spb/uqo-research-db

1\documentclass[11pt]{article}23% ---------- packages ----------4\usepackage[utf8]{inputenc}5\usepackage[T1]{fontenc}6\usepackage{lmodern}7\usepackage[margin=1in]{geometry}8\usepackage{amsmath,amssymb}9\usepackage{graphicx}10\usepackage{booktabs}11\usepackage{multirow}12\usepackage{array}13\usepackage{caption}14\usepackage{subcaption}15\usepackage{float}16\usepackage{setspace}17\usepackage{natbib}18\usepackage{xcolor}19\usepackage{titlesec}20\usepackage[hidelinks,colorlinks=true,linkcolor=blue!50!black,citecolor=blue!40!black,urlcolor=blue!50!black]{hyperref}2122\graphicspath{{../figures/}}23\captionsetup{font=small,labelfont=bf}24\onehalfspacing25\titleformat{\section}{\large\bfseries}{\thesection}{0.6em}{}26\titleformat{\subsection}{\normalsize\bfseries}{\thesubsection}{0.6em}{}2728\newcommand{\R}{$R^2$}29\newcommand{\fig}[1]{Figure~\ref{#1}}30\newcommand{\tab}[1]{Table~\ref{#1}}3132% ---------- title ----------33\title{\bfseries The Corporate Project Economy:\\[0.3em]34\large Large-Scale Extraction and Econometric Analysis of 19{,}000 Strategic35Initiatives from Two Decades of SEC Filings (2010--2026)}3637\author{%38  \begin{tabular}{c}39    {\large Manel Kammoun}\\[-0.05cm]40    {\small\href{mailto:manel.kammoun@uqo.ca}{manel.kammoun@uqo.ca}}\\[0.35cm]41    {\large Djerrycharli Tandjambianda}\\[-0.05cm]42    {\small\href{mailto:djerrycharli.tandjambianda@uqo.ca}{djerrycharli.tandjambianda@uqo.ca}}\\[0.35cm]43    {\large Simon-Pierre Boucher}\\[-0.05cm]44    {\small\href{mailto:simon-pierre.boucher@uqo.ca}{simon-pierre.boucher@uqo.ca}}45  \end{tabular}\\[0.5cm]46  \small D\'epartement des sciences administratives\\47  \small Universit\'e du Qu\'ebec en Outaouais\\[0.3cm]48  \small Gatineau -- Pavillon Alexandre-Tach\'e\\49  \small 283, boulevard Alexandre-Tach\'e\\50  \small Gatineau, Qu\'ebec, Canada J9A 1L8%51}5253\date{\today}5455\begin{document}56\maketitle5758\begin{abstract}59\noindent60Corporate strategy is enacted through \emph{projects} --- plant builds, data61centers, acquisitions, research programs, digital migrations --- yet these62initiatives are scattered, unstructured, and buried inside the narrative63sections of regulatory filings. We introduce SPID (the SEC Project Intelligence64Database), a pipeline that converts the raw text of corporate filings into a65structured, project-centric panel, and we use it to build the largest66machine-extracted census of corporate projects we are aware of: \textbf{19{,}22767resolved projects} drawn from \textbf{39{,}930 mentions} across68\textbf{153{,}894} filing sections (10-K, 10-Q and 8-K) filed by \textbf{50069large U.S.\ public companies} between 2010 and 2026. A large-language-model70extractor (\texttt{gpt-4o-mini}, JSON mode) maps each section to a controlled7121-category project taxonomy with quantitative attributes (announced capital,72status, locations, technologies, partners), which are then resolved into73firm-level project portfolios with timelines, a 32{,}226-node knowledge graph,74and 384-dimensional semantic embeddings. Using this panel we estimate a battery75of models: fixed-effects Poisson and negative-binomial count models of76project-announcement intensity; hedonic log-capital regressions with firm and77year effects; a logistic technology-diffusion model with an estimated inflection78in \textbf{2023}; a difference-in-differences event study of COVID-19 and79corporate digitalization; Cox proportional-hazards models of project lifecycles;80community detection over a technology co-occurrence network; unsupervised81clustering of project embeddings; and a gradient-boosted classifier of82``mega-projects'' that attains a cross-validated AUC of \textbf{0.74} from83text-derived features alone. We document strong persistence in project activity,84a sharp post-2020 acceleration of announcements, an energy-transition technology85cluster at the center of the corporate-innovation network, and project86lifecycles that range from a median of four years for M\&A integration to twelve87years for digital and AI programs. Beyond the substantive findings, the paper88demonstrates that LLM-based information extraction can turn the unstructured89regulatory record into research-grade panel data for empirical90finance and the economics of innovation.9192\vspace{0.5em}93\noindent\textbf{Keywords:} corporate investment, text-as-data, large language94models, information extraction, innovation, SEC filings, knowledge graphs,95panel econometrics.\\96\noindent\textbf{JEL:} G31, O30, C55, D22, M15.97\end{abstract}9899\clearpage100\tableofcontents101\clearpage102103% =====================================================================104\section{Introduction}105% =====================================================================106The modern public corporation is, in operational terms, a portfolio of107\emph{projects}. New products are launched, plants and data centers are built,108firms are acquired and integrated, enterprise systems are migrated to the cloud,109research pipelines are advanced, and entire business lines are reoriented around110new technologies such as artificial intelligence. These initiatives are the111mechanism through which strategy becomes capital expenditure and, eventually,112realized returns. Yet for the empirical researcher they are remarkably hard to113observe. Financial statements report aggregate capital expenditure, not the114projects that compose it; press releases are selective and unstructured; and the115richest single source --- the narrative discussion inside regulatory filings ---116is written in prose, not in tables.117118This paper has two goals. The first is methodological: to show that recent119advances in large language models (LLMs) make it possible to convert the120unstructured regulatory record into structured, research-grade panel data at121scale and at low cost. The second is substantive: to use the resulting data to122characterize the \emph{corporate project economy} of large U.S.\ firms over the1232010--2026 period and to estimate a series of econometric models that speak to124classic questions in corporate investment and the economics of innovation.125126We construct SPID, the SEC Project Intelligence Database, on top of a 500-company127corpus of filings downloaded from EDGAR. The pipeline (i) sections raw 10-K HTML128into its project-dense items (Business, Risk Factors, Properties, MD\&A), (ii)129sends every project-relevant text window to an LLM extractor that returns a130strict-JSON project record against a 21-category taxonomy, (iii) resolves the131resulting mentions into firm-level projects with timelines using stable hashed132identifiers, and (iv) builds a knowledge graph and a set of semantic embeddings133on top of the resolved projects. The end product is a panel of134\textbf{19{,}227 projects} and \textbf{39{,}930 dated mentions} spanning135\textbf{500 firms}, \textbf{11 GICS sectors}, \textbf{21 project types}, and136\textbf{3{,}080 distinct locations}, of which \textbf{9{,}321 (48.5\%)} carry an137explicit dollar amount with a median announced value of \textbf{US\$718 million}.138139The scale and structure of this panel allow us to ask questions that aggregate140data cannot answer. \emph{How persistent is project activity within firms, and141did it change after 2020?} We estimate fixed-effects Poisson and142negative-binomial models and find strong, statistically robust persistence and a143significant post-COVID jump in announcement intensity, against a mild negative144secular time trend once firm composition is held fixed. \emph{What kinds of145projects command the most capital?} Hedonic log-capital regressions with firm146and year fixed effects explain roughly a third of the variation in announced147amounts and reveal large premia for capital-intensive project families relative148to M\&A integration. \emph{How fast is corporate AI adoption diffusing?} A149logistic diffusion model places the inflection point of cumulative AI adoption150in \textbf{2023}, coincident with the generative-AI boom, and the AI share of151new project announcements rises from 0.7\% in 2015 to 4.7\% in 2025. \emph{Did152the pandemic differentially accelerate the digitalization of physically153intensive incumbents?