MD\&A), (ii)
sends every project-relevant text window to an LLM extractor that returns a
strict-JSON project record against a 21-category taxonomy, (iii) resolves the
resulting mentions into firm-level projects with timelines using stable hashed
identifiers, and (iv) builds a knowledge graph and a set of semantic embeddings
on top of the resolved projects. The end product is a panel of
\textbf{19{,}227 projects} and \textbf{39{,}930 dated mentions} spanning
\textbf{500 firms}, \textbf{11 GICS sectors}, \textbf{21 project types}, and
\textbf{3{,}080 distinct locations}, of which \textbf{9{,}321 (48.5\%)} carry an
explicit dollar amount with a median announced value of \textbf{US\$718 million}.

The scale and structure of this panel allow us to ask questions that aggregate
data cannot answer. \emph{How persistent is project activity within firms, and
did it change after 2020?} We estimate fixed-effects Poisson and
negative-binomial models and find strong, statistically robust persistence and a
significant post-COVID jump in announcement intensity, against a mild negative
secular time trend once firm composition is held fixed. \emph{What kinds of
projects command the most capital?} Hedonic log-capital regressions with firm
and year fixed effects explain roughly a third of the variation in announced
amounts and reveal large premia for capital-intensive project families relative
to M\&A integration. \emph{How fast is corporate AI adoption diffusing?} A
logistic diffusion model places the inflection point of cumulative AI adoption
in \textbf{2023}, coincident with the generative-AI boom, and the AI share of
new project announcements rises from 0.7\% in 2015 to 4.7\% in 2025. \emph{Did
the pandemic differentially accelerate the digitalization of physically
intensive incumbents?} A difference-in-differences event study with clean
parallel pre-trends returns a precisely estimated null, cautioning against
the popular narrative of pandemic-driven digital convergence. \emph{How long do
projects live?} Cox and Kaplan--Meier models show that M\&A integrations resolve
fastest (median four years) while digital and AI programs persist longest
(median twelve years), and that larger projects complete more slowly.

We complement these regressions with network and representation-learning
analyses. A technology co-occurrence network exhibits strong community structure
(modularity 0.52) organized into four ``technology ecosystems,'' with
energy-transition technologies --- solar, wind, battery storage --- occupying the
most central positions in the corporate-innovation graph. Unsupervised
clustering of the 384-dimensional project embeddings recovers six interpretable
themes that map cleanly onto the taxonomy while crossing sector boundaries.
Finally, a gradient-boosted classifier predicts whether a project is a
``mega-project'' (above the \$718m median) with a cross-validated AUC of 0.74
using only text-derived structural features, with project type, announcement
year and sector the most informative predictors.

\paragraph{Contributions.} We make four contributions. First, we provide a
\emph{reproducible, low-cost methodology} for turning the narrative sections of
regulatory filings into structured project-level panel data using an LLM
extractor with a controlled taxonomy and deterministic resolution. Second, we
deliver a \emph{novel dataset} --- to our knowledge the largest
machine-extracted corporate-project census --- together with a knowledge graph
and embeddings that support search, analytics, and downstream modeling. Third,
we present a \emph{comprehensive econometric analysis} that combines count
models, hedonic regressions, diffusion curves, difference-in-differences,
survival analysis, network science, and machine learning on a single coherent
panel. Fourth, we surface \emph{substantive findings} about persistence, the
post-2020 surge, AI diffusion timing, project longevity, and the centrality of
the energy transition that are difficult or impossible to obtain from
conventional structured data.

\paragraph{Caveats up front.} Our census is an extraction from disclosure, not a
ground-truth registry. Announced amounts are self-reported and heterogeneous in
scope; the headline aggregate of announced capital should be read as an
\emph{upper-bound order of magnitude} rather than a netted investment total,
because the same multi-year program can be mentioned (and counted) in several
filings and because large M\&A enterprise values dominate the sum. Throughout,
we therefore emphasize \emph{within-firm} and \emph{relative} comparisons, which
are robust to these level issues, and we treat the dollar figures as ordinal
signals of project scale. Section~\ref{sec:limitations} discusses measurement,
selection, and extraction error in detail.

The remainder of the paper proceeds as follows.
Section~\ref{sec:related} situates the work in the text-as-data and corporate
investment literatures. Section~\ref{sec:data} describes the corpus, the
extraction pipeline, and the resolution procedure. Section~\ref{sec:descriptives}
presents descriptive statistics and stylized facts. Section~\ref{sec:framework}
lays out the econometric framework. Sections~\ref{sec:counts}--\ref{sec:ml}
report results for each family of models. Section~\ref{sec:robustness} collects
robustness checks, and Section~\ref{sec:limitations} discusses limitations.
Section~\ref{sec:conclusion} concludes.

% =====================================================================
\section{Related Literature}\label{sec:related}
% =====================================================================
Our work sits at the intersection of three literatures: text-as-data methods in
economics and finance, the empirical analysis of corporate investment, and the
measurement of technological change and innovation.

\paragraph{Text as data.} A now-large body of work extracts economic signals
from corporate disclosure. Early contributions used dictionary methods and
readability indices to study tone and sentiment in 10-K filings
\citep{loughran2011liability,li2008annual}, while later work measured risk and
attention from the co-occurrence of words and topics
\citep{hoberg2016text,hassan2019firm}. \citet{gentzkow2019text} survey the
econometrics of text and emphasize the move from counting words to learned
representations. We extend this trajectory in two ways. First, rather than
scoring documents, we perform \emph{structured information extraction}: each text
window is mapped to a typed record with named entities and quantities. Second,
we use a generative LLM as the extraction engine, which removes much of the
brittle feature engineering required by earlier rule-based systems while
retaining a controlled output schema.

\paragraph{LLMs for extraction and measurement.} Recent papers show that LLMs
can classify and extract economic content with accuracy approaching that of
trained human readers, at a fraction of the cost
\citep{korinek2023language,bybee2023ghost}. The methodological question is no
longer whether extraction is feasible but how to make it \emph{reproducible,
auditable, and cheap at corpus scale}. Our pipeline addresses this with a
deterministic taxonomy, JSON-mode decoding, confidence thresholds, and stable
hashed mention identifiers that make the entire process incremental and
resumable. The marginal cost of extracting a project mention is on the order of
a tenth of a U.S.\ cent, which is what makes a census of this size practical.

