Source-linked AI summary
INDRA: A New AI Tool for Exploring Tobacco, Fossil Fuel, and Chemical Industry Archives
Daniel Akselrad, Robert N. Proctor
TL;DR
Massive industry archives remain difficult for general-purpose LLMs to access and interpret reliably. INDRA federates these collections and applies a closed evidentiary sandbox, provenance tagging, and deterministic safeguards; preliminary testing found reduced fabrication and misgrounding compared with off-the-shelf Claude Sonnet 5. The platform supports auditable archival investigation, but researchers must select subsets because only about 500,000 tokens can be reliably analyzed at once, and outputs remain nondeterministic.
Problem
Massive industry archives are effectively inaccessible to general-purpose LLMs, which lack direct document access and are vulnerable to hallucination and other defects.
Method
INDRA federates archival collections and uses a closed evidentiary sandbox, provenance tagging, and a system-level protocol enforced with deterministic scripts.
Results
INDRA’s combined safeguards significantly reduce fabrication and misgrounding compared with off-the-shelf Claude Sonnet 5 when given the same documents.
Takeaways & Limitations
INDRA enables multi-archival investigations through outputs designed to be checked rather than trusted, while making knowledge-production conditions more visible and auditable.
Takeaways & Limitations
Only 500,000 tokens can be reliably analyzed at once, so researchers must select a document subset and different searches can produce different results.
Abstract
from arXiv · showhide
Five decades of litigation have disgorged hundreds of millions of pages of formerly secret business records from the tobacco industry, along with documents from the makers of drugs, chemicals, food, firearms, and fossil fuels. Yet these archives have been effectively inaccessible to general-purpose large language models (LLMs) because they have never been compiled into an LLM-readable corpus. Chatbots may be familiar with some of the materials contained in such archives but, with no direct access to the documents, they are vulnerable to hallucination and other defects. Here we introduce INDRA, a research platform designed to remedy such failures by embedding the conventions of archival historiography into a system-level protocol governing every output. The platform federates UCSF's Industry Documents Library, Columbia and CUNY's ToxicDocs, Stanford's SRITA, and other heretofore siloed collections, and provides three interlinked safeguards: (1) a closed evidentiary sandbox confines the model to a user-selected corpus, blocking retrieval from external sources that could introduce bias; (2) real-time provenance tagging marks the boundary between archival evidence and parametric inference; and (3) a system-level protocol enforced by deterministic scripts guides the structure of every output. Together these safeguards prevent the model from conflating "the documents say X" with "I think X" or "I learned X from prior training." The result is an LLM-powered research partner enabling massive multi-archival investigations, a tool whose outputs are designed to be checked rather than trusted, and whose architecture makes the conditions of knowledge production visible and auditable. Three case studies demonstrate the method's analytical value and limitations, including what we call the Heraclitus effect, the steppingstone dilemma, and the gullibility (or mafia) problem.
1 Introduction & Background
INDRA addresses the scale, access, interpretation, and reliability problems that hinder research on adversarial industry archives. It federates collections and constrains LLM outputs through evidentiary, provenance, and protocol-based safeguards designed for verification rather than trust.
- 1 Introduction & Background: The archives contain hundreds of millions of pages and hundreds of billions of tokens, while existing search, modeling, and coding methods either lack interpretation or cannot scale.The Industry Documents Library alone contains about 150 million pages and would take a quick reader thousands of years to digest.
- 1 Introduction & Background: INDRA federates previously siloed industry archives, enabling researchers to search and analyze corporate documents across collections through one portal.The platform supports cross-corpus pattern recognition and investigations of networks, strategies, rhetorics, and historical change.
- 1 Introduction & Background: Its safeguards combine a closed user-selected corpus, provenance tagging, bounded working memory, quotation verification, OCR assessment, and deterministic output checks.The system also reports when evidence is absent rather than filling gaps with unsupported claims.
- 1 Introduction & Background: Preliminary testing found that INDRA’s combined safeguards reduced fabrication and misgrounding relative to off-the-shelf Claude Sonnet 5, while the system refused uncertain outputs.The platform is presented as a research partner whose outputs are designed to be checked, not trusted, and whose limitations include the Heraclitus effect, steppingstone dilemma, and gullibility problem.
- 1 Introduction & Background: General-purpose LLMs lack direct access to many archives and can hallucinate, fabricate citations, and lose information when context windows become overloaded.These systems may rely on impoverished abridgments and attentional biases toward document beginnings and endings.
- 1 Introduction & Background: INDRA embeds archival historiography into a system-level protocol that makes the distinction between archival evidence and model inference visible.This addresses a gap in prior tools that provide citations without adequately flagging interpretation, domain knowledge, or untraceable confabulation.
2 Methods
INDRA combines a closed evidentiary sandbox, sentence-level provenance tagging, and deterministic safeguards to support auditable analysis of federated industry archives. Its search tools pair open-ended semantic retrieval with precise Boolean querying while accounting for OCR and prompt-injection risks.
- 2.1 The Evidentiary Sandbox: INDRA confines analysis to user-selected OCR text, blocks external retrieval, and marks claims based on domain knowledge or interpretation.The sandbox makes the selected archives the model’s evidentiary base, while provenance tags distinguish archival evidence from parametric contributions.
- 2.2 Provenance Tagging: Sentence-level provenance tags label domain knowledge, interpretation, and perspective shifts, while OCR warnings expose uncertainty in the source material.These tags make visible how assertions were derived and whether underlying OCR quality may affect interpretation.