} A difference-in-differences event study with clean154parallel pre-trends returns a precisely estimated null, cautioning against155the popular narrative of pandemic-driven digital convergence. \emph{How long do156projects live?} Cox and Kaplan--Meier models show that M\&A integrations resolve157fastest (median four years) while digital and AI programs persist longest158(median twelve years), and that larger projects complete more slowly.159160We complement these regressions with network and representation-learning161analyses. A technology co-occurrence network exhibits strong community structure162(modularity 0.52) organized into four ``technology ecosystems,'' with163energy-transition technologies --- solar, wind, battery storage --- occupying the164most central positions in the corporate-innovation graph. Unsupervised165clustering of the 384-dimensional project embeddings recovers six interpretable166themes that map cleanly onto the taxonomy while crossing sector boundaries.167Finally, a gradient-boosted classifier predicts whether a project is a168``mega-project'' (above the \$718m median) with a cross-validated AUC of 0.74169using only text-derived structural features, with project type, announcement170year and sector the most informative predictors.171172\paragraph{Contributions.} We make four contributions. First, we provide a173\emph{reproducible, low-cost methodology} for turning the narrative sections of174regulatory filings into structured project-level panel data using an LLM175extractor with a controlled taxonomy and deterministic resolution. Second, we176deliver a \emph{novel dataset} --- to our knowledge the largest177machine-extracted corporate-project census --- together with a knowledge graph178and embeddings that support search, analytics, and downstream modeling. Third,179we present a \emph{comprehensive econometric analysis} that combines count180models, hedonic regressions, diffusion curves, difference-in-differences,181survival analysis, network science, and machine learning on a single coherent182panel. Fourth, we surface \emph{substantive findings} about persistence, the183post-2020 surge, AI diffusion timing, project longevity, and the centrality of184the energy transition that are difficult or impossible to obtain from185conventional structured data.186187\paragraph{Caveats up front.} Our census is an extraction from disclosure, not a188ground-truth registry. Announced amounts are self-reported and heterogeneous in189scope; the headline aggregate of announced capital should be read as an190\emph{upper-bound order of magnitude} rather than a netted investment total,191because the same multi-year program can be mentioned (and counted) in several192filings and because large M\&A enterprise values dominate the sum. Throughout,193we therefore emphasize \emph{within-firm} and \emph{relative} comparisons, which194are robust to these level issues, and we treat the dollar figures as ordinal195signals of project scale. Section~\ref{sec:limitations} discusses measurement,196selection, and extraction error in detail.197198The remainder of the paper proceeds as follows.199Section~\ref{sec:related} situates the work in the text-as-data and corporate200investment literatures. Section~\ref{sec:data} describes the corpus, the201extraction pipeline, and the resolution procedure. Section~\ref{sec:descriptives}202presents descriptive statistics and stylized facts. Section~\ref{sec:framework}203lays out the econometric framework. Sections~\ref{sec:counts}--\ref{sec:ml}204report results for each family of models. Section~\ref{sec:robustness} collects205robustness checks, and Section~\ref{sec:limitations} discusses limitations.206Section~\ref{sec:conclusion} concludes.207208% =====================================================================209\section{Related Literature}\label{sec:related}210% =====================================================================211Our work sits at the intersection of three literatures: text-as-data methods in212economics and finance, the empirical analysis of corporate investment, and the213measurement of technological change and innovation.214215\paragraph{Text as data.} A now-large body of work extracts economic signals216from corporate disclosure. Early contributions used dictionary methods and217readability indices to study tone and sentiment in 10-K filings218\citep{loughran2011liability,li2008annual}, while later work measured risk and219attention from the co-occurrence of words and topics220\citep{hoberg2016text,hassan2019firm}. \citet{gentzkow2019text} survey the221econometrics of text and emphasize the move from counting words to learned222representations. We extend this trajectory in two ways. First, rather than223scoring documents, we perform \emph{structured information extraction}: each text224window is mapped to a typed record with named entities and quantities. Second,225we use a generative LLM as the extraction engine, which removes much of the226brittle feature engineering required by earlier rule-based systems while227retaining a controlled output schema.228229\paragraph{LLMs for extraction and measurement.} Recent papers show that LLMs230can classify and extract economic content with accuracy approaching that of231trained human readers, at a fraction of the cost232\citep{korinek2023language,bybee2023ghost}. The methodological question is no233longer whether extraction is feasible but how to make it \emph{reproducible,234auditable, and cheap at corpus scale}. Our pipeline addresses this with a235deterministic taxonomy, JSON-mode decoding, confidence thresholds, and stable236hashed mention identifiers that make the entire process incremental and237resumable. The marginal cost of extracting a project mention is on the order of238a tenth of a U.S.\ cent, which is what makes a census of this size practical.239240\paragraph{Corporate investment.} The empirical investment literature has long241worked with aggregate capital expenditure and Tobin's $q$242\citep{hayashi1982tobin,fazzari1988financing}, and more recently with the243rising importance of intangible and innovation-related investment244\citep{peters2017intangible,crouzet2019understanding}. A persistent245difficulty is that the unit of decision --- the project --- is not observed.246Our data make the project the unit of analysis, allowing us to study the247\emph{extensive margin} (how many initiatives a firm announces) separately from248the \emph{intensive margin} (how large each is), and to characterize the249\emph{composition} of investment across project types and technologies.250251\paragraph{Innovation and diffusion.} The economics of technology diffusion has252modeled adoption with logistic and Bass-type curves253\citep{griliches1957hybrid,bass1969new}, and a recent literature measures the254diffusion of artificial intelligence across firms using job postings, patents,255and earnings calls \citep{babina2024artificial,acemoglu2022artificial}. We256contribute a disclosure-based measure of AI \emph{projects} and estimate its257diffusion path directly, finding an inflection in 2023 that aligns with the258arrival of large generative models.259260\paragraph{Networks and representations.} Finally, we connect to work that261represents firms and technologies as networks \citep{hidalgo2007product} and as262learned embeddings. Our knowledge graph and sentence-embedding clustering let us263recover technology ecosystems and latent project themes that cut across the264formal taxonomy, complementing the regression evidence with structure that is265not imposed a priori.266267% =====================================================================268\section{Data and Extraction Pipeline}\label{sec:data}269% =====================================================================270\subsection{Source corpus}271The underlying corpus is a 500-company subset of U.S.