\paragraph{Corporate investment.} The empirical investment literature has long
worked with aggregate capital expenditure and Tobin's $q$
\citep{hayashi1982tobin,fazzari1988financing}, and more recently with the
rising importance of intangible and innovation-related investment
\citep{peters2017intangible,crouzet2019understanding}. A persistent
difficulty is that the unit of decision --- the project --- is not observed.
Our data make the project the unit of analysis, allowing us to study the
\emph{extensive margin} (how many initiatives a firm announces) separately from
the \emph{intensive margin} (how large each is), and to characterize the
\emph{composition} of investment across project types and technologies.

\paragraph{Innovation and diffusion.} The economics of technology diffusion has
modeled adoption with logistic and Bass-type curves
\citep{griliches1957hybrid,bass1969new}, and a recent literature measures the
diffusion of artificial intelligence across firms using job postings, patents,
and earnings calls \citep{babina2024artificial,acemoglu2022artificial}. We
contribute a disclosure-based measure of AI \emph{projects} and estimate its
diffusion path directly, finding an inflection in 2023 that aligns with the
arrival of large generative models.

\paragraph{Networks and representations.} Finally, we connect to work that
represents firms and technologies as networks \citep{hidalgo2007product} and as
learned embeddings. Our knowledge graph and sentence-embedding clustering let us
recover technology ecosystems and latent project themes that cut across the
formal taxonomy, complementing the regression evidence with structure that is
not imposed a priori.

% =====================================================================
\section{Data and Extraction Pipeline}\label{sec:data}
% =====================================================================
\subsection{Source corpus}
The underlying corpus is a 500-company subset of U.S.\ SEC filings retrieved from
EDGAR, comprising raw filing HTML and a pre-parsed store of filing metadata and
sectioned text. Two annual and quarterly forms (10-Q and 8-K) were already
sectioned upstream; the project-dense 10-K filings were retained as raw HTML. As
the first stage of SPID we ingested and sectioned \textbf{7{,}654} 10-K filings,
extracting their Business (Item~1), Risk Factors (Item~1A), Properties (Item~2),
and Management's Discussion and Analysis (Item~7/7A) sections into
\textbf{27{,}164} clean text sections. Unioned with the pre-parsed 10-Q and 8-K
sections, the analysis feed contains \textbf{153{,}894} filing sections (with at
least 30 words) eligible for extraction: 103{,}926 from 8-K, 27{,}164 from 10-K,
and 22{,}804 from 10-Q filings.

\subsection{LLM extraction}
Each eligible section is passed to an extraction model
(\texttt{gpt-4o-mini}, temperature 0, JSON mode). To bound cost and focus the
model, we first apply a deterministic ``project gate'' --- a regular-expression
filter for project-signal vocabulary --- and send only the windows of text
around matches, capped at roughly 14{,}000 characters per call. The model is
instructed to return a strict-JSON array of project objects, each with a
\texttt{project\_name}, a \texttt{project\_type} drawn from a fixed 21-category
taxonomy (Table~\ref{tab:bytype}), a one-sentence grounded \texttt{description},
an \texttt{objective}, a \texttt{status} (planned / in-progress / completed /
mentioned), an optional \texttt{amount\_usd}, and arrays of
\texttt{locations}, \texttt{technologies}, \texttt{partners}, \texttt{benefits},
\texttt{risks}, and \texttt{years}, together with a self-reported
\texttt{confidence} in $[0,1]$. The system prompt forbids speculation and
instructs the model to prefer a small number of high-quality records over many
vague ones; records below a confidence of 0.45 are discarded.

This stage produced \textbf{39{,}930 project mentions}. Because the model emits
typed quantities, downstream analysis requires no further parsing: amounts are
numeric, statuses and types are categorical, and list fields are stored as
native arrays in a columnar database.

\subsection{Resolution, graph, and embeddings}
The same physical project is typically mentioned in several filings across
years. We resolve mentions into projects by grouping on firm, project type, and
a normalized project anchor, and we assign each project a timeline, a canonical
location, an aggregated announced amount, first/last-seen dates, and an average
extraction confidence. Resolution collapses the 39{,}930 mentions into
\textbf{19{,}227 distinct projects}; the mention timeline is retained for the
survival analysis. From the resolved projects we build a directed knowledge
graph with \textbf{32{,}226 nodes} (19{,}227 projects, 7{,}640 partners, 3{,}080
locations, 1{,}779 technologies, 500 companies) and \textbf{33{,}176 edges}
(\texttt{owns}, \texttt{uses}, \texttt{located\_in}), and we encode every
project's text into a 384-dimensional sentence embedding
(MiniLM, \texttt{all-MiniLM-L6-v2}) for semantic search and clustering.

\subsection{The economics of extraction}
A defining feature of the approach is its cost structure. Each section incurs at
most one model call, and the deterministic project gate suppresses calls on
sections with no project vocabulary; the focused-window construction caps the
prompt at roughly 14{,}000 characters regardless of section length. The marginal
cost of processing a section is therefore on the order of a tenth of a U.S.\
cent, and the full-corpus pass over 153{,}894 sections completed in a few hours
on commodity hardware with bounded memory, batched into 34 firm-cohorts with
checkpointing so that the run is fully resumable. This matters substantively:
the feasibility of a \emph{census} (as opposed to a sample) is precisely what
distinguishes LLM extraction from earlier hand-coding or rule-based approaches,
and it is what allows us to populate a 7{,}370-observation firm-year panel with
project-level granularity. The same OpenAI-compatible interface can target a
self-hosted open-weights model, so the pipeline is not tied to any single
provider, and the per-section cost can in principle be driven to near zero on
local infrastructure.

\subsection{Analysis datasets}
For econometric analysis we derive three datasets from the resolved store: (i) a
project-level cross-section (19{,}227 rows) with engineered flags for AI and
digital orientation, project age, and technology counts; (ii) a mention-level
event table (39{,}930 rows) with filing dates and forms; and (iii) a balanced
\textbf{firm-year panel} of \textbf{7{,}370} observations spanning 500 firms and
the years 2010--2025 (the partial year 2026 is excluded from the panel), in
which each firm-year records the count of newly announced projects, the number
that are AI- or digital-oriented, announced capital, and project-type
composition. Announcement year is taken from the project's first observed year.
\tab{tab:summary} reports summary statistics for the main variables.

\begin{table}[H]
\centering
\caption{Summary statistics. Amounts are in millions of USD; duration is the
span between a project's first and last appearance in filings.}
\label{tab:summary}
\small
\input{../tables/tab_summary_stats}
\end{table}

% =====================================================================
\section{Descriptive Evidence and Stylized Facts}\label{sec:descriptives}
% =====================================================================
We begin with the broad contours of the corporate project economy.