- 2.3 System Prompt: A system protocol requires document-specific citations, prohibits references to unloaded documents, and uses assistant prefilling and a deterministic watchdog to check outputs.The watchdog also helps defend against prompt injection by treating document contents as data rather than instructions.
- 2.3 System Prompt: INDRA’s safeguards are intended to reduce fabrication and misgrounding while making negative findings and the boundary between archival evidence and model inference explicit.The platform is designed to support outputs that can be checked rather than simply trusted.
- 2.4 Trimodal Search: Semantic search uses a BGE-M3-indexed heterogeneous graph, while Boolean search targets exact phrases and document classes that semantic retrieval may miss.The two methods are complementary: semantic retrieval casts a wider conceptual net, whereas Boolean search narrows the query to specified wording or metadata.
- 2.4 Trimodal Search: A multi-agent pipeline ranks graph-retrieved documents and distributes groups of up to 500,000 tokens to six or more analysis agents.The pipeline orchestrates retrieval and parallel analysis across the federated archives.
3 Case Studies
INDRA’s case studies use targeted archival searches to reconstruct tobacco training, fossil-fuel climate rhetoric, and paraquat risk communication. Across these examples, the platform surfaces patterns in corporate messaging, institutional networks, and contrasts between internal knowledge and public claims.
- 3.1 Case Study 1: The “College of Tobacco Knowledge”: INDRA reconstructs the College of Tobacco Knowledge as an industry-wide network linking cigarette companies, lawyers, public-relations agents, and allied organizations.The platform identifies attendees by workplace and connects them to the Institute’s broader “tobacco family.”
- 3.1 Case Study 1: The “College of Tobacco Knowledge”: The college redirected discussion of smoking and health toward industry-friendly claims while training participants in secrecy, litigation preparation, and “issues management.”INDRA retrieves examples of denial and of instructions to keep the college’s activities confidential.
- 3.2 Case Study 2: Climate Change Denial at Exxon: Ecology’s KWIC analysis traces how Exxon’s climate vocabulary changed over time and identifies coordinated denial rhetoric in employee publications during the 1990s.The analysis searches contextual snippets to examine word use across documents and periods.
- 3.2 Case Study 2: Climate Change Denial at Exxon: Across Exxon’s fossil-fuel magazines, “sound science” and selective statistical language supported skepticism about climate change while avoiding terms such as “climate science” and “anthropogenic.”The corpus contains zero occurrences of “climate science” and zero of “anthropogenic” in the cited collections.
- 3.2 Case Study 2: Climate Change Denial at Exxon: ExxonMobil’s later disaster-relief messaging framed petroleum-based products as corporate rescue in response to supposedly natural catastrophes.The case study links this framing to the company’s employee magazines and climate communications.
4 Discussion
INDRA expands archival research by letting scholars interrogate federated collections through evidence-based, cross-corpus analysis rather than conventional keyword pathways. Its constrained tools support new forms of pattern, rhetoric, and network analysis while leaving future capabilities open.
- INDRA lets researchers weave threads across multiple archives into evidence-based, verifiable interpretations and identify rhetorical patterns, concealment, and collaboration networks.
- The platform’s value is framed as a new kind of constrained close reading that increases interpretive power and speed while preserving researcher involvement.
- The platform decouples massive online archives from gated access systems and enables commands ranging from rhetorical typologies to analyses of emotion, metaphor, and corporate communication.
- INDRA’s archival exploration extends beyond text retrieval toward identifying visual elements, geolocating documents by typewriter, and analyzing affect in testimony.
5 Limitations
INDRA remains bounded by token limits, nondeterministic and path-dependent conversations, susceptibility to deceptive archives, and leakage from prior knowledge. These constraints reinforce the need for repeated testing and active scholarly oversight.
- 500,000 tokens can be reliably analyzed at once, so researchers must select subsets whose search terms and corpus boundaries can produce different results.The authors compare this capacity with roughly five scholarly books and expect the constraint to diminish as context windows expand.
- The Heraclitus effect means identical questions over identical documents can yield different emphases because LLM generation is fundamentally nondeterministic.INDRA mitigates this with grounded generation, verbatim extraction, and watchdog scripts, while repeated testing remains good research practice.
- The steppingstone dilemma makes early conversational framing shape later model behavior, including through sycophancy and path dependency.Restarting sessions provides a form of “virtuous forgetting” that can interrupt an established conversational path.
- INDRA may accept dishonest industry language at face value, especially when adversarial archives are filled with propaganda and guarded wording.The authors connect this gullibility to the difficulty of recognizing eavescasting and to the legal fiction that documents speak for themselves.
- Prior internet knowledge cannot be fully excluded from the evidentiary sandbox, so domain knowledge contamination remains visible but not eliminated.Such prior knowledge can sometimes help establish chronology, but it also means the sandbox is not an airtight seal.
- Humans share many of these limitations, including selection bias, gullibility, information overload, and lost-in-the-middle neglect.
6 Conclusion
INDRA is presented as a constrained, provenance-conscious way to read massive archives while preserving the scholar’s role. Its safeguards address contamination risks, but the archive still requires verification and human interpretation.
- INDRA combines an evidentiary sandbox, provenance tagging, watchdog scripts, and transparency tracking to make archival interpretation more faithful and auditable.
- Adversarial archives require attention to how documents were produced and surfaced because industry talking points can generate misleading inferences without safeguards.
- The platform should complement rather than replace trained scholars because broader social, political, and natural context lies outside the model’s context window.The authors characterize AI systems as powerful but credulous assistants whose outputs require verification and repeated testing.