\ SEC filings retrieved from272EDGAR, comprising raw filing HTML and a pre-parsed store of filing metadata and273sectioned text. Two annual and quarterly forms (10-Q and 8-K) were already274sectioned upstream; the project-dense 10-K filings were retained as raw HTML. As275the first stage of SPID we ingested and sectioned \textbf{7{,}654} 10-K filings,276extracting their Business (Item~1), Risk Factors (Item~1A), Properties (Item~2),277and Management's Discussion and Analysis (Item~7/7A) sections into278\textbf{27{,}164} clean text sections. Unioned with the pre-parsed 10-Q and 8-K279sections, the analysis feed contains \textbf{153{,}894} filing sections (with at280least 30 words) eligible for extraction: 103{,}926 from 8-K, 27{,}164 from 10-K,281and 22{,}804 from 10-Q filings.282283\subsection{LLM extraction}284Each eligible section is passed to an extraction model285(\texttt{gpt-4o-mini}, temperature 0, JSON mode). To bound cost and focus the286model, we first apply a deterministic ``project gate'' --- a regular-expression287filter for project-signal vocabulary --- and send only the windows of text288around matches, capped at roughly 14{,}000 characters per call. The model is289instructed to return a strict-JSON array of project objects, each with a290\texttt{project\_name}, a \texttt{project\_type} drawn from a fixed 21-category291taxonomy (Table~\ref{tab:bytype}), a one-sentence grounded \texttt{description},292an \texttt{objective}, a \texttt{status} (planned / in-progress / completed /293mentioned), an optional \texttt{amount\_usd}, and arrays of294\texttt{locations}, \texttt{technologies}, \texttt{partners}, \texttt{benefits},295\texttt{risks}, and \texttt{years}, together with a self-reported296\texttt{confidence} in $[0,1]$. The system prompt forbids speculation and297instructs the model to prefer a small number of high-quality records over many298vague ones; records below a confidence of 0.45 are discarded.299300This stage produced \textbf{39{,}930 project mentions}. Because the model emits301typed quantities, downstream analysis requires no further parsing: amounts are302numeric, statuses and types are categorical, and list fields are stored as303native arrays in a columnar database.304305\subsection{Resolution, graph, and embeddings}306The same physical project is typically mentioned in several filings across307years. We resolve mentions into projects by grouping on firm, project type, and308a normalized project anchor, and we assign each project a timeline, a canonical309location, an aggregated announced amount, first/last-seen dates, and an average310extraction confidence. Resolution collapses the 39{,}930 mentions into311\textbf{19{,}227 distinct projects}; the mention timeline is retained for the312survival analysis. From the resolved projects we build a directed knowledge313graph with \textbf{32{,}226 nodes} (19{,}227 projects, 7{,}640 partners, 3{,}080314locations, 1{,}779 technologies, 500 companies) and \textbf{33{,}176 edges}315(\texttt{owns}, \texttt{uses}, \texttt{located\_in}), and we encode every316project's text into a 384-dimensional sentence embedding317(MiniLM, \texttt{all-MiniLM-L6-v2}) for semantic search and clustering.318319\subsection{The economics of extraction}320A defining feature of the approach is its cost structure. Each section incurs at321most one model call, and the deterministic project gate suppresses calls on322sections with no project vocabulary; the focused-window construction caps the323prompt at roughly 14{,}000 characters regardless of section length. The marginal324cost of processing a section is therefore on the order of a tenth of a U.S.\325cent, and the full-corpus pass over 153{,}894 sections completed in a few hours326on commodity hardware with bounded memory, batched into 34 firm-cohorts with327checkpointing so that the run is fully resumable. This matters substantively:328the feasibility of a \emph{census} (as opposed to a sample) is precisely what329distinguishes LLM extraction from earlier hand-coding or rule-based approaches,330and it is what allows us to populate a 7{,}370-observation firm-year panel with331project-level granularity. The same OpenAI-compatible interface can target a332self-hosted open-weights model, so the pipeline is not tied to any single333provider, and the per-section cost can in principle be driven to near zero on334local infrastructure.335336\subsection{Analysis datasets}337For econometric analysis we derive three datasets from the resolved store: (i) a338project-level cross-section (19{,}227 rows) with engineered flags for AI and339digital orientation, project age, and technology counts; (ii) a mention-level340event table (39{,}930 rows) with filing dates and forms; and (iii) a balanced341\textbf{firm-year panel} of \textbf{7{,}370} observations spanning 500 firms and342the years 2010--2025 (the partial year 2026 is excluded from the panel), in343which each firm-year records the count of newly announced projects, the number344that are AI- or digital-oriented, announced capital, and project-type345composition. Announcement year is taken from the project's first observed year.346\tab{tab:summary} reports summary statistics for the main variables.347348\begin{table}[H]349\centering350\caption{Summary statistics. Amounts are in millions of USD; duration is the351span between a project's first and last appearance in filings.}352\label{tab:summary}353\small354\input{../tables/tab_summary_stats}355\end{table}356357% =====================================================================358\section{Descriptive Evidence and Stylized Facts}\label{sec:descriptives}359% =====================================================================360We begin with the broad contours of the corporate project economy.361362\paragraph{Volume and trend.} \fig{fig:mentions} plots the number of extracted363project mentions per year by filing form. The series is large and stable ---364between roughly 2{,}000 and 2{,}800 mentions per year --- with a mild upward365drift (a compound annual growth rate of about 2.2\% over 2010--2025) and a366visible step-up after 2020. 8-K filings (event-driven disclosures) and 10-K367filings (annual reports) are the dominant sources, consistent with the fact that368new initiatives are announced in 8-Ks and reviewed comprehensively in 10-Ks.369370\begin{figure}[H]371\centering372\includegraphics[width=0.78\textwidth]{fig_mentions_by_year.pdf}373\caption{Project mentions extracted per year, by filing form. The corpus is374roughly balanced across years, with 8-K and 10-K filings supplying the bulk of375project signals.}376\label{fig:mentions}377\end{figure}378379\paragraph{Composition.} \fig{fig:typedist} and \tab{tab:bytype} show how380projects distribute across the 21-category taxonomy. M\&A integration is by far381the most frequent category (7{,}274 projects), followed by capital-expenditure382programs (2{,}085), plant construction (1{,}252), and partnerships (1{,}239).383The long tail includes specialized categories such as drug pipelines, ERP384implementations, and data-center builds. The dominance of M\&A integration385reflects both the economic importance of acquisitions among large firms and the386prominence of integration narratives in their filings.387388\begin{figure}[H]389\centering390\includegraphics[width=0.74\textwidth]{fig_type_distribution.pdf}391\caption{Distribution of resolved projects across the 21-category taxonomy.}392\label{fig:typedist}393\end{figure}394395\begin{table}[H]396\centering397\caption{Projects by type: counts, aggregate announced capital, AI share, and398mean extraction confidence. Capital is the sum of disclosed amounts and is an399upper bound (see Section~\ref{sec:limitations}).}400\label{tab:bytype}401\small402\input{../tables/tab_by_type}403\end{table}404405\paragraph{Sectoral structure.} The composition of projects varies sharply406across sectors. \fig{fig:heatmap} row-normalizes the project-type mix within each407GICS sector. Utilities and Energy concentrate in plant construction and408energy-transition projects; Information Technology and Communication Services in409product launches and technology deployment; Health Care in R\&D and drug410pipelines; and Financials and Consumer sectors in M\&A integration and411geographic expansion. \tab{tab:bysector} reports the sector-level totals: every412sector is well populated, with Industrials (2{,}708 projects), Health Care413(2{,}475), and Information Technology (2{,}225) the most project-active.414415\begin{figure}[H]416\centering417\includegraphics[width=0.92\textwidth]{fig_sector_type_heatmap.pdf}418\caption{Project-type composition by GICS sector (row-normalized within-sector419shares). Darker cells indicate a larger share of a sector's projects in that420category.