\paragraph{Volume and trend.} \fig{fig:mentions} plots the number of extracted
project mentions per year by filing form. The series is large and stable ---
between roughly 2{,}000 and 2{,}800 mentions per year --- with a mild upward
drift (a compound annual growth rate of about 2.2\% over 2010--2025) and a
visible step-up after 2020. 8-K filings (event-driven disclosures) and 10-K
filings (annual reports) are the dominant sources, consistent with the fact that
new initiatives are announced in 8-Ks and reviewed comprehensively in 10-Ks.

\begin{figure}[H]
\centering
\includegraphics[width=0.78\textwidth]{fig_mentions_by_year.pdf}
\caption{Project mentions extracted per year, by filing form. The corpus is
roughly balanced across years, with 8-K and 10-K filings supplying the bulk of
project signals.}
\label{fig:mentions}
\end{figure}

\paragraph{Composition.} \fig{fig:typedist} and \tab{tab:bytype} show how
projects distribute across the 21-category taxonomy. M\&A integration is by far
the most frequent category (7{,}274 projects), followed by capital-expenditure
programs (2{,}085), plant construction (1{,}252), and partnerships (1{,}239).
The long tail includes specialized categories such as drug pipelines, ERP
implementations, and data-center builds. The dominance of M\&A integration
reflects both the economic importance of acquisitions among large firms and the
prominence of integration narratives in their filings.

\begin{figure}[H]
\centering
\includegraphics[width=0.74\textwidth]{fig_type_distribution.pdf}
\caption{Distribution of resolved projects across the 21-category taxonomy.}
\label{fig:typedist}
\end{figure}

\begin{table}[H]
\centering
\caption{Projects by type: counts, aggregate announced capital, AI share, and
mean extraction confidence. Capital is the sum of disclosed amounts and is an
upper bound (see Section~\ref{sec:limitations}).}
\label{tab:bytype}
\small
\input{../tables/tab_by_type}
\end{table}

\paragraph{Sectoral structure.} The composition of projects varies sharply
across sectors. \fig{fig:heatmap} row-normalizes the project-type mix within each
GICS sector. Utilities and Energy concentrate in plant construction and
energy-transition projects; Information Technology and Communication Services in
product launches and technology deployment; Health Care in R\&D and drug
pipelines; and Financials and Consumer sectors in M\&A integration and
geographic expansion. \tab{tab:bysector} reports the sector-level totals: every
sector is well populated, with Industrials (2{,}708 projects), Health Care
(2{,}475), and Information Technology (2{,}225) the most project-active.

\begin{figure}[H]
\centering
\includegraphics[width=0.92\textwidth]{fig_sector_type_heatmap.pdf}
\caption{Project-type composition by GICS sector (row-normalized within-sector
shares). Darker cells indicate a larger share of a sector's projects in that
category.}
\label{fig:heatmap}
\end{figure}

\begin{table}[H]
\centering
\caption{Sector breakdown: number of firms, projects, aggregate announced
capital, and AI share.}
\label{tab:bysector}
\small
\input{../tables/tab_by_sector}
\end{table}

\paragraph{Capital and geography.} \fig{fig:capex} overlays announced capital and
project counts by announcement year. Announced capital is lumpy and dominated by
a handful of very large transactions in particular years, whereas the project
\emph{count} is smooth, underscoring why we model the extensive and intensive
margins separately. \fig{fig:locations} shows the geographic concentration of
projects, with the United States and a familiar set of global hubs accounting
for most located projects.

\begin{figure}[H]
\centering
\begin{subfigure}{0.52\textwidth}
\includegraphics[width=\textwidth]{fig_capex_by_year.pdf}
\caption{Announced capital and project count by year.}
\label{fig:capex}
\end{subfigure}\hfill
\begin{subfigure}{0.46\textwidth}
\includegraphics[width=\textwidth]{fig_top_locations.pdf}
\caption{Top 15 project locations.}
\label{fig:locations}
\end{subfigure}
\caption{Capital and geography of the corporate project economy.}
\end{figure}

\subsection{Six stylized facts}
The descriptive evidence can be distilled into six stylized facts that motivate
the formal analysis and that, to our knowledge, have not previously been
documented at this granularity:
\begin{enumerate}
\item \textbf{Project activity is voluminous and steadily growing.} Large public
firms disclose on the order of 2{,}000--2{,}800 distinct project signals per year,
growing at roughly 2.2\% annually, with a level step-up after 2020.
\item \textbf{Acquisitions dominate the project mix.} M\&A integration is the
single largest category (38\% of projects), reflecting both its economic weight
and its disclosure salience among incumbents.
\item \textbf{Sectors specialize sharply.} Project composition is highly
sector-specific --- energy and utilities in physical builds, technology in
deployments, health care in pipelines --- with little cross-sector convergence in
composition.
\item \textbf{Capital is concentrated in physical projects.} Disclosed amounts
are heavily right-skewed and dominated by capital-intensive physical families;
the median disclosed project is \$718m but the distribution spans four orders of
magnitude.
\item \textbf{AI is frequent but small.} AI-oriented projects are proliferating
rapidly in \emph{count} yet remain individually modest in disclosed capital.
\item \textbf{Disclosure salience tracks scale.} The number of filings that
mention a project rises monotonically with its announced amount, validating both
the resolution procedure and the economic content of the extracted records.
\end{enumerate}

% =====================================================================
\section{Econometric Framework}\label{sec:framework}
% =====================================================================
The panel structure of the data supports a sequence of complementary models. We
summarize the specifications here and report results in the following sections.

\paragraph{(i) Count models.} Let $y_{it}$ denote the number of projects firm
$i$ announces in year $t$. We model the conditional mean as
\begin{equation}
\mathbb{E}[y_{it}\mid x_{it}] = \exp\!\big(\alpha_i + \delta_t + x_{it}'\beta\big),
\label{eq:poisson}
\end{equation}
estimated by pseudo-Poisson maximum likelihood (PPML) with firm effects
$\alpha_i$ and year effects $\delta_t$. PPML is consistent under arbitrary
distributional misspecification given a correctly specified mean, which makes it
the natural workhorse for non-negative count outcomes. Because raw project counts
are overdispersed, we also estimate a negative-binomial model and a pooled
Poisson with sector and year effects. Standard errors are clustered by firm.

