Deep-research engine

The adaptive core of the infrastructure

The deep-research engine serves as the adaptive core of the infrastructure. It organizes successive inquiry across literature, primary sources, and the state-of-the-art technical repositories, combining contributions from multiple LLMs under researcher direction. Evidence and conclusions accumulate in a shared knowledge vault for retrieval and reconsideration.

Development agents build and maintain the engine, while research agents investigate methods and capabilities. The engine uses findings from this continuing inquiry to build and revise tools, workflows, and resources, enabling the infrastructure to adapt to changing research needs and advances in AI.

What has been built

Research protocols, a scout skill and software helpers support successive inquiry. A methods review informed source and comparison checks in the Grounded Theory toolkit. The successor helper has been tested on synthetic inputs only.

What comes next

Adapt a shareable engine, integrate and evaluate it, and use continuing inquiry into models and methods to revise tools, workflows and resources.

From a source to a synthesis

Explore 77 works and 141 located claims about theory-building methods and software. Follow one source to see how its evidence informs a design decision.

Each work links its inspected versions, claims, exact source locators and evidence limits.

CellEvidence classFind

Loading the records…

Use the evidence map

Download the evidence, claims and citing passages. The appendix provides source addresses and the inputs for the offline validator.

Read the AI prompt

# Explore the deep-research engine

Read https://textdata.info/engine/ and the component folder at https://github.com/joyzhzh/textdata/tree/main/appendix/01-engine.

Start with `README.md`, `example/README.md`, and `tool/README.md`. Use `example/campaign/EVIDENCE_LEDGER.jsonl`, `example/campaign/FINDINGS.md`, and `example/campaign/SYNTHESIS.md` to explain one concrete example: Langley’s 1999 article (`work-doi:10.5465/amr.1999.2553248`), its two abstract-level claims (`clm-grf-005-01` and `clm-grf-005-02`), and how they inform the choice of theory-building methods.

Help me explore the evidence and synthesis in the page, or inspect the same files locally. If a checkout and Python are available, offer the existing commands in the README to check the supplied example. Explain what those checks establish. Distinguish the working scout/validator and completed evidence map from the proposed integrated engine; retain the limits of abstract-level evidence, incomplete search coverage and synthetic-only successor tests. Do not imply that the validator conducts a new inquiry or verifies every source interpretation.

My research question or aspect to explore: [insert here]

Evidence informing tools

The engine’s methods inquiry informed the Grounded Theory toolkit’s source and comparison checks: identify the lines supporting an incident, state what a comparison changed, and explain responses to negative cases. Further sampling remains a proposal for the researcher.

Explore the Grounded Theory toolkit

Run and extend

The appendix includes scout instructions, a search contract, an offline validator and the worked example. Python 3.9 or later is sufficient for the checks; no model or network is needed. A new inquiry requires a researcher or agent with source access.

The evidence map is English-primary, public-access and effort-bounded, observed through 28 July 2026. It is not saturated. Many claims rely on abstracts or software documentation; candidate software was not run, and a human has not checked every claim against its source.

Adapting a shareable engine, integrating it and evaluating it remain proposed. The successor helper is not included and has only synthetic-input evidence from 42 unit tests. Unattended operation and wide external adoption are unproved.

Open the engine appendix

Back to the overview