}421\label{fig:heatmap}422\end{figure}423424\begin{table}[H]425\centering426\caption{Sector breakdown: number of firms, projects, aggregate announced427capital, and AI share.}428\label{tab:bysector}429\small430\input{../tables/tab_by_sector}431\end{table}432433\paragraph{Capital and geography.} \fig{fig:capex} overlays announced capital and434project counts by announcement year. Announced capital is lumpy and dominated by435a handful of very large transactions in particular years, whereas the project436\emph{count} is smooth, underscoring why we model the extensive and intensive437margins separately. \fig{fig:locations} shows the geographic concentration of438projects, with the United States and a familiar set of global hubs accounting439for most located projects.440441\begin{figure}[H]442\centering443\begin{subfigure}{0.52\textwidth}444\includegraphics[width=\textwidth]{fig_capex_by_year.pdf}445\caption{Announced capital and project count by year.}446\label{fig:capex}447\end{subfigure}\hfill448\begin{subfigure}{0.46\textwidth}449\includegraphics[width=\textwidth]{fig_top_locations.pdf}450\caption{Top 15 project locations.}451\label{fig:locations}452\end{subfigure}453\caption{Capital and geography of the corporate project economy.}454\end{figure}455456\subsection{Six stylized facts}457The descriptive evidence can be distilled into six stylized facts that motivate458the formal analysis and that, to our knowledge, have not previously been459documented at this granularity:460\begin{enumerate}461\item \textbf{Project activity is voluminous and steadily growing.} Large public462firms disclose on the order of 2{,}000--2{,}800 distinct project signals per year,463growing at roughly 2.2\% annually, with a level step-up after 2020.464\item \textbf{Acquisitions dominate the project mix.} M\&A integration is the465single largest category (38\% of projects), reflecting both its economic weight466and its disclosure salience among incumbents.467\item \textbf{Sectors specialize sharply.} Project composition is highly468sector-specific --- energy and utilities in physical builds, technology in469deployments, health care in pipelines --- with little cross-sector convergence in470composition.471\item \textbf{Capital is concentrated in physical projects.} Disclosed amounts472are heavily right-skewed and dominated by capital-intensive physical families;473the median disclosed project is \$718m but the distribution spans four orders of474magnitude.475\item \textbf{AI is frequent but small.} AI-oriented projects are proliferating476rapidly in \emph{count} yet remain individually modest in disclosed capital.477\item \textbf{Disclosure salience tracks scale.} The number of filings that478mention a project rises monotonically with its announced amount, validating both479the resolution procedure and the economic content of the extracted records.480\end{enumerate}481482% =====================================================================483\section{Econometric Framework}\label{sec:framework}484% =====================================================================485The panel structure of the data supports a sequence of complementary models. We486summarize the specifications here and report results in the following sections.487488\paragraph{(i) Count models.} Let $y_{it}$ denote the number of projects firm489$i$ announces in year $t$. We model the conditional mean as490\begin{equation}491\mathbb{E}[y_{it}\mid x_{it}] = \exp\!\big(\alpha_i + \delta_t + x_{it}'\beta\big),492\label{eq:poisson}493\end{equation}494estimated by pseudo-Poisson maximum likelihood (PPML) with firm effects495$\alpha_i$ and year effects $\delta_t$. PPML is consistent under arbitrary496distributional misspecification given a correctly specified mean, which makes it497the natural workhorse for non-negative count outcomes. Because raw project counts498are overdispersed, we also estimate a negative-binomial model and a pooled499Poisson with sector and year effects. Standard errors are clustered by firm.500501\paragraph{(ii) Hedonic capital regressions.} For the 9{,}321 projects with a502disclosed amount we estimate503\begin{equation}504\log(\text{amount}_{j}) = \gamma_{\text{type}(j)} + \theta_{\text{status}(j)}505+ \lambda\,\text{AI}_j + \mu' z_j + \alpha_{i(j)} + \delta_{t(j)} + \varepsilon_j,506\label{eq:hedonic}507\end{equation}508where $z_j$ collects project covariates (technology count, filing salience) and509$\alpha_{i(j)}$, $\delta_{t(j)}$ are firm and year effects. The coefficients510$\gamma_{\text{type}}$ are interpreted as log-capital premia relative to a511baseline category (M\&A integration).512513\paragraph{(iii) Diffusion.} We model AI adoption at the firm-year level with a514logistic regression for the probability of announcing at least one AI project,515and we fit a logistic (S-curve) diffusion model516$F(t)=K/(1+e^{-r(t-t_0)})$ to the cumulative share of firms that have ever517announced an AI project, recovering the saturation $K$, speed $r$, and518inflection $t_0$.519520\paragraph{(iv) Difference-in-differences.} To test whether COVID-19521differentially accelerated digitalization, we define treatment as above-median522\emph{physical-project intensity} in the pre-period (2015--2019) and estimate a523dynamic event study,524\begin{equation}525\text{DigitalShare}_{it} = \alpha_i + \delta_t526+ \sum_{\tau\neq 2019}\pi_\tau\,(\text{Treat}_i\times \mathbf{1}[t=\tau])527+ u_{it},528\label{eq:did}529\end{equation}530with 2019 omitted, alongside a canonical $2\times2$ specification. The leads531$\pi_\tau$ for $\tau<2019$ provide a parallel-trends test.532533\paragraph{(v) Survival.} We treat each multi-filing project as a spell that534begins at first appearance and ``fails'' when it reaches completed status,535right-censoring projects last seen as planned or in progress. We estimate536Kaplan--Meier survival functions by project family and a Cox proportional-hazards537model with project covariates.538539\paragraph{(vi) Networks and representations.} We analyze the degree540distribution of the knowledge graph, detect communities in a technology541co-occurrence network by greedy modularity maximization, and cluster the542384-dimensional project embeddings with $K$-means, selecting $K$ by silhouette.543544\paragraph{(vii) Prediction.} Finally, we train a gradient-boosted tree545classifier to predict ``mega-projects'' (announced amount above the sample546median) from text-derived structural features, evaluated by 5-fold547cross-validated ROC-AUC with permutation importance.548549% =====================================================================550\section{The Extensive Margin: Count Models}\label{sec:counts}551% =====================================================================552\tab{tab:counts} reports the count-model estimates. Three results stand out.553554First, project activity is \textbf{strongly persistent}. In the pooled Poisson555the coefficient on the lagged project count is $0.100$ ($p<0.01$); within firms556(column 2, firm and year effects) it remains positive and highly significant at557$0.028$, an incidence-rate ratio of about $1.03$ per additional prior-year558project. Persistence survives the inclusion of firm effects, indicating that it559reflects within-firm dynamics --- ongoing programs spawning follow-on560initiatives --- rather than fixed cross-firm heterogeneity alone.561562Second, there is a \textbf{significant post-2020 increase} in announcement563intensity. In specifications with sector (rather than firm) effects, the564post-COVID indicator carries a coefficient of $0.208$ ($p<0.01$), implying565roughly a 23\% higher announcement rate after 2020, even as the secular year566trend is mildly negative ($-0.031$ per year). The juxtaposition --- a negative567trend but a positive level shift after 2020 --- suggests a structural break568rather than a smooth acceleration.569570Third, the data are \textbf{strongly overdispersed}. A Cameron--Trivedi auxiliary571regression rejects equidispersion overwhelmingly ($t=18.3$, $p<10^{-70}$), and572the negative-binomial dispersion parameter is $\hat\alpha=0.50$. Overdispersion573justifies the negative-binomial specification (column 3) and cautions against574naive Poisson standard errors; our clustered, PPML-based inference is robust to575it. \fig{fig:countcoef} visualizes the within-firm effects of persistence and576accumulated AI experience.577578\begin{table}[H]579\centering580\caption{Panel count models of project-announcement intensity. Dependent581variable: number of projects announced by firm $i$ in year $t$. Cluster-robust582standard errors (by firm) in parentheses.583$^{*}p<0.1$, $^{**}p<0.05$, $^{***}p<0.01$.