\paragraph{(ii) Hedonic capital regressions.} For the 9{,}321 projects with a
disclosed amount we estimate
\begin{equation}
\log(\text{amount}_{j}) = \gamma_{\text{type}(j)} + \theta_{\text{status}(j)}
+ \lambda\,\text{AI}_j + \mu' z_j + \alpha_{i(j)} + \delta_{t(j)} + \varepsilon_j,
\label{eq:hedonic}
\end{equation}
where $z_j$ collects project covariates (technology count, filing salience) and
$\alpha_{i(j)}$, $\delta_{t(j)}$ are firm and year effects. The coefficients
$\gamma_{\text{type}}$ are interpreted as log-capital premia relative to a
baseline category (M\&A integration).

\paragraph{(iii) Diffusion.} We model AI adoption at the firm-year level with a
logistic regression for the probability of announcing at least one AI project,
and we fit a logistic (S-curve) diffusion model
$F(t)=K/(1+e^{-r(t-t_0)})$ to the cumulative share of firms that have ever
announced an AI project, recovering the saturation $K$, speed $r$, and
inflection $t_0$.

\paragraph{(iv) Difference-in-differences.} To test whether COVID-19
differentially accelerated digitalization, we define treatment as above-median
\emph{physical-project intensity} in the pre-period (2015--2019) and estimate a
dynamic event study,
\begin{equation}
\text{DigitalShare}_{it} = \alpha_i + \delta_t
+ \sum_{\tau\neq 2019}\pi_\tau\,(\text{Treat}_i\times \mathbf{1}[t=\tau])
+ u_{it},
\label{eq:did}
\end{equation}
with 2019 omitted, alongside a canonical $2\times2$ specification. The leads
$\pi_\tau$ for $\tau<2019$ provide a parallel-trends test.

\paragraph{(v) Survival.} We treat each multi-filing project as a spell that
begins at first appearance and ``fails'' when it reaches completed status,
right-censoring projects last seen as planned or in progress. We estimate
Kaplan--Meier survival functions by project family and a Cox proportional-hazards
model with project covariates.

\paragraph{(vi) Networks and representations.} We analyze the degree
distribution of the knowledge graph, detect communities in a technology
co-occurrence network by greedy modularity maximization, and cluster the
384-dimensional project embeddings with $K$-means, selecting $K$ by silhouette.

\paragraph{(vii) Prediction.} Finally, we train a gradient-boosted tree
classifier to predict ``mega-projects'' (announced amount above the sample
median) from text-derived structural features, evaluated by 5-fold
cross-validated ROC-AUC with permutation importance.

% =====================================================================
\section{The Extensive Margin: Count Models}\label{sec:counts}
% =====================================================================
\tab{tab:counts} reports the count-model estimates. Three results stand out.

First, project activity is \textbf{strongly persistent}. In the pooled Poisson
the coefficient on the lagged project count is $0.100$ ($p<0.01$); within firms
(column 2, firm and year effects) it remains positive and highly significant at
$0.028$, an incidence-rate ratio of about $1.03$ per additional prior-year
project. Persistence survives the inclusion of firm effects, indicating that it
reflects within-firm dynamics --- ongoing programs spawning follow-on
initiatives --- rather than fixed cross-firm heterogeneity alone.

Second, there is a \textbf{significant post-2020 increase} in announcement
intensity. In specifications with sector (rather than firm) effects, the
post-COVID indicator carries a coefficient of $0.208$ ($p<0.01$), implying
roughly a 23\% higher announcement rate after 2020, even as the secular year
trend is mildly negative ($-0.031$ per year). The juxtaposition --- a negative
trend but a positive level shift after 2020 --- suggests a structural break
rather than a smooth acceleration.

Third, the data are \textbf{strongly overdispersed}. A Cameron--Trivedi auxiliary
regression rejects equidispersion overwhelmingly ($t=18.3$, $p<10^{-70}$), and
the negative-binomial dispersion parameter is $\hat\alpha=0.50$. Overdispersion
justifies the negative-binomial specification (column 3) and cautions against
naive Poisson standard errors; our clustered, PPML-based inference is robust to
it. \fig{fig:countcoef} visualizes the within-firm effects of persistence and
accumulated AI experience.

\begin{table}[H]
\centering
\caption{Panel count models of project-announcement intensity. Dependent
variable: number of projects announced by firm $i$ in year $t$. Cluster-robust
standard errors (by firm) in parentheses.
$^{*}p<0.1$, $^{**}p<0.05$, $^{***}p<0.01$.}
\label{tab:counts}
\small
\input{../tables/tab_count_models}
\end{table}

\begin{figure}[H]
\centering
\includegraphics[width=0.66\textwidth]{fig_count_coef.pdf}
\caption{Within-firm drivers of project-announcement intensity (Poisson with
firm and year effects, 95\% confidence intervals).}
\label{fig:countcoef}
\end{figure}

The accumulated-AI-experience term is positive and significant in the pooled and
negative-binomial models but attenuates to insignificance within firms,
indicating that the cross-sectional association between AI experience and
overall project activity is largely a firm-level (selection) phenomenon rather
than a within-firm causal channel --- AI-active firms are simply more
project-active in general.

% =====================================================================
\section{The Intensive Margin: What Drives Announced Capital?}\label{sec:amount}
% =====================================================================
\tab{tab:amount} reports the hedonic log-capital regressions of
equation~\eqref{eq:hedonic}. With firm and year effects (column 2) the model
explains $R^2=0.318$ of the variation in log announced amount --- substantial for
project-level data --- and the project-type fixed effects are jointly the
dominant component.

Several findings are noteworthy. The number of distinct technologies attached to
a project is positively associated with its size, consistent with larger
projects being more technologically complex, though the within-firm estimate is
imprecise. Filing salience (the log number of filings mentioning a project) is
strongly positively associated with announced capital: larger projects are
discussed more often, which both validates the resolution procedure and reflects
genuine economic salience. Project \emph{status} matters: relative to the
omitted category, projects observed as in-progress or completed differ
systematically in disclosed scale from those merely planned or mentioned.

