Words about research agents
- Summary
- What a chat AI returns when you ask "tell me about X" without retrieval — the model's compressed memory of its training data, stated confidently. Not anchored in any specific source.
- Research agent
- A model + a loop + retrieval tools (web search, fetch URL, file reader) — an agent that can go out and bring real sources into the conversation, instead of answering only from what it remembers from training.
- Retrieval
- The action of going out to a real source — running a search, opening a URL, reading a file — and bringing the contents back. The thing that makes a research agent different from a summarizer.
- The four moves
- Every directed research task has four moves: scope (what's the question), retrieve (gather candidate sources), triangulate (compare and weigh them), synthesize (turn findings into a usable output). You own scope. You audit retrieve. You check triangulate. You read synthesize line by line.
Words about sources and triangulation
- Primary source
- The original record of a thing — a bill's text, a scientific paper reporting original research, a court opinion, a product's own pricing page. Closest to the claim itself.
- Secondary source
- A competent interpretation of primary sources by someone who did real work — investigative journalism, a literature review, a reputable industry report. Cites primary sources and explains them.
- Tertiary source
- A summary of primary and secondary sources without original analysis — Wikipedia, encyclopedia entries, textbooks, overview sites. Useful for orientation; rarely the right place to land an important claim.
- Triangulation
- The move that turns a set of retrieved sources into an answer. Asks three questions per claim: is this supported by more than one independent source? do the sources agree or disagree? if it rests on one source, what tier is it?
- Independent sources
- Sources that arrived at a finding without copying from each other. Three news articles all citing one press release are one source, not three. Counting independence is what gives triangulation its weight.
Words about fabrication risk
- Fabricated citation
- A citation in research output that is not real or does not support the claim it's attached to. Four common patterns: - Invented source — the paper, author, journal, or DOI does not exist. - Real source, fabricated quote — the source exists; the quotation is not in it. - Real source, misrepresented finding — the source exists; the brief misstates what it says. - Dead URL or wrong URL — the URL 404s, or it loads a page that's not what was claimed.
- Open-before-cite rule
- The operational safety norm of Module 4: every source that appears in a final research output must have been opened, skimmed, and confirmed by the student before the output ships. No exceptions for short sources, news URLs, or "the agent said it was good."
- Source-list.md
- A markdown file in your topic folder that tracks every source the agent retrieved, with one row per source: URL, title, author, date, tier, and status (not-yet-opened, opened-and-confirmed, opened-and-dropped). The status column is your attestation that you read the source. A template lives at /resources/module-04/source-list-template.md.
Words about scoping (Lesson 4.2)
- Scoping brief
- A one-page document with five sections — Question, Reader & Decision, Answer Shape, Out of Scope, Constraints — written before any retrieval starts. Defines what a successful research run looks like.
- Question
- The actual interrogative the brief is going to answer. Ends with a question mark and names specifics. "Solar panels" is a topic; "Does installing solar at 1234 Main pay back within ten years given Ohio rates?" is a question.
- Reader and decision
- Who is going to read the output and what they're going to do with it. Tells the agent — and you — what a useful answer looks like.
- Answer shape
- A one-paragraph sketch of the form the answer will take. "A two-paragraph comparison ending with a recommendation" is an answer shape; "tell me what you find" is not.
- Out of scope
- Explicit list of topics the brief is not covering, even though they're adjacent. Defends against scope creep before it starts.
- Constraints
- Time budget and source preferences. "Three hours of my time; primary sources from each publisher's website required; no social-media threads" is a usable constraint set.
Words about synthesis (Lesson 4.4)
- Synthesis
- The move that turns a set of triangulated claims into a single document a reader can use. Has three jobs: shape (what order), selection (what's in vs. out), connection (the prose between claims).
- Synthesis plan
- A one-page outline with six parts — question, thesis, structure, evidence map, uncertainty ledger, reader-next — that you write. Different from the synthesis brief: the plan is what you write, the brief is what the agent writes against your plan.
- Synthesis brief
- The actual prose document, drafted by the agent against your synthesis plan, then read line by line by you. 1,000–2,000 words for most topics.
- Thesis (in this course)
- The one-sentence answer the rest of the brief argues toward. (Same idea as a thesis in an English-class essay.) Write it before the agent drafts. If you can't write it, your triangulation didn't produce a verdict and you need either more evidence or a narrower question.
- Evidence map
- Per-section listing of the two to four sourced claims that will appear in each section, with source tags. The structure that prevents fact-free connector prose.
- Uncertainty ledger
- A short list of things the brief is not confident about — disagreements between sources, single-source claims, areas the triangulation left open. These get named in the brief, near the relevant claim, not in a hedged postscript at the end.
- Confidence tiers
- Low / medium / high. A short, honest vocabulary for marking individual claims when the evidence is mixed or thin. A brief that uses these is more trustworthy than one that pretends to uniform certainty.
The three synthesis failure modes
- False voice
- The brief smooths real disagreement between sources into confident single-voice prose. "Research suggests the effect is real" when two studies say yes and one says no. Fix: name the disagreement where the claim appears.
- Fact-free connector
- A paragraph mostly made of transitional sentences ("It is worth noting that...", "In light of these findings...") with no sourced claim. Reads smoothly, says nothing. Fix: every paragraph carries at least one sourced claim.
- Single-source-as-consensus
- A claim built on one source but phrased in plural-voice ("Experts agree that..."). Fix: name the source inline ("Thompson (2024) argues that...").
Words about output shapes (Lesson 4.5)
- Scoping brief (output 1)
- One page. Defines the work before the work runs. Reader: often you, sometimes a teacher or parent.
- Fact-check memo (output 2)
- 400–800 words. Verdict on a single bounded claim with five sections — claim, verdict, evidence, confidence, what-would-change-your-mind. Reader: a decision-maker who needs a verdict, not an overview.
- Synthesis brief (output 3)
- 1,000–2,000 words. Reasoned case on a larger question. Reader: someone who needs a complete, defendable answer.
- What-would-change-your-mind line
- The integrity check at the end of a fact-check memo. One sentence: what new evidence would move you off your verdict. If you can't name anything, your verdict is held religiously, not on evidence.
- Open-before-cite audit
- The five-step pre-shipping pass on a complete brief: list all citations, cross-check against source-list.md, sample-read three citations at random, check thesis/body alignment, run the synthesis-failure-mode pass. Mandatory before ship.
Forward references
- MCP (Model Context Protocol)
- A standard way to give AI agents extra tools — web search, file fetching, custom data lookups, and so on. Module 7 covers MCP in depth and walks through authoring your own. For Module 4, you only need to know that MCP tools are pre-built capabilities you can enable on your agent. The easiest way to confirm what tools your agent has is to ask the agent directly: "List the tools you currently have access to."
- Quarterly refresh
- Every 90 days, a course maintainer re-tests every Recipe Book entry against the current tool versions and either marks it fresh, patches it, or retires it. Student feedback (flags, patches, new recipes) feeds this refresh. See /recipe-book/_contribute.md for how to contribute.