Ad hoc research with AI tools produces ad hoc results. One person scopes the question tightly and verifies carefully; another fires a vague prompt and ships the first answer; nobody can say afterward what was actually checked. A named model fixes this by giving the work a shape anyone can follow and anyone can audit. This article introduces SOURCE, a six-stage model for AI-assisted research.
SOURCE stands for Scope, Origin, Use, Reconcile, Confirm, and Encode. The stages run in order, and each produces an artifact the next stage consumes: a scoped question, chosen origins of evidence, the tool's raw use, reconciled disagreements, confirmed claims, and an encoded audit trail. The name is a mnemonic, not magic. Its job is to make sure no stage gets silently skipped.
Treat SOURCE as a default structure you scale to the stakes. A client-facing recommendation earns the full discipline. A throwaway lookup runs a lightweight version. Either way, the stages keep the work honest and reproducible, which is the difference between research you can defend and a confident guess.
The deeper reason to use a named model is auditability. Without a structure, "we researched it" is an assertion nobody can check. With SOURCE, the same sentence decomposes into six questions a reviewer can actually ask: Did you scope it to a decision? Where did the evidence come from? What did the tool produce? How did you reconcile disagreement? What did you confirm? What did you encode? Each question maps to an artifact, so the answer is either present or visibly missing. That is the difference between a claim and evidence, and it is what makes research survive scrutiny instead of collapsing under it.
Scope: Define the Decision Before the Query
Anchor to a Choice
The Scope stage produces one written sentence: the decision this research will inform. Without it, research drifts toward whatever the tool finds easy to discuss. With it, every later stage has a standard to judge answers against. This is the cheapest stage and the one that prevents the most wasted effort.
Narrow the Question
A scoped question is specific enough that depth is forced. "What did Gmail's 2024 sender requirements mandate for a 60,000-contact list?" beats "tell me about email deliverability." Vague scope guarantees shallow output, the first failure named in When a Research Assistant Hands You a Confident Wrong Answer.
Origin: Choose Where Evidence Comes From
Match Tool to Question Shape
The Origin stage decides which tool or tools fit. Time-sensitive and factual points to live retrieval; deep reasoning over known material points to a document-grounded tool; broad orientation points to synthesis. A mismatch produces weak answers that look fine. The fit decision is mapped in Mapping the Landscape of AI Research Assistants.
Decide on One Tool or Two
If the decision is high-stakes, Origin commits to triangulating across two tools so disagreement can surface later. If it is low-stakes, one tool is enough. This is where the rigor-to-stakes tradeoff gets made explicit.
Use: Run the Tool and Capture the Raw Output
Treat the First Answer as a Draft
The Use stage produces the tool's raw response, explicitly marked as a draft rather than a result. The fluency is seductive; the model writes a shaky inference and a solid fact in the same assured voice. Use captures the output without yet trusting it.
Force Self-Disclosure
Within Use, ask the tool to rate its confidence per claim, name its weakest assumption, and list what it could not verify. That self-disclosure becomes the to-do list the later stages work through.
Reconcile: Resolve Disagreement and Gaps
Read the Disagreement First
When two tools were used, Reconcile reads where they diverge before anything else, because disagreement points straight at the contested part of the topic. This is the highest-value step in the model, the same pattern shown in Inside Three Research Workflows Rebuilt Around AI.
Identify the Load-Bearing Claim
Reconcile also names the single claim the decision rests on. Everything downstream concentrates verification effort there rather than spreading it thinly across the whole output.
Confirm: Verify What Will Ship
Trace to Primary Sources
The Confirm stage traces every load-bearing claim to a primary source, read in context and dated within the relevant window. A link nobody read is not confirmation. This stage is where the per-answer checklist from Vetting an AI Research Tool Before You Trust Its Output gets executed.
Mark the Unconfirmable
Anything that cannot be confirmed gets flagged as such rather than quietly shipped at full confidence. A finding labeled "unverified" is honest; an unverified finding presented as fact is a liability.
Encode: Save the Trail
Capture Prompt, Sources, and Date
The Encode stage produces a one-page record: the prompt, the source list, the date, and the decision. When a finding is challenged months later, this is how it gets defended in minutes. Research you cannot reproduce is not research.