Strikingly, AI projects are \emph{not} larger than comparable projects; if
anything the point estimate is negative ($-0.18$ in log points) and statistically
insignificant ($p=0.48$). This is an economically meaningful null: in this
period AI initiatives are numerous and strategically prominent but individually
modest in disclosed capital relative to plant builds, acquisitions, and
infrastructure programs. \fig{fig:amountcoef} plots the estimated capital premia
by project type relative to M\&A integration; capital-intensive families ---
infrastructure, plant construction, data centers, energy transition --- carry
the largest premia, while software, partnership, and cost-reduction projects sit
below the baseline.

\begin{table}[H]
\centering
\caption{Hedonic regressions for log announced capital (projects with a
disclosed amount, $N=9{,}321$). Baseline project type: M\&A integration.
Cluster-robust SE (by firm). $^{*}p<0.1$, $^{**}p<0.05$, $^{***}p<0.01$.}
\label{tab:amount}
\small
\input{../tables/tab_amount_models}
\end{table}

\begin{figure}[H]
\centering
\includegraphics[width=0.74\textwidth]{fig_amount_coef.pdf}
\caption{Announced-capital premia by project type relative to M\&A integration
(firm and year effects, 95\% CI). Capital-intensive physical projects command
the largest premia.}
\label{fig:amountcoef}
\end{figure}

% =====================================================================
\section{Technology Diffusion: The Rise of Corporate AI}\label{sec:diffusion}
% =====================================================================
\fig{fig:diffusion} summarizes the diffusion analysis. The left panel plots the
cumulative share of firms that have ever announced an AI project together with
the fitted logistic curve. The estimated saturation is $K=0.42$, the speed is
$r=0.21$, and the inflection point is \textbf{$t_0=2023.1$} --- precisely when
generative AI entered the corporate mainstream. The right panel shows the AI
\emph{share} of all new project announcements rising more than sixfold, from
0.7\% in 2015 to 4.7\% in 2025, with a clear acceleration after 2022.

\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{fig_ai_diffusion.pdf}
\caption{AI-project diffusion. Left: cumulative share of firms ever announcing an
AI project, with a fitted logistic (S-curve) diffusion model; the vertical line
marks the estimated inflection year ($\approx$2023). Right: AI share of all new
project announcements by year.}
\label{fig:diffusion}
\end{figure}

The firm-year logit (\tab{tab:ailogit}) confirms a positive secular trend in the
probability of AI announcement (average marginal effect $+0.25$ percentage points
per year) and a positive but imprecisely estimated post-COVID shift. Larger
project portfolios are associated with higher AI-adoption probability,
consistent with scale and absorptive capacity facilitating experimentation. That
the diffusion is still on the steep part of its S-curve --- the fitted saturation
of 0.42 is far from unity --- implies that corporate AI adoption, measured by
disclosed projects, had substantial room to run at the end of the sample.

\begin{table}[H]
\centering
\caption{Firm-year logit for the probability of announcing $\geq1$ AI project.
Average marginal effects in the right column. Cluster-robust SE (by firm).
$^{*}p<0.1$, $^{**}p<0.05$, $^{***}p<0.01$.}
\label{tab:ailogit}
\small
\input{../tables/tab_ai_logit}
\end{table}

% =====================================================================
\section{COVID-19 and Digitalization: A Difference-in-Differences Test}\label{sec:did}
% =====================================================================
A widespread narrative holds that the pandemic forced a rapid digital
convergence, with physically intensive incumbents racing to catch up. We test
this directly. Treatment is defined as above-median physical-project intensity in
2015--2019 (223 treated vs.\ 250 control firms); the outcome is the digital
share of a firm-year's project announcements.

\fig{fig:did} plots the dynamic treatment effects from
equation~\eqref{eq:did}. The pre-2019 leads are jointly indistinguishable from
zero ($F$-test $p=0.78$), supporting the parallel-trends assumption, and the
post-2020 lags hover around zero. The canonical $2\times2$ estimate
(\tab{tab:did}) is $-0.010$ (s.e.\ $0.014$, $p=0.49$) --- a precisely estimated
null. We find \emph{no evidence} that physically intensive incumbents
differentially accelerated their digitalization after COVID-19 relative to
already-digital firms. If anything, both groups digitalized in parallel. This
tempers the convergence narrative: the pandemic may have raised digital activity
broadly, but it did not measurably close the cross-firm gap in project
composition.

\begin{figure}[H]
\centering
\includegraphics[width=0.78\textwidth]{fig_did_event.pdf}
\caption{Event study of COVID-19 and digitalization. Points are
treatment$\times$year coefficients on the digital share of announcements
relative to 2019 (95\% CI). Flat pre-trends and post-period coefficients near
zero indicate no differential digital acceleration of physically intensive
incumbents.}
\label{fig:did}
\end{figure}

\begin{table}[H]
\centering
\caption{Difference-in-differences: COVID-19 and digital project share.
Cluster-robust SE (by firm).}
\label{tab:did}
\small
\input{../tables/tab_did}
\end{table}

% =====================================================================
\section{Project Lifecycles: Survival Analysis}\label{sec:survival}
% =====================================================================
How long do corporate projects live? Among the 5{,}905 projects tracked across at
least two filings (2{,}664 of which reach completed status), median lifecycle
length varies dramatically by family. \fig{fig:km} plots Kaplan--Meier survival
curves --- the probability that a project has \emph{not} yet been completed as a
function of months since its first appearance. M\&A integrations resolve fastest
(median $\approx$49 months), capital and construction projects next ($\approx$84
months), while R\&D / product programs ($\approx$132 months) and especially
digital / AI initiatives ($\approx$144 months) persist far longer. The ordering
is intuitive: acquisitions have a natural completion event, whereas digital
transformation and AI are open-ended, continuously renewed programs that rarely
declare victory.

\begin{figure}[H]
\centering
\includegraphics[width=0.78\textwidth]{fig_survival_km.pdf}
\caption{Kaplan--Meier survival curves by project family. M\&A integrations
complete fastest; digital and AI programs persist longest.}
\label{fig:km}
\end{figure}

The Cox proportional-hazards model (\tab{tab:cox}) quantifies the covariate
effects. Larger projects complete more slowly: a one-log-point increase in
announced amount lowers the completion hazard by about 4\% (hazard ratio $0.96$,
$p<10^{-10}$), consistent with bigger initiatives being more complex and
longer-lived. The AI indicator is not significantly associated with the
completion hazard once project type and scale are controlled (HR $1.06$,
$p=0.74$), reinforcing that the long lifecycles of AI programs operate through
their \emph{type} (open-ended digital programs) rather than an independent AI
effect.

\begin{table}[H]
\centering
\caption{Cox proportional-hazards model of time to project completion (hazard
ratios). Project-type controls included. $^{*}p<0.1$, $^{**}p<0.05$,
$^{***}p<0.01$.}
\label{tab:cox}
\small
\input{../tables/tab_cox}
\end{table}