}584\label{tab:counts}585\small586\input{../tables/tab_count_models}587\end{table}588589\begin{figure}[H]590\centering591\includegraphics[width=0.66\textwidth]{fig_count_coef.pdf}592\caption{Within-firm drivers of project-announcement intensity (Poisson with593firm and year effects, 95\% confidence intervals).}594\label{fig:countcoef}595\end{figure}596597The accumulated-AI-experience term is positive and significant in the pooled and598negative-binomial models but attenuates to insignificance within firms,599indicating that the cross-sectional association between AI experience and600overall project activity is largely a firm-level (selection) phenomenon rather601than a within-firm causal channel --- AI-active firms are simply more602project-active in general.603604% =====================================================================605\section{The Intensive Margin: What Drives Announced Capital?}\label{sec:amount}606% =====================================================================607\tab{tab:amount} reports the hedonic log-capital regressions of608equation~\eqref{eq:hedonic}. With firm and year effects (column 2) the model609explains $R^2=0.318$ of the variation in log announced amount --- substantial for610project-level data --- and the project-type fixed effects are jointly the611dominant component.612613Several findings are noteworthy. The number of distinct technologies attached to614a project is positively associated with its size, consistent with larger615projects being more technologically complex, though the within-firm estimate is616imprecise. Filing salience (the log number of filings mentioning a project) is617strongly positively associated with announced capital: larger projects are618discussed more often, which both validates the resolution procedure and reflects619genuine economic salience. Project \emph{status} matters: relative to the620omitted category, projects observed as in-progress or completed differ621systematically in disclosed scale from those merely planned or mentioned.622623Strikingly, AI projects are \emph{not} larger than comparable projects; if624anything the point estimate is negative ($-0.18$ in log points) and statistically625insignificant ($p=0.48$). This is an economically meaningful null: in this626period AI initiatives are numerous and strategically prominent but individually627modest in disclosed capital relative to plant builds, acquisitions, and628infrastructure programs. \fig{fig:amountcoef} plots the estimated capital premia629by project type relative to M\&A integration; capital-intensive families ---630infrastructure, plant construction, data centers, energy transition --- carry631the largest premia, while software, partnership, and cost-reduction projects sit632below the baseline.633634\begin{table}[H]635\centering636\caption{Hedonic regressions for log announced capital (projects with a637disclosed amount, $N=9{,}321$). Baseline project type: M\&A integration.638Cluster-robust SE (by firm). $^{*}p<0.1$, $^{**}p<0.05$, $^{***}p<0.01$.}639\label{tab:amount}640\small641\input{../tables/tab_amount_models}642\end{table}643644\begin{figure}[H]645\centering646\includegraphics[width=0.74\textwidth]{fig_amount_coef.pdf}647\caption{Announced-capital premia by project type relative to M\&A integration648(firm and year effects, 95\% CI). Capital-intensive physical projects command649the largest premia.}650\label{fig:amountcoef}651\end{figure}652653% =====================================================================654\section{Technology Diffusion: The Rise of Corporate AI}\label{sec:diffusion}655% =====================================================================656\fig{fig:diffusion} summarizes the diffusion analysis. The left panel plots the657cumulative share of firms that have ever announced an AI project together with658the fitted logistic curve. The estimated saturation is $K=0.42$, the speed is659$r=0.21$, and the inflection point is \textbf{$t_0=2023.1$} --- precisely when660generative AI entered the corporate mainstream. The right panel shows the AI661\emph{share} of all new project announcements rising more than sixfold, from6620.7\% in 2015 to 4.7\% in 2025, with a clear acceleration after 2022.663664\begin{figure}[H]665\centering666\includegraphics[width=0.95\textwidth]{fig_ai_diffusion.pdf}667\caption{AI-project diffusion. Left: cumulative share of firms ever announcing an668AI project, with a fitted logistic (S-curve) diffusion model; the vertical line669marks the estimated inflection year ($\approx$2023). Right: AI share of all new670project announcements by year.}671\label{fig:diffusion}672\end{figure}673674The firm-year logit (\tab{tab:ailogit}) confirms a positive secular trend in the675probability of AI announcement (average marginal effect $+0.25$ percentage points676per year) and a positive but imprecisely estimated post-COVID shift. Larger677project portfolios are associated with higher AI-adoption probability,678consistent with scale and absorptive capacity facilitating experimentation. That679the diffusion is still on the steep part of its S-curve --- the fitted saturation680of 0.42 is far from unity --- implies that corporate AI adoption, measured by681disclosed projects, had substantial room to run at the end of the sample.682683\begin{table}[H]684\centering685\caption{Firm-year logit for the probability of announcing $\geq1$ AI project.686Average marginal effects in the right column. Cluster-robust SE (by firm).687$^{*}p<0.1$, $^{**}p<0.05$, $^{***}p<0.01$.}688\label{tab:ailogit}689\small690\input{../tables/tab_ai_logit}691\end{table}692693% =====================================================================694\section{COVID-19 and Digitalization: A Difference-in-Differences Test}\label{sec:did}695% =====================================================================696A widespread narrative holds that the pandemic forced a rapid digital697convergence, with physically intensive incumbents racing to catch up. We test698this directly. Treatment is defined as above-median physical-project intensity in6992015--2019 (223 treated vs.\ 250 control firms); the outcome is the digital700share of a firm-year's project announcements.701702\fig{fig:did} plots the dynamic treatment effects from703equation~\eqref{eq:did}. The pre-2019 leads are jointly indistinguishable from704zero ($F$-test $p=0.78$), supporting the parallel-trends assumption, and the705post-2020 lags hover around zero. The canonical $2\times2$ estimate706(\tab{tab:did}) is $-0.010$ (s.e.\ $0.014$, $p=0.49$) --- a precisely estimated707null. We find \emph{no evidence} that physically intensive incumbents708differentially accelerated their digitalization after COVID-19 relative to709already-digital firms. If anything, both groups digitalized in parallel. This710tempers the convergence narrative: the pandemic may have raised digital activity711broadly, but it did not measurably close the cross-firm gap in project712composition.713714\begin{figure}[H]715\centering716\includegraphics[width=0.78\textwidth]{fig_did_event.pdf}717\caption{Event study of COVID-19 and digitalization. Points are718treatment$\times$year coefficients on the digital share of announcements719relative to 2019 (95\% CI). Flat pre-trends and post-period coefficients near720zero indicate no differential digital acceleration of physically intensive721incumbents.}722\label{fig:did}723\end{figure}724725\begin{table}[H]726\centering727\caption{Difference-in-differences: COVID-19 and digital project share.728Cluster-robust SE (by firm).