Make Encoding Automatic
Encode should be a template, not a chore. Friction is why trails get skipped, so removing the friction is what makes the final stage actually happen. Whether the whole model is working is the subject of Knowing Whether Your AI Research Workflow Actually Works.
Running SOURCE at Two Different Scales
The Full Version for High-Stakes Work
For a client-facing recommendation, run all six stages with care. Scope to the exact decision, commit to two tools in Origin, capture and self-disclose in Use, read the disagreement in Reconcile, trace every load-bearing claim in Confirm, and save a complete trail in Encode. The investment is minutes to an hour, and it buys a finding you can defend against any challenge.
The Lightweight Version for Quick Lookups
For a throwaway internal question, the model compresses to three moves: scope it in a sentence, run one tool and read the answer as a draft, and gut-check anything that surprises you. Skipping Reconcile, full Confirm, and Encode is correct here, because the stakes do not justify them. The point of a named model is not to make every task heavy; it is to make sure you skip stages on purpose rather than by accident. Choosing where that line falls is the stakes-based judgment explored in Depth, Speed, and Cost in AI Research Software.
Common Failure Points and What They Look Like
Skipping Scope
The most common breakdown is starting at Use, firing a vague prompt before scoping to a decision. The symptom is a sprawling, interesting answer that does not actually resolve anything, because there was no decision for it to serve. When research feels productive but never concludes, the missing stage is almost always Scope. Adding the one-sentence decision up front is the cheapest fix in the whole model.
Collapsing Use Into Confirm
Another frequent failure is treating the tool's first answer as already confirmed, fusing two stages that must stay separate. Use produces a draft; Confirm tests it. When those collapse, unverified output ships under the impression that running the tool was itself a form of checking. Keeping them as distinct stages, with the draft explicitly marked as untrusted until Confirm runs, is what prevents the quiet, confident error described in When a Research Assistant Hands You a Confident Wrong Answer.
Why Naming the Stages Helps
The value of giving each step a name is that a skipped stage becomes nameable, and a nameable gap is a fixable one. "We skipped Confirm" is a precise diagnosis someone can act on. "The research was a bit sloppy" is not. The model's vocabulary turns vague unease about a research process into specific, correctable observations, which is what lets a team improve rather than just worry.
Frequently Asked Questions
Do I have to run all six stages every time?
No. Scale SOURCE to the stakes. A client-facing recommendation runs all six; a quick internal lookup might run Scope, Use, and a light Confirm. The model is a default structure you adjust, not a mandatory ceremony.
Which stage do teams most often skip, and what does it cost?
Confirm and Encode, because the output already looks finished. Skipping Confirm ships unverified claims; skipping Encode means you cannot defend a finding later. Both failures are quiet until they are expensive.
How is SOURCE different from just being careful?
Carefulness is not auditable. SOURCE decomposes "we did the research" into six questions a reviewer can actually ask, each mapped to an artifact that is either present or visibly missing. That is the difference between a claim and evidence.
What is the single most valuable stage?
Reconcile, when two tools were used, because reading the disagreement is the most concentrated signal you get about where the topic is uncertain. Scope is the cheapest high-value stage; Reconcile is the highest-value one.
Can a solo researcher use this, or is it for teams?
Both. A solo researcher uses SOURCE to keep their own work honest and reproducible. A team uses it to make research auditable across people, so "I checked it" becomes a set of artifacts anyone can inspect.
Where does the tool actually save time in this model?
In Origin and Use, where it gathers and structures fast. Reconcile, Confirm, and Encode are human-owned and take real time. The model captures the tool's speed while protecting the parts where speed would be dangerous.
Key Takeaways
- SOURCE gives AI-assisted research a fixed, auditable shape: Scope, Origin, Use, Reconcile, Confirm, Encode.
- Each stage produces an artifact the next consumes, turning "we did the research" into questions a reviewer can check.
- Scope anchors every answer to a decision; Reconcile reads tool disagreement as the highest-value signal.
- Confirm traces load-bearing claims to dated primary sources and flags anything it cannot verify.
- Scale the model to the stakes, and make the Encode audit trail an automatic template rather than a chore.