% =====================================================================
\section{The Innovation Network}\label{sec:network}
% =====================================================================
The knowledge graph offers a structural view of the corporate project economy.
\fig{fig:degree} plots its degree distribution on log-log axes: the distribution
is heavy-tailed with an estimated tail slope of $-1.55$, the signature of a
scale-free-like structure in which a few hub nodes (large diversified firms,
ubiquitous technologies) connect to a great many projects while most nodes have
low degree.

\begin{figure}[H]
\centering
\includegraphics[width=0.64\textwidth]{fig_degree_dist.pdf}
\caption{Degree distribution of the 32{,}226-node project knowledge graph
(log-log). The heavy tail is consistent with a scale-free-like organization.}
\label{fig:degree}
\end{figure}

To uncover technology ecosystems we build a co-occurrence network in which two
technologies are linked when they appear in the same project, restricted to the
60 most frequent technologies. Greedy modularity maximization recovers
\textbf{four communities} with a high modularity of \textbf{0.52}, indicating
well-separated ecosystems. \fig{fig:technet} visualizes the network;
\tab{tab:network} lists the most central technologies. Strikingly, the highest
eigenvector-centrality nodes are all \emph{energy-transition} technologies ---
solar (0.57), renewable energy (0.49), wind (0.44), battery storage (0.26),
natural gas, energy storage --- which form the gravitational center of the
corporate-innovation graph. Software/SaaS, cloud, AI, and semiconductors form a
second, digital ecosystem; manufacturing/automation and connectivity form
others. The centrality of the energy transition reflects both its capital
intensity and its cross-sector reach: utilities, industrials, materials, and
consumer firms all participate in it, whereas digital technologies, though
pervasive, are more modular and less tightly co-bundled.

\begin{figure}[H]
\centering
\includegraphics[width=0.90\textwidth]{fig_tech_network.pdf}
\caption{Technology co-occurrence network (top 60 technologies, edges for
co-use in $\geq2$ projects). Node color denotes community; node size scales with
project frequency. Four ecosystems emerge (modularity 0.52), with
energy-transition technologies most central.}
\label{fig:technet}
\end{figure}

\begin{table}[H]
\centering
\caption{Most central technologies in the co-occurrence network (eigenvector
centrality) with project frequency and community membership.}
\label{tab:network}
\small
\input{../tables/tab_network}
\end{table}

% =====================================================================
\section{Latent Themes: Embedding the Project Corpus}\label{sec:embeddings}
% =====================================================================
The formal taxonomy is imposed at extraction time. To recover \emph{emergent}
structure we cluster the 384-dimensional MiniLM project embeddings with
$K$-means, selecting $K=6$ by silhouette (\fig{fig:silhouette}). Although the
silhouette values are modest --- expected for dense sentence embeddings of
heterogeneous business text --- the resulting clusters are interpretable and
economically coherent (\tab{tab:clusters}). They span an M\&A/technology cluster,
a health-care partnership and AI cluster, an energy/utilities construction
cluster centered on renewables and storage, an industrial capex/ERP cluster, a
consumer geographic-expansion cluster, and a consumer-staples cost-reduction and
automation cluster.

\begin{figure}[H]
\centering
\includegraphics[width=0.82\textwidth]{fig_embedding_tsne.pdf}
\caption{Two-dimensional t-SNE projection of project embeddings (stratified
sample of 5{,}000 projects), colored by $K$-means cluster. Themes separate while
overlapping at the boundaries, reflecting the continuous semantic space of
corporate initiatives.}
\label{fig:tsne}
\end{figure}

\begin{figure}[H]
\centering
\includegraphics[width=0.56\textwidth]{fig_silhouette.pdf}
\caption{Silhouette score versus number of clusters; $K=6$ is selected.}
\label{fig:silhouette}
\end{figure}

\begin{table}[H]
\centering
\caption{Embedding clusters: size, dominant project type, top technologies, and
aggregate announced capital. Clusters cross sector boundaries, capturing themes
the taxonomy alone does not.}
\label{tab:clusters}
\small
\input{../tables/tab_clusters}
\end{table}

\fig{fig:tsne} shows the t-SNE projection of a sample of projects. The clusters
separate visibly while overlapping at their margins, consistent with corporate
initiatives occupying a continuous semantic space rather than discrete bins.
That the unsupervised themes largely \emph{recover} the supervised taxonomy ---
while recombining it across sectors --- is an internal-validity check on the
extraction: the embeddings and the LLM labels agree on the broad organization of
the corpus.

% =====================================================================
\section{Predicting Mega-Projects}\label{sec:ml}
% =====================================================================
Can the scale of a project be predicted from its qualitative description alone?
We train a gradient-boosted tree classifier to distinguish ``mega-projects''
(announced amount above the \$718m median) from smaller ones, using only
text-derived structural features --- project type, sector, status, technology
count, filing and mention counts, AI flag, presence of a location, announcement
year, and extraction confidence --- with no leakage of the amount itself.

The model attains a 5-fold cross-validated \textbf{ROC-AUC of 0.743} and average
precision of 0.722 (\fig{fig:roc}, \tab{tab:ml}). Permutation importance ranks
\emph{project type} first, followed by \emph{announcement year}, \emph{sector},
and \emph{filing count}. The predictive value of year and sector indicates that
mega-projects cluster in particular periods and industries (e.g., recent
data-center and energy build-outs), while the importance of filing count
confirms that disclosure salience tracks economic scale. That qualitative text
features alone recover three-quarters of the discriminative information about
project scale is itself evidence that the extracted attributes carry real
economic content.

\begin{figure}[H]
\centering
\includegraphics[width=0.95\textwidth]{fig_ml_roc.pdf}
\caption{Mega-project classifier. Left: 5-fold cross-validated ROC curve
(AUC = 0.74). Right: permutation feature importance (drop in AUC).}
\label{fig:roc}
\end{figure}

\begin{table}[H]
\centering
\caption{Gradient-boosted mega-project classifier: feature-group importance and
performance ($N=9{,}321$ projects with disclosed amounts).}
\label{tab:ml}
\small
\input{../tables/tab_ml}
\end{table}

% =====================================================================
\section{Robustness and Validation}\label{sec:robustness}
% =====================================================================
We summarize the checks that support the analysis.

\paragraph{Extraction quality.} The mean extraction confidence is 0.78 and the
median 0.90; records below 0.45 are dropped at the source. The agreement between
the unsupervised embedding clusters and the supervised taxonomy
(Section~\ref{sec:embeddings}) provides an internal-consistency check: two
independent representations of the same text organize it the same way. Spot
auditing of high-amount records confirms that disclosed figures are grounded in
the underlying text rather than hallucinated, a behavior the JSON-mode prompt
explicitly enforces by forbidding unsupported numbers.

