Build Your Personal Assistant Operating System

4.3

Ground Every Claim in a Source You Can Check

Confident text without sources is a liability you cannot afford

You ask the assistant to research a topic for a client presentation. It produces a polished three-paragraph summary with clear conclusions. The summary reads well. The problem is that you cannot verify any of it. Which claims come from published sources? Which are the assistant's inferences? Which are outdated? You would have to research the topic yourself to find out, which defeats the purpose of asking the assistant.

Source grounding is the of making every factual claim traceable to a source you can check. When the assistant produces research, analysis, or recommendations, every claim should link to its evidence: the article it came from, the data point it references, or the passage it quotes. Claims without sources are clearly marked as the assistant's reasoning, and the reader knows the difference.

This matters for every that produces recommendations, analysis, or briefings. The morning brief, the email , the reading pipeline, the , and the growth tracker all make claims about your data. Source grounding ensures those claims are verifiable.

A stylized teaching image showing research text with claim labels linked to source cards and an uncertainty section at the bottom — Source grounding links every factual claim to checkable evidence and separates what is known from what is inferred.

Source-Grounded Research

Promise	Research briefs and analysis where every factual claim links to a source, uncertain claims are flagged, and the reader can verify any statement without starting from scratch.
Sources	Articles, reports, databases, highlights from the reading pipeline, and data from other modules. Your specific research question or topic.
Output	A research brief with inline source citations, a source list with access information, confidence levels for each claim, and an explicit uncertainty section.
Saves to memory	Research briefs with sources, saved for reference. Key findings linked to the reading pipeline. Uncertainty flags available for future review.
Approval boundary	Reads and synthesizes sources freely. Presents findings for your review. Never publishes, shares, or presents research without your approval. Never states uncertain findings as facts.
Failure behavior	If a source is inaccessible (paywalled, broken link, deleted), cites the source and notes that the content could not be verified. If a claim cannot be sourced, moves it to the uncertainty section.
Review criteria	A reader should be able to verify any factual claim by following the source link. No claim in the main body should lack a source. The uncertainty section should be honest about what the research does not cover.
Graduation trigger	When research briefs consistently contain sources you can verify and uncertainty sections that are genuinely useful, the is working.
Privacy tier	Inherits from the sources. Research involving confidential data or client information inherits a higher .

Three tiers of evidence keep claims honest

Every claim the assistant makes falls into one of three evidence tiers. The tier determines how the claim should be presented in the research brief.

Three tiers of evidence

Tier	What it means	How to present it
Sourced	The claim is supported by a specific, citable source the reader can verify.	State the claim and cite the source: 'Remote teams report 23% fewer interruptions (Buffer State of Remote Work, 2025).'
Inferred	The claim is a reasonable conclusion from multiple sources, but no single source states it directly.	State the claim and note the inference: 'Combining the interruption data with the productivity study suggests a net gain, though no single study confirms this directly.'
Uncertain	The claim is plausible but lacks strong evidence, or the sources conflict.	Move to the uncertainty section: 'Whether this net gain holds for teams under ten people is unclear. Two studies suggest yes; one small study found no difference.'

The uncertainty section is the most valuable part of a source-grounded brief. It tells you where your knowledge is thin, where sources conflict, and where the assistant is speculating. Without it, the brief presents everything with equal confidence, which makes the weakest claims indistinguishable from the strongest.

Uncertainty language prevents false confidence

The assistant's default behavior is to write with confidence. When it says 'studies show that X,' it sounds like the evidence is settled. In practice, 'studies show' might mean one study from 2019 with a small sample, or it might mean a robust meta-analysis from last year. The language is identical. The evidence is worlds apart.

Source grounding requires calibrated language. The research brief should use different phrasing depending on the strength of the evidence:

Strong evidence: 'According to [source], X.' 'Multiple studies confirm X.' 'Data from [source] shows X.'
Moderate evidence: 'One study suggests X.' 'Limited data indicates X.' '[Source] found X, though the sample was small.'
Weak evidence: 'It is plausible that X, but the evidence is thin.' 'No reliable source confirms X.' 'The assistant infers X from [reasoning], but this has not been independently verified.'

When you train the assistant to use calibrated language, the research brief becomes a document you can trust at the sentence level. Strong-evidence sentences are reliable. Weak-evidence sentences tell you where to dig deeper. The language itself carries the confidence information.