}729\label{tab:did}730\small731\input{../tables/tab_did}732\end{table}733734% =====================================================================735\section{Project Lifecycles: Survival Analysis}\label{sec:survival}736% =====================================================================737How long do corporate projects live? Among the 5{,}905 projects tracked across at738least two filings (2{,}664 of which reach completed status), median lifecycle739length varies dramatically by family. \fig{fig:km} plots Kaplan--Meier survival740curves --- the probability that a project has \emph{not} yet been completed as a741function of months since its first appearance. M\&A integrations resolve fastest742(median $\approx$49 months), capital and construction projects next ($\approx$84743months), while R\&D / product programs ($\approx$132 months) and especially744digital / AI initiatives ($\approx$144 months) persist far longer. The ordering745is intuitive: acquisitions have a natural completion event, whereas digital746transformation and AI are open-ended, continuously renewed programs that rarely747declare victory.748749\begin{figure}[H]750\centering751\includegraphics[width=0.78\textwidth]{fig_survival_km.pdf}752\caption{Kaplan--Meier survival curves by project family. M\&A integrations753complete fastest; digital and AI programs persist longest.}754\label{fig:km}755\end{figure}756757The Cox proportional-hazards model (\tab{tab:cox}) quantifies the covariate758effects. Larger projects complete more slowly: a one-log-point increase in759announced amount lowers the completion hazard by about 4\% (hazard ratio $0.96$,760$p<10^{-10}$), consistent with bigger initiatives being more complex and761longer-lived. The AI indicator is not significantly associated with the762completion hazard once project type and scale are controlled (HR $1.06$,763$p=0.74$), reinforcing that the long lifecycles of AI programs operate through764their \emph{type} (open-ended digital programs) rather than an independent AI765effect.766767\begin{table}[H]768\centering769\caption{Cox proportional-hazards model of time to project completion (hazard770ratios). Project-type controls included. $^{*}p<0.1$, $^{**}p<0.05$,771$^{***}p<0.01$.}772\label{tab:cox}773\small774\input{../tables/tab_cox}775\end{table}776777% =====================================================================778\section{The Innovation Network}\label{sec:network}779% =====================================================================780The knowledge graph offers a structural view of the corporate project economy.781\fig{fig:degree} plots its degree distribution on log-log axes: the distribution782is heavy-tailed with an estimated tail slope of $-1.55$, the signature of a783scale-free-like structure in which a few hub nodes (large diversified firms,784ubiquitous technologies) connect to a great many projects while most nodes have785low degree.786787\begin{figure}[H]788\centering789\includegraphics[width=0.64\textwidth]{fig_degree_dist.pdf}790\caption{Degree distribution of the 32{,}226-node project knowledge graph791(log-log). The heavy tail is consistent with a scale-free-like organization.}792\label{fig:degree}793\end{figure}794795To uncover technology ecosystems we build a co-occurrence network in which two796technologies are linked when they appear in the same project, restricted to the79760 most frequent technologies. Greedy modularity maximization recovers798\textbf{four communities} with a high modularity of \textbf{0.52}, indicating799well-separated ecosystems. \fig{fig:technet} visualizes the network;800\tab{tab:network} lists the most central technologies. Strikingly, the highest801eigenvector-centrality nodes are all \emph{energy-transition} technologies ---802solar (0.57), renewable energy (0.49), wind (0.44), battery storage (0.26),803natural gas, energy storage --- which form the gravitational center of the804corporate-innovation graph. Software/SaaS, cloud, AI, and semiconductors form a805second, digital ecosystem; manufacturing/automation and connectivity form806others. The centrality of the energy transition reflects both its capital807intensity and its cross-sector reach: utilities, industrials, materials, and808consumer firms all participate in it, whereas digital technologies, though809pervasive, are more modular and less tightly co-bundled.810811\begin{figure}[H]812\centering813\includegraphics[width=0.90\textwidth]{fig_tech_network.pdf}814\caption{Technology co-occurrence network (top 60 technologies, edges for815co-use in $\geq2$ projects). Node color denotes community; node size scales with816project frequency. Four ecosystems emerge (modularity 0.52), with817energy-transition technologies most central.}818\label{fig:technet}819\end{figure}820821\begin{table}[H]822\centering823\caption{Most central technologies in the co-occurrence network (eigenvector824centrality) with project frequency and community membership.}825\label{tab:network}826\small827\input{../tables/tab_network}828\end{table}829830% =====================================================================831\section{Latent Themes: Embedding the Project Corpus}\label{sec:embeddings}832% =====================================================================833The formal taxonomy is imposed at extraction time. To recover \emph{emergent}834structure we cluster the 384-dimensional MiniLM project embeddings with835$K$-means, selecting $K=6$ by silhouette (\fig{fig:silhouette}). Although the836silhouette values are modest --- expected for dense sentence embeddings of837heterogeneous business text --- the resulting clusters are interpretable and838economically coherent (\tab{tab:clusters}). They span an M\&A/technology cluster,839a health-care partnership and AI cluster, an energy/utilities construction840cluster centered on renewables and storage, an industrial capex/ERP cluster, a841consumer geographic-expansion cluster, and a consumer-staples cost-reduction and842automation cluster.843844\begin{figure}[H]845\centering846\includegraphics[width=0.82\textwidth]{fig_embedding_tsne.pdf}847\caption{Two-dimensional t-SNE projection of project embeddings (stratified848sample of 5{,}000 projects), colored by $K$-means cluster. Themes separate while849overlapping at the boundaries, reflecting the continuous semantic space of850corporate initiatives.}851\label{fig:tsne}852\end{figure}853854\begin{figure}[H]855\centering856\includegraphics[width=0.56\textwidth]{fig_silhouette.pdf}857\caption{Silhouette score versus number of clusters; $K=6$ is selected.}858\label{fig:silhouette}859\end{figure}860861\begin{table}[H]862\centering863\caption{Embedding clusters: size, dominant project type, top technologies, and864aggregate announced capital. Clusters cross sector boundaries, capturing themes865the taxonomy alone does not.}866\label{tab:clusters}867\small868\input{../tables/tab_clusters}869\end{table}870871\fig{fig:tsne} shows the t-SNE projection of a sample of projects. The clusters872separate visibly while overlapping at their margins, consistent with corporate873initiatives occupying a continuous semantic space rather than discrete bins.874That the unsupervised themes largely \emph{recover} the supervised taxonomy ---875while recombining it across sectors --- is an internal-validity check on the876extraction: the embeddings and the LLM labels agree on the broad organization of877the corpus.878879% =====================================================================880\section{Predicting Mega-Projects}\label{sec:ml}881% =====================================================================882Can the scale of a project be predicted from its qualitative description alone?883We train a gradient-boosted tree classifier to distinguish ``mega-projects''884(announced amount above the \$718m median) from smaller ones, using only885text-derived structural features --- project type, sector, status, technology886count, filing and mention counts, AI flag, presence of a location, announcement887year, and extraction confidence --- with no leakage of the amount itself.888889The model attains a 5-fold cross-validated \textbf{ROC-AUC of 0.743} and average890precision of 0.722 (\fig{fig:roc}, \tab{tab:ml}). Permutation importance ranks891\emph{project type} first, followed by \emph{announcement year}, \emph{sector},892and \emph{filing count}. The predictive value of year and sector indicates that893mega-projects cluster in particular periods and industries (e.g., recent894data-center and energy build-outs), while the importance of filing count895confirms that disclosure salience tracks economic scale. That qualitative text896features alone recover three-quarters of the discriminative information about897project scale is itself evidence that the extracted attributes carry real898economic content.899900\begin{figure}[H]901\centering902\includegraphics[width=0.95\textwidth]{fig_ml_roc.pdf}903\caption{Mega-project classifier. Left: 5-fold cross-validated ROC curve904(AUC = 0.74). Right: permutation feature importance (drop in AUC).