\paragraph{Estimator robustness.} Because project counts are overdispersed, we
report pooled Poisson, firm-fixed-effects PPML, and negative-binomial estimates
side by side; the sign and significance of the persistence and post-2020 effects
are stable across all three (\tab{tab:counts}). All panel models use
cluster-robust standard errors by firm. The hedonic capital results are
qualitatively identical under sector effects and under firm effects
(\tab{tab:amount}), and the project-type ranking of premia is unchanged.

\paragraph{Parallel trends.} The credibility of the difference-in-differences
rests on parallel pre-trends, which we cannot reject ($F$-test $p=0.78$);
the dynamic event-study coefficients are flat before 2020
(\fig{fig:did}).

\paragraph{Level vs.\ composition.} Throughout we privilege within-firm and
relative comparisons over levels. The aggregate announced-capital figure is
sensitive to a small number of very large transactions and to repeated mentions
of multi-year programs; none of our regression conclusions depends on it, because
the hedonic model uses log amounts with firm and year effects and the count
models use the extensive margin.

\paragraph{Sample.} The 500-firm corpus is weighted toward large, mature public
companies, which is appropriate for studying the project economy of incumbents
but limits external validity to small and private firms. Results should be read
as describing large-cap U.S.\ public companies over 2010--2026.

% =====================================================================
\section{Limitations}\label{sec:limitations}
% =====================================================================
Several limitations warrant emphasis. \emph{First}, our census measures
\emph{disclosure of} projects, not projects themselves; initiatives that are
never discussed in filings are invisible, and disclosure incentives vary across
firms and over time. \emph{Second}, announced amounts are self-reported,
inconsistently scoped (a figure may be total program cost, annual spend, or deal
value), and frequently absent (only 48.5\% of projects carry a number); we
mitigate this with logs, fixed effects, and an explicit focus on relative
magnitudes. \emph{Third}, LLM extraction can err --- both false positives
(spurious projects) and false negatives (missed initiatives) --- and although
the confidence filter and grounding constraints reduce these, we do not have a
hand-labeled gold standard large enough to estimate precision and recall
precisely; constructing one is a priority for future work. \emph{Fourth},
resolution merges mentions heuristically; aggressive merging understates project
counts while conservative merging overstates them. \emph{Fifth}, the survival
analysis infers completion from disclosed status, which is itself a noisy signal.
Finally, our identification in the difference-in-differences is quasi-experimental
and the treatment is defined on pre-period behavior; we interpret the null
cautiously, as the absence of a \emph{differential} effect rather than the
absence of any pandemic effect.

% =====================================================================
\section{Discussion and Implications}\label{sec:discussion}
% =====================================================================
The results have implications for research methodology, for the empirical study
of corporate investment, and for practice.

\paragraph{For methodology.} The paper is, in part, an existence proof: a single
researcher, using a commodity LLM and open-source statistical tooling, can
construct a population-scale, project-level panel from the unstructured
regulatory record and subject it to the full apparatus of modern applied
econometrics. The binding constraint in text-based empirical work has shifted
from \emph{extraction} to \emph{validation and design}. Because extraction is now
cheap and reproducible, the scarce inputs are a credible ground-truth benchmark
and a research design that isolates economic mechanisms from disclosure
artifacts. Our internal-consistency checks --- the agreement of unsupervised
embeddings with the supervised taxonomy, the monotone relation between disclosure
salience and scale, and the stability of estimates across estimators --- are a
template, but they are not a substitute for a hand-labeled gold standard, which
remains the most valuable complementary investment.

\paragraph{For the study of investment.} Observing the project as a unit allows a
clean separation of the extensive margin (how many initiatives a firm pursues)
from the intensive margin (how large each is) and the compositional margin (which
kinds). Our evidence suggests these margins are governed by different forces:
the extensive margin is persistent and firm-driven with a macro level shift after
2020; the intensive margin is dominated by project \emph{type} and physical
capital intensity; and the compositional margin is strongly sectoral and, in the
case of AI, diffusing along a classic S-curve. A natural agenda is to link these
margins to financing conditions, governance, and ex-post outcomes --- for
instance, whether project-announcement intensity predicts future capital
expenditure, productivity, or returns, and whether the market prices the project
portfolio beyond aggregate guidance.

\paragraph{For the AI-diffusion debate.} Our disclosure-based measure complements
job-posting and patent-based measures and adds a distinct vantage point: the
\emph{strategic} framing of AI as a corporate project. The 2023 inflection and
the still-low saturation imply that, as of the sample's end, corporate AI was
early in its diffusion and that the count surge had not yet translated into large
disclosed capital commitments --- a pattern consistent with an experimentation
phase preceding heavy build-out. Whether the subsequent data-center and compute
investments convert AI from a high-count, low-capital category into a high-capital
one is a question the same pipeline can answer as new filings arrive.

\paragraph{For practice.} For analysts and policymakers, a structured project
census enables monitoring that aggregate statistics cannot: tracking
energy-transition build-outs across sectors in near real time, mapping
partnership and supply-chain networks, or flagging firms whose project mix is
shifting. The knowledge graph and embeddings make the corpus queryable both
structurally and semantically, turning a static disclosure archive into a living
intelligence layer.

% =====================================================================
\section{Conclusion}\label{sec:conclusion}
% =====================================================================
We have shown that the narrative sections of SEC filings, long treated as
qualitative context, can be converted into structured, research-grade panel data
on corporate projects at scale and at trivial marginal cost, and that the
resulting data support a rich program of empirical analysis. From 153{,}894
filing sections we extracted 39{,}930 project mentions and resolved them into
19{,}227 projects across 500 firms, building a knowledge graph and semantic
embeddings on top. The econometric analysis documents strong within-firm
persistence of project activity and a significant post-2020 level shift; large
capital premia for physically intensive project families and a surprising null
for the scale of AI projects; an AI-diffusion path inflecting in 2023; a
precisely estimated null for differential pandemic-driven digitalization;
project lifecycles ranging from four years for acquisitions to twelve for digital
programs; an energy-transition technology ecosystem at the center of the
innovation network; and a text-only classifier that predicts project scale with
an AUC of 0.74.

