spb/firc0001_uqo

1% Owner: Agent 10 (Stage E writing). Every generated exhibit is \input exactly once in the prose at the anchor2% where it is discussed; please do not add a second \input of the same float (duplicate labels). — 2026-11-053% ============================================================================4\section{Related Literature}\label{sec:literature}5% ============================================================================67\paragraph{Text as data in accounting and finance.} The narrative content of SEC filings has become a standard empirical input. \citet{loughran2011} showed that general-purpose sentiment dictionaries misclassify financial text and introduced finance-specific word lists; \citet{loughran2016} survey the subsequent decade of textual analysis in accounting and finance, and \citet{gentzkow2019} provide the general econometric framework for treating text as data, including the measurement-error and validation issues that any text-derived variable inherits. \citet{dyer2017} document how 10-K disclosure has lengthened and become more repetitive over time, a trend that matters for any extractor that reads the same passages year after year; whether repeated boilerplate produces false-positive projects in SPID has not been measured and is one of the labels of the annotation protocol (\Cref{sec:validation}). \citet{cohen2020} show that changes in otherwise repetitive filing language carry information, which motivates our emphasis on \emph{repeated} mentions of the same initiative across filings rather than on single occurrences. \citet{grimmer2013} offer the general methodological warning that automated text measures must be validated against human judgment on the specific task at hand, which is the stance adopted in \Cref{sec:validation}.89\paragraph{Measuring corporate activity from disclosures.} Closest to our construct are text-based measures of what firms do rather than how they feel. \citet{hoberg2016} build product-market classifications from 10-K business descriptions; \citet{hobergmaksimovic2015} identify financially constrained firms from the language of the liquidity discussion; \citet{hassan2019} measure firm-level political risk from the share of earnings-call text devoted to politics; \citet{sautner2023} measure firm-level climate exposure from calls and validate it against green hiring and patenting; \citet{bellstam2021} score innovation from analyst reports for firms without patents; \citet{li2021} measure corporate culture with word embeddings; \citet{kelly2021} measure the novelty of patents from their text, and \citet{kalyani2025} track the emergence and diffusion of technologies by discovering bigrams in patents and counting them in earnings calls and job postings. Two recent papers obtain firm-level measures directly from a generative model: \citet{kim2023} summarize earnings-call transcripts into political, climate and AI risk scores, and \citet{jha2024} prompt a chat model to read managers' investment plans from calls, producing a firm-quarter score that is validated against CFO survey responses and predicts capital expenditure several quarters ahead. SPID differs from all of these in its unit of observation. Rather than a firm-level score summarizing the intensity of a theme, it targets \emph{discrete initiatives}---objects with a name, a type, a status, cited years, a location, partners and an amount---that can be followed from filing to filing. Relative to firm-level text scores \citep{hoberg2016,hassan2019,sautner2023,jha2024} and to event-level extraction from 8-Ks \citep{crane2024,dolphin2026}, SPID's unit is the initiative itself, followed across 10-K, 10-Q and 8-K filings; unlike those studies it does not yet report accuracy against human labels, which \Cref{sec:validation} specifies. \Cref{tab:positioning} places the database next to the closest prior work on the dimensions that matter for a user: unit, sources, extractor, provenance granularity and the validation reported.1011\begin{table}[H]12\centering\scriptsize13\caption{Positioning of SPID relative to the closest text-based measures, extraction systems and corpora}\label{tab:positioning}14\begin{threeparttable}15\begin{tabularx}{\textwidth}{@{}>{\raggedright\arraybackslash}p{2.6cm}>{\raggedright\arraybackslash}p{2.5cm}>{\raggedright\arraybackslash}p{2.1cm}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{2.2cm}>{\raggedright\arraybackslash}X@{}}16\toprule17Study & Unit & Sources & Method & Provenance & Validation reported \\18\midrule19\citet{hoberg2016} & firm-pair similarity; industry classification & 10-K business descriptions & word-vector similarity & firm-year & against SIC/NAICS classifications on explaining fundamentals \\20\citet{hassan2019} & firm-quarter political-risk score & earnings calls & pattern-based bigram share & firm-quarter & against volatility, lobbying and donations; human audit of snippets \\21\citet{sautner2023} & firm-year climate exposure & earnings calls & machine-learning keyword discovery & firm-year & against green jobs and patents \\22\citet{kim2023} & firm-level risk measures & earnings calls & GPT summarization & firm-call & against volatility and investment outcomes \\23\citet{jha2024} & firm-quarter investment score & earnings calls & ChatGPT prompting & firm-quarter & against CFO surveys; predicts capex up to nine quarters ahead \\24\citet{crane2024} & layoff event & Form 8-K & text extraction & filing & against WARN notices (administrative source) \\25\citet{dolphin2026} & 8-K event, 119-type three-tier taxonomy & Form 8-K & two-stage LLM, labels traceable to the source text & passage & label reliability evaluated \\26\citet{loukas2021} & item-segmented corpus & all 10-K filings 1993--2020 & HTML parsing, Item segmentation & item & corpus release, no extraction \\27\citet{bettencourt2026} & layout-faithful corpus & SEC filings, many forms & reconstruction into structured text with tables preserved & document & corpus release, no extraction \\28SPID (this paper) & initiative mention, linked into projects across filings & 10-K, 10-Q, 8-K primary documents 2010--2026, \SPIDuniverseN\ issuers & gpt-4o-mini in JSON-object mode behind a lexical gate and focus windows; deterministic key & section (accession, section id); no character span & mechanical grounding of structured fields, 8-K item-code consistency, rules-vs-LLM agreement (\Cref{sec:validation}); no human labels yet \\29\bottomrule30\end{tabularx}31\begin{tablenotes}\footnotesize32\item Content of the rows other than SPID is taken from the cited works as summarized in the literature review of this revision; it describes the design of each study, not its results. ``Provenance'' is the finest unit at which an output can be traced to its source. The row for SPID describes the build at revision \texttt{2324b58} after the Stage D audits.33\end{tablenotes}34\end{threeparttable}35\end{table}3637\paragraph{Large language models as measurement instruments.} The extraction step relies on an instruction-tuned LLM used in a constrained-output mode: a fixed system prompt, a JSON-object response format whose field list is stated in the prompt rather than enforced by a schema, temperature zero, and an explicit instruction to extract only initiatives grounded in the supplied passage. \citet{ludwig2025} give the econometric framework within which such an output can become a regressor: a measurement generated by a language model is usable for estimation only with a validation subsample that identifies its error structure, and prediction uses require that the model's training data not contain the outcome. SPID currently has the first half of that framework---the instrument and its ledger---and not the second; \Cref{sec:validation} is written as the specification of that validation subsample. \citet{dekok2025} sets out the practical protocol for generative models in accounting research (prompt and model versioning, temperature, repeated runs, validation), several items of which the current build does not meet: no prompt hash or model snapshot is recorded per run and the June model snapshot is unknown (\Cref{sec:architecture}). \citet{gilardi2023} show that zero-shot LLM annotation can match or exceed crowd workers on text-classification tasks, which is the comparison the human-versus-model benchmark of \Cref{sec:validation} is designed to make for this construct; \citet{afzal2026} report accuracy for a generic LLM framework extracting structured fields from narrative financial reports. Closest in design are \citet{dolphin2026}, who label 8-K disclosures against a fine-grained taxonomy with labels constrained to be traceable to the source text and evaluate their reliability, and \citet{crane2024}, who extract one event type from 8-Ks and validate it against an administrative source; both report the reliability that SPID does not yet have. The model is asked to prefer fewer, concrete projects to many vague ones, to return an empty list when no concrete project is described, and to attach a self-reported confidence. We make no claim that this confidence is calibrated: verbalized confidence scores of language models are known to be poorly calibrated \citep{yang2024,ye2024}, a missing confidence is defaulted to a fixed value by the post-processing (\Cref{sec:construct}), and calibration is one of the items of the validation protocol. Semantic search over projects uses sentence embeddings \citep{reimers2019}.3839\paragraph{Open EDGAR corpora.} The corpus layer of SPID re-implements a step that prior work has already published: EDGAR-CORPUS \citep{loukas2021} provides every 10-K from 1993 to 2020 segmented by Item, and the Stanford EDGAR Filings Dataset \citep{bettencourt2026} reconstructs filings into layout-faithful text in which tables are preserved. SPID's parser covers 10-K, 10-Q and 8-K primary documents from 2010 to 2026 for its universe, flattens tables to tab-separated text and stores no character offsets into the raw HTML (\Cref{sec:architecture}); for the S\&P~500 block the 10-K items were prepared by an earlier parser that discarded tables altogether (\Cref{sec:architecture}). A user who needs table-level fidelity should prefer those corpora as the text source and use SPID for its project layer.4041\paragraph{Entity resolution and record linkage.} Linking repeated mentions of the same initiative across years is a record-linkage problem in the sense of \citet{fellegi1969}; \citet{christen2012} reviews the modern toolbox of blocking, comparison and classification. The current SPID resolver is a deterministic blocking rule (issuer, project type and a location or name token) rather than a probabilistic linker, and \Cref{sec:resolution} shows how much the project count depends on it. The migration to an explainable, reversible linkage with match scores and human adjudication of ambiguous clusters is part of the roadmap; pre-trained language models have been shown to improve entity matching over string-similarity baselines \citep{li2020}, and generative models can be used for the same task \citep{peeters2023}, which is the direction the proposed \texttt{project\_mention\_links} table is designed to accommodate.4243\paragraph{Infrastructure.} The database is implemented as a single analytical file in DuckDB \citep{raasveldt2019}, with Parquet exports; acquisition respects the SEC's published fair-access policy \citep{secfairaccess}, and the Russell~3000 is defined following FTSE Russell's published construction methodology \citep{ftserussell2026}, approximated here through a fund-holdings proxy whose coverage of the index is quantified in \Cref{sec:sample}.44