}905\label{fig:roc}906\end{figure}907908\begin{table}[H]909\centering910\caption{Gradient-boosted mega-project classifier: feature-group importance and911performance ($N=9{,}321$ projects with disclosed amounts).}912\label{tab:ml}913\small914\input{../tables/tab_ml}915\end{table}916917% =====================================================================918\section{Robustness and Validation}\label{sec:robustness}919% =====================================================================920We summarize the checks that support the analysis.921922\paragraph{Extraction quality.} The mean extraction confidence is 0.78 and the923median 0.90; records below 0.45 are dropped at the source. The agreement between924the unsupervised embedding clusters and the supervised taxonomy925(Section~\ref{sec:embeddings}) provides an internal-consistency check: two926independent representations of the same text organize it the same way. Spot927auditing of high-amount records confirms that disclosed figures are grounded in928the underlying text rather than hallucinated, a behavior the JSON-mode prompt929explicitly enforces by forbidding unsupported numbers.930931\paragraph{Estimator robustness.} Because project counts are overdispersed, we932report pooled Poisson, firm-fixed-effects PPML, and negative-binomial estimates933side by side; the sign and significance of the persistence and post-2020 effects934are stable across all three (\tab{tab:counts}). All panel models use935cluster-robust standard errors by firm. The hedonic capital results are936qualitatively identical under sector effects and under firm effects937(\tab{tab:amount}), and the project-type ranking of premia is unchanged.938939\paragraph{Parallel trends.} The credibility of the difference-in-differences940rests on parallel pre-trends, which we cannot reject ($F$-test $p=0.78$);941the dynamic event-study coefficients are flat before 2020942(\fig{fig:did}).943944\paragraph{Level vs.\ composition.} Throughout we privilege within-firm and945relative comparisons over levels. The aggregate announced-capital figure is946sensitive to a small number of very large transactions and to repeated mentions947of multi-year programs; none of our regression conclusions depends on it, because948the hedonic model uses log amounts with firm and year effects and the count949models use the extensive margin.950951\paragraph{Sample.} The 500-firm corpus is weighted toward large, mature public952companies, which is appropriate for studying the project economy of incumbents953but limits external validity to small and private firms. Results should be read954as describing large-cap U.S.\ public companies over 2010--2026.955956% =====================================================================957\section{Limitations}\label{sec:limitations}958% =====================================================================959Several limitations warrant emphasis. \emph{First}, our census measures960\emph{disclosure of} projects, not projects themselves; initiatives that are961never discussed in filings are invisible, and disclosure incentives vary across962firms and over time. \emph{Second}, announced amounts are self-reported,963inconsistently scoped (a figure may be total program cost, annual spend, or deal964value), and frequently absent (only 48.5\% of projects carry a number); we965mitigate this with logs, fixed effects, and an explicit focus on relative966magnitudes. \emph{Third}, LLM extraction can err --- both false positives967(spurious projects) and false negatives (missed initiatives) --- and although968the confidence filter and grounding constraints reduce these, we do not have a969hand-labeled gold standard large enough to estimate precision and recall970precisely; constructing one is a priority for future work. \emph{Fourth},971resolution merges mentions heuristically; aggressive merging understates project972counts while conservative merging overstates them. \emph{Fifth}, the survival973analysis infers completion from disclosed status, which is itself a noisy signal.974Finally, our identification in the difference-in-differences is quasi-experimental975and the treatment is defined on pre-period behavior; we interpret the null976cautiously, as the absence of a \emph{differential} effect rather than the977absence of any pandemic effect.978979% =====================================================================980\section{Discussion and Implications}\label{sec:discussion}981% =====================================================================982The results have implications for research methodology, for the empirical study983of corporate investment, and for practice.984985\paragraph{For methodology.} The paper is, in part, an existence proof: a single986researcher, using a commodity LLM and open-source statistical tooling, can987construct a population-scale, project-level panel from the unstructured988regulatory record and subject it to the full apparatus of modern applied989econometrics. The binding constraint in text-based empirical work has shifted990from \emph{extraction} to \emph{validation and design}. Because extraction is now991cheap and reproducible, the scarce inputs are a credible ground-truth benchmark992and a research design that isolates economic mechanisms from disclosure993artifacts. Our internal-consistency checks --- the agreement of unsupervised994embeddings with the supervised taxonomy, the monotone relation between disclosure995salience and scale, and the stability of estimates across estimators --- are a996template, but they are not a substitute for a hand-labeled gold standard, which997remains the most valuable complementary investment.998999\paragraph{For the study of investment.} Observing the project as a unit allows a1000clean separation of the extensive margin (how many initiatives a firm pursues)1001from the intensive margin (how large each is) and the compositional margin (which1002kinds). Our evidence suggests these margins are governed by different forces:1003the extensive margin is persistent and firm-driven with a macro level shift after10042020; the intensive margin is dominated by project \emph{type} and physical1005capital intensity; and the compositional margin is strongly sectoral and, in the1006case of AI, diffusing along a classic S-curve. A natural agenda is to link these1007margins to financing conditions, governance, and ex-post outcomes --- for1008instance, whether project-announcement intensity predicts future capital1009expenditure, productivity, or returns, and whether the market prices the project1010portfolio beyond aggregate guidance.10111012\paragraph{For the AI-diffusion debate.} Our disclosure-based measure complements1013job-posting and patent-based measures and adds a distinct vantage point: the1014\emph{strategic} framing of AI as a corporate project. The 2023 inflection and1015the still-low saturation imply that, as of the sample's end, corporate AI was1016early in its diffusion and that the count surge had not yet translated into large1017disclosed capital commitments --- a pattern consistent with an experimentation1018phase preceding heavy build-out. Whether the subsequent data-center and compute1019investments convert AI from a high-count, low-capital category into a high-capital1020one is a question the same pipeline can answer as new filings arrive.10211022\paragraph{For practice.