Beyond the specific findings, the broader message is that LLM-based information
extraction reshapes what is measurable in empirical finance and the economics of
innovation. The project --- the actual unit of corporate decision-making --- can
now be observed at population scale. Natural next steps include constructing a
human-labeled validation set to benchmark extraction precision and recall,
extending the corpus to additional forms (S-1, DEF 14A, 20-F) and to the full
universe of filers, linking projects to ex-post financial and market outcomes,
and using the project panel to revisit classic questions about investment,
financing constraints, and the returns to corporate innovation.

\clearpage
\appendix
% =====================================================================
\section{Project Taxonomy and Variable Definitions}\label{app:vars}
% =====================================================================
The extraction taxonomy comprises 21 mutually exclusive project types:
manufacturing expansion, plant construction, data center, cloud migration, AI
initiative, digital transformation, ERP implementation, R\&D program, product
launch, drug pipeline, M\&A integration, geographic expansion, capex program,
supply chain, automation, sustainability, energy transition, infrastructure,
cost reduction, partnership, and technology deployment. Key engineered variables
are defined as follows. \textbf{AI project}: a project whose type is
\emph{AI initiative} or whose technology list contains an AI term (AI, artificial
intelligence, machine/deep learning, generative, LLM, neural).
\textbf{Digital project}: an AI project or one whose type is among cloud
migration, ERP, data center, automation, digital transformation, or technology
deployment. \textbf{Announcement year}: the first calendar year in which the
project appears in any filing. \textbf{Duration}: days between a project's first
and last appearance. \textbf{Announced amount}: the resolved (aggregated) USD
figure disclosed for the project, when present.

% =====================================================================
\section{Pipeline and Reproducibility}\label{app:repro}
% =====================================================================
The extraction backend is OpenAI-compatible and was run with
\texttt{gpt-4o-mini} at temperature 0 in JSON mode. The pipeline is incremental
and resumable: mention identifiers are stable SHA-1 hashes of
(accession, section, type, anchor), so re-runs use \texttt{INSERT OR REPLACE}
semantics and never double-count. The full corpus run processed the 153{,}894
sections in 34 firm-batches with checkpointing; resolution, graph construction,
embedding, and all analyses are deterministic given the resolved store. Analysis
code is organized as ten numbered scripts (dataset construction; descriptives;
count models; hedonic models; diffusion; difference-in-differences; survival;
network; embeddings; prediction), each emitting figures, \LaTeX\ table fragments,
and a machine-readable results file. Estimation uses
\texttt{statsmodels} (Poisson, negative binomial, logit, OLS, Cox PHReg),
\texttt{scipy} (logistic curve fitting), \texttt{networkx} (community detection
and centrality), and \texttt{scikit-learn} ($K$-means, t-SNE, gradient boosting,
cross-validation).

\begin{thebibliography}{99}
\small
\bibitem[Acemoglu et al.(2022)]{acemoglu2022artificial} Acemoglu, D., Autor, D.,
Hazell, J., \& Restrepo, P. (2022). Artificial intelligence and jobs: Evidence
from online vacancies. \emph{Journal of Labor Economics}, 40(S1), S293--S340.

\bibitem[Babina et al.(2024)]{babina2024artificial} Babina, T., Fedyk, A., He,
A., \& Hodson, J. (2024). Artificial intelligence, firm growth, and product
innovation. \emph{Journal of Financial Economics}, 151, 103745.

\bibitem[Bass(1969)]{bass1969new} Bass, F. M. (1969). A new product growth model
for consumer durables. \emph{Management Science}, 15(5), 215--227.

\bibitem[Bybee et al.(2023)]{bybee2023ghost} Bybee, L., Kelly, B., \& Su, Y.
(2023). Narrative-driven fluctuations in sentiment: Evidence from large language
models. Working paper.

\bibitem[Crouzet \& Eberly(2019)]{crouzet2019understanding} Crouzet, N., \&
Eberly, J. (2019). Understanding weak capital investment: The role of market
concentration and intangibles. \emph{NBER Working Paper} 25869.

\bibitem[Fazzari et al.(1988)]{fazzari1988financing} Fazzari, S. M., Hubbard,
R. G., \& Petersen, B. C. (1988). Financing constraints and corporate
investment. \emph{Brookings Papers on Economic Activity}, 1988(1), 141--206.

\bibitem[Gentzkow et al.(2019)]{gentzkow2019text} Gentzkow, M., Kelly, B., \&
Taddy, M. (2019). Text as data. \emph{Journal of Economic Literature}, 57(3),
535--574.

\bibitem[Griliches(1957)]{griliches1957hybrid} Griliches, Z. (1957). Hybrid corn:
An exploration in the economics of technological change. \emph{Econometrica},
25(4), 501--522.

\bibitem[Hassan et al.(2019)]{hassan2019firm} Hassan, T. A., Hollander, S., van
Lent, L., \& Tahoun, A. (2019). Firm-level political risk: Measurement and
effects. \emph{Quarterly Journal of Economics}, 134(4), 2135--2202.

\bibitem[Hayashi(1982)]{hayashi1982tobin} Hayashi, F. (1982). Tobin's marginal
$q$ and average $q$: A neoclassical interpretation. \emph{Econometrica}, 50(1),
213--224.

\bibitem[Hidalgo et al.(2007)]{hidalgo2007product} Hidalgo, C. A., Klinger, B.,
Barab\'asi, A.-L., \& Hausmann, R. (2007). The product space conditions the
development of nations. \emph{Science}, 317(5837), 482--487.

\bibitem[Hoberg \& Phillips(2016)]{hoberg2016text} Hoberg, G., \& Phillips, G.
(2016). Text-based network industries and endogenous product differentiation.
\emph{Journal of Political Economy}, 124(5), 1423--1465.

\bibitem[Korinek(2023)]{korinek2023language} Korinek, A. (2023). Language models
and cognitive automation for economic research. \emph{NBER Working Paper} 30957.

\bibitem[Li(2008)]{li2008annual} Li, F. (2008). Annual report readability,
current earnings, and earnings persistence. \emph{Journal of Accounting and
Economics}, 45(2--3), 221--247.

\bibitem[Loughran \& McDonald(2011)]{loughran2011liability} Loughran, T., \&
McDonald, B. (2011). When is a liability not a liability? Textual analysis,
dictionaries, and 10-Ks. \emph{Journal of Finance}, 66(1), 35--65.

\bibitem[Peters \& Taylor(2017)]{peters2017intangible} Peters, R. H., \& Taylor,
L. A. (2017). Intangible capital and the investment-$q$ relation. \emph{Journal
of Financial Economics}, 123(2), 251--272.
\end{thebibliography}

\end{document}