The source-trail test catches claims that look cited but are not

After the assistant produces a research brief, run this check: pick three factual claims at random and follow their source trails. Can you find the original source? Does the source say what the brief claims it says? Is the source current?

If even one claim fails this test, the brief has a grounding problem. Either the assistant is paraphrasing loosely, citing sources it has not verified, or combining information from multiple sources and attributing the synthesis to a single source.

How this module breaks

What goes wrong	How you notice	What to correct once	What rule to save
The assistant launders a weak inference through confident language, making it read like a sourced fact.	A claim like 'research consistently shows that X' turns out to rest on a single blog post. The confident language hid the weak evidence.	Add a rule: every claim using language like 'research shows,' 'studies confirm,' or 'evidence suggests' must cite a specific source in the same sentence. General claims without a specific citation are moved to the uncertainty section.	Confidence language requires a specific citation in the same sentence. No claim can use 'research shows' or 'studies confirm' without naming the source.
The uncertainty section is empty because the assistant treated every claim as well-sourced.	The brief reads as if every question about the topic is settled. You discover later that several claims are contested or based on limited evidence.	Require the uncertainty section to contain at least one entry. If the assistant reports no uncertainty, ask it to identify the weakest claim in the brief and explain why it might be wrong.	Every research brief has a non-empty uncertainty section. If no uncertainty exists, the assistant identifies the weakest claim and explains its limitations.
Sources are outdated but still cited with confidence.	The brief cites a 2019 study as if its findings are current. The field has changed since then, and more recent data contradicts the claim.	Add a recency check: flag any source older than two years. For fast-moving fields, flag sources older than one year. The assistant notes the age and recommends checking for newer data.	Sources older than two years are flagged with their publication date. The assistant notes whether newer data may exist and recommends verification.

Your turn

Produce a source-grounded research brief

You'll produce a research brief on one real topic with inline citations, confidence levels, and an explicit uncertainty section.

Why this exercise matters

Research without source trails is indistinguishable from the assistant's guesses. Source grounding turns a polished summary into a document you can verify and trust.

You’ll leave with

A research brief with every factual claim cited.
Three tiers of evidence clearly distinguished (sourced, inferred, uncertain).
An uncertainty section identifying gaps and conflicts.
A source list with access information.

Use the prompt in order

1
State your research question and any sources you already have
Tell the assistant what you need to know and why. Provide any sources you have already read. The assistant will build on your existing knowledge rather than starting from scratch.
2
Require citations and evidence tiers
Tell the assistant to cite every factual claim, classify each as sourced, inferred, or uncertain, and use calibrated language that matches the evidence strength.
3
Review the uncertainty section first
The uncertainty section tells you where the research is thin. If the section is empty, push back. If the section identifies real gaps, decide whether those gaps matter for your decision.
4
Run the source-trail test on three claims
Pick three factual claims at random. Follow the citation. Verify that the source says what the brief claims. If any claim fails, ask the assistant to re-check all citations.

Prompt details

Each detail will be inferred, or you will be asked to clarify in the chat.

3/3 details

3 details will be inferred, or you will be asked to clarify in the chat

Starter prompt text

Open the full text if you want to check what will be copied.

Show starter prompt

Research this topic and produce a brief where every factual claim is linked to a source I can verify.

Topic: Infer from what you know about me or ask if it is unclear

Sources I already have:
Infer from what you know about me or ask if it is unclear

For every factual claim:
- Cite the specific source (author, title, date, link if available).
- Mark as sourced (directly stated in source), inferred (concluded from
  multiple sources), or uncertain (plausible but weakly supported).

- Use calibrated language: 'according to Infer from what you know about me or ask if it is unclear' for strong evidence,
  'one study suggests' for moderate, 'it is plausible but unconfirmed'
  for weak.


End the brief with an Uncertainty section: what this research does not cover, where sources conflict, and where the evidence is thin.

Include a source list at the end with enough information for me to find each source.

What the answer should give you

A research brief with: five sourced claims with inline citations, two inferred conclusions with reasoning explained, an uncertainty section with three gaps identified (limited data on firms under 10 people, conflicting evidence on onboarding length, no studies specific to your industry). Source list with six entries including titles, authors, dates, and links.