} For analysts and policymakers, a structured project1023census enables monitoring that aggregate statistics cannot: tracking1024energy-transition build-outs across sectors in near real time, mapping1025partnership and supply-chain networks, or flagging firms whose project mix is1026shifting. The knowledge graph and embeddings make the corpus queryable both1027structurally and semantically, turning a static disclosure archive into a living1028intelligence layer.10291030% =====================================================================1031\section{Conclusion}\label{sec:conclusion}1032% =====================================================================1033We have shown that the narrative sections of SEC filings, long treated as1034qualitative context, can be converted into structured, research-grade panel data1035on corporate projects at scale and at trivial marginal cost, and that the1036resulting data support a rich program of empirical analysis. From 153{,}8941037filing sections we extracted 39{,}930 project mentions and resolved them into103819{,}227 projects across 500 firms, building a knowledge graph and semantic1039embeddings on top. The econometric analysis documents strong within-firm1040persistence of project activity and a significant post-2020 level shift; large1041capital premia for physically intensive project families and a surprising null1042for the scale of AI projects; an AI-diffusion path inflecting in 2023; a1043precisely estimated null for differential pandemic-driven digitalization;1044project lifecycles ranging from four years for acquisitions to twelve for digital1045programs; an energy-transition technology ecosystem at the center of the1046innovation network; and a text-only classifier that predicts project scale with1047an AUC of 0.74.10481049Beyond the specific findings, the broader message is that LLM-based information1050extraction reshapes what is measurable in empirical finance and the economics of1051innovation. The project --- the actual unit of corporate decision-making --- can1052now be observed at population scale. Natural next steps include constructing a1053human-labeled validation set to benchmark extraction precision and recall,1054extending the corpus to additional forms (S-1, DEF 14A, 20-F) and to the full1055universe of filers, linking projects to ex-post financial and market outcomes,1056and using the project panel to revisit classic questions about investment,1057financing constraints, and the returns to corporate innovation.10581059\clearpage1060\appendix1061% =====================================================================1062\section{Project Taxonomy and Variable Definitions}\label{app:vars}1063% =====================================================================1064The extraction taxonomy comprises 21 mutually exclusive project types:1065manufacturing expansion, plant construction, data center, cloud migration, AI1066initiative, digital transformation, ERP implementation, R\&D program, product1067launch, drug pipeline, M\&A integration, geographic expansion, capex program,1068supply chain, automation, sustainability, energy transition, infrastructure,1069cost reduction, partnership, and technology deployment. Key engineered variables1070are defined as follows. \textbf{AI project}: a project whose type is1071\emph{AI initiative} or whose technology list contains an AI term (AI, artificial1072intelligence, machine/deep learning, generative, LLM, neural).1073\textbf{Digital project}: an AI project or one whose type is among cloud1074migration, ERP, data center, automation, digital transformation, or technology1075deployment. \textbf{Announcement year}: the first calendar year in which the1076project appears in any filing. \textbf{Duration}: days between a project's first1077and last appearance. \textbf{Announced amount}: the resolved (aggregated) USD1078figure disclosed for the project, when present.10791080% =====================================================================1081\section{Pipeline and Reproducibility}\label{app:repro}1082% =====================================================================1083The extraction backend is OpenAI-compatible and was run with1084\texttt{gpt-4o-mini} at temperature 0 in JSON mode. The pipeline is incremental1085and resumable: mention identifiers are stable SHA-1 hashes of1086(accession, section, type, anchor), so re-runs use \texttt{INSERT OR REPLACE}1087semantics and never double-count. The full corpus run processed the 153{,}8941088sections in 34 firm-batches with checkpointing; resolution, graph construction,1089embedding, and all analyses are deterministic given the resolved store. Analysis1090code is organized as ten numbered scripts (dataset construction; descriptives;1091count models; hedonic models; diffusion; difference-in-differences; survival;1092network; embeddings; prediction), each emitting figures, \LaTeX\ table fragments,1093and a machine-readable results file. Estimation uses1094\texttt{statsmodels} (Poisson, negative binomial, logit, OLS, Cox PHReg),1095\texttt{scipy} (logistic curve fitting), \texttt{networkx} (community detection1096and centrality), and \texttt{scikit-learn} ($K$-means, t-SNE, gradient boosting,1097cross-validation).10981099\begin{thebibliography}{99}1100\small1101\bibitem[Acemoglu et al.(2022)]{acemoglu2022artificial} Acemoglu, D., Autor, D.,1102Hazell, J., \& Restrepo, P. (2022). Artificial intelligence and jobs: Evidence1103from online vacancies. \emph{Journal of Labor Economics}, 40(S1), S293--S340.11041105\bibitem[Babina et al.(2024)]{babina2024artificial} Babina, T., Fedyk, A., He,1106A., \& Hodson, J. (2024). Artificial intelligence, firm growth, and product1107innovation. \emph{Journal of Financial Economics}, 151, 103745.11081109\bibitem[Bass(1969)]{bass1969new} Bass, F. M. (1969). A new product growth model1110for consumer durables. \emph{Management Science}, 15(5), 215--227.11111112\bibitem[Bybee et al.(2023)]{bybee2023ghost} Bybee, L., Kelly, B., \& Su, Y.1113(2023). Narrative-driven fluctuations in sentiment: Evidence from large language1114models. Working paper.11151116\bibitem[Crouzet \& Eberly(2019)]{crouzet2019understanding} Crouzet, N., \&1117Eberly, J. (2019). Understanding weak capital investment: The role of market1118concentration and intangibles. \emph{NBER Working Paper} 25869.11191120\bibitem[Fazzari et al.(1988)]{fazzari1988financing} Fazzari, S. M., Hubbard,1121R. G., \& Petersen, B. C. (1988). Financing constraints and corporate1122investment. \emph{Brookings Papers on Economic Activity}, 1988(1), 141--206.11231124\bibitem[Gentzkow et al.(2019)]{gentzkow2019text} Gentzkow, M., Kelly, B., \&1125Taddy, M. (2019). Text as data. \emph{Journal of Economic Literature}, 57(3),1126535--574.11271128\bibitem[Griliches(1957)]{griliches1957hybrid} Griliches, Z. (1957). Hybrid corn:1129An exploration in the economics of technological change. \emph{Econometrica},113025(4), 501--522.11311132\bibitem[Hassan et al.(2019)]{hassan2019firm} Hassan, T. A., Hollander, S., van1133Lent, L., \& Tahoun, A. (2019). Firm-level political risk: Measurement and1134effects. \emph{Quarterly Journal of Economics}, 134(4), 2135--2202.11351136\bibitem[Hayashi(1982)]{hayashi1982tobin} Hayashi, F. (1982). Tobin's marginal1137$q$ and average $q$: A neoclassical interpretation. \emph{Econometrica}, 50(1),1138213--224.11391140\bibitem[Hidalgo et al.(2007)]{hidalgo2007product} Hidalgo, C. A., Klinger, B.,1141Barab\'asi, A.-L., \& Hausmann, R. (2007). The product space conditions the1142development of nations. \emph{Science}, 317(5837), 482--487.11431144\bibitem[Hoberg \& Phillips(2016)]{hoberg2016text} Hoberg, G., \& Phillips, G.1145(2016). Text-based network industries and endogenous product differentiation.1146\emph{Journal of Political Economy}, 124(5), 1423--1465.11471148\bibitem[Korinek(2023)]{korinek2023language} Korinek, A. (2023). Language models1149and cognitive automation for economic research. \emph{NBER Working Paper} 30957.11501151\bibitem[Li(2008)]{li2008annual} Li, F. (2008). Annual report readability,1152current earnings, and earnings persistence. \emph{Journal of Accounting and1153Economics}, 45(2--3), 221--247.11541155\bibitem[Loughran \& McDonald(2011)]{loughran2011liability} Loughran, T., \&1156McDonald, B. (2011). When is a liability not a liability? Textual analysis,1157dictionaries, and 10-Ks. \emph{Journal of Finance}, 66(1), 35--65.11581159\bibitem[Peters \& Taylor(2017)]{peters2017intangible} Peters, R. H., \& Taylor,1160L. A. (2017). Intangible capital and the investment-$q$ relation. \emph{Journal1161of Financial Economics}, 123(2), 251--272.1162\end{thebibliography}11631164\end{document}1165