Research in the Fieldbook

Research belongs here when it helps a reader make a better decision about working with AI. A paper can suggest a useful question. A small study can test one. A practical chapter can explain what to do with the result while keeping its conditions visible.

This section has two kinds of work: notes on published papers and reports of our own investigations. Proposals give an untested question a home without presenting it as a finding. These are editorial publications, not a claim of peer review. Keep each page's status visible before its conclusions.

Start with a question

Identify the decision an addition could change. “Which model should I use?” is usually too broad for one small comparison. “Can this session retrieve the current decision from the brief it receives?” gives us something more specific to inspect.

Use the experiments chapter for comparisons and the evidence note for observations. The research template adds publication identity, methods, review status, and claim boundaries.

Reading notes

A reading note links a primary paper, credits its authors, identifies the version and publication status, and describes its question and relevant evidence in original language. Read the method and limitations as well as the abstract. Keep the authors' result separate from our interpretation and proposed application.

Start with A long brief still needs a retrieval check. Its source was inspected; its experiments were not rerun for this book.

Add another note when its paper supplies a distinct question, comparison, or limitation worth teaching. Interesting titles can stay in the roadmap until there is a useful reason to examine one. Link to the paper rather than copying text, figures, or data without checking their rights.

Proposals

A proposal states a question, expected observation, comparison, inputs, scoring, and stopping conditions before any results exist. Explain what would count against the preferred explanation.

Where does the current decision sit? is our first proposed protocol. It asks how placement of a fictional decision record affects retrieval from a supplied brief. It has not been executed, frozen, or preregistered. Its expected call count is a planning quantity, not a completed cohort.

Preparing a proposal does not start provider calls, automations, or a study. Agree on the environment, cost, and execution scope when the work is requested. Keep the proposal useful even if its next step is to reduce or abandon it.

Original reports and research papers

An original report needs the question, background, method, inputs, baselines, scoring, outcomes, interpretation, and limits. Keep tasks, attempts, failures, exclusions, and reasons available in publishable form. Tie the report to exact protocol and artifact identities. A chart illustrating one success does not supply the missing attempts.

Use a report when actual evidence answers a bounded question. A larger research paper can build on that report when the contribution needs a sustained argument, related work, multiple comparisons, or a fuller account of uncertainty. Length and academic formatting do not establish scientific validity.

State who performed the work and what review occurred. Preserve collaboration and authorship. A project's published evaluation remains reported evidence until an independent reproduction is carried out and recorded. Essays and owner accounts retain their own boundaries.

There are no original study results in this section yet. The digest walkthrough's acceptance tests establish its fictional program behavior; they are not a model, skill, or context experiment.

Turn a result into a practice carefully

Connect a study to the chapter it can inform. State which task, model, environment, and conditions support the result and what needs checking before transferring it. If evidence contradicts current advice, preserve the disagreement and update the interpretation openly.

Ask what the reader can now decide or inspect that the current book leaves unclear. When the answer is still speculative, publish the question as a proposal or keep it on the roadmap.