A pharmacovigilance writer is reviewing an AI-drafted PSUR signal section, paragraph by paragraph, the way she has reviewed prose her whole career: she reads each sentence, judges whether it sounds right, and moves on when it does. The third paragraph reads beautifully and states that the cumulative exposure since the international birth date is 1.2 million patient-years, a figure that is wrong by a factor she will not discover until the assessor questions it months later, because the sentence sounded exactly as authoritative as the true ones around it. The error was not that she missed a typo. It was that she read the paragraph before she checked its sources, and by the time she read it, the fluent prose had already convinced her eye. The show-your-sources pattern inverts that order. It forces the model to attach a source to every factual claim, it rejects any sentence that arrives without one, and it puts the source-link check before the writer reads the paragraph, so the writer verifies the claims while they are still naked facts and not yet dressed in persuasive prose. This lesson builds that pattern and the verification cadence that makes it work.

Why Fluency Defeats the Careful Reader

The reason a careful reader misses an invented number is mechanical, and it was established in the first level: the model writes a true value and a fabricated one in exactly the same even, confident tone, because confidence in an LLM is a property of the prose, not of the underlying truth. There is no linguistic tell. A hedged human writer signals uncertainty with words like "approximately" or "preliminary"; the model tends to state everything with the same flat assurance, so the reader cannot triage by tone and instead reads every claim as equally settled. When the prose is fluent, the reader's natural posture is trust, and trust is precisely the wrong posture for content that has to reconcile to source.

This is why rereading is a weak control. Rereading checks the prose against itself, against its own internal logic and flow, which the model is excellent at making coherent even when the facts are wrong. A self-consistent paragraph can be entirely fabricated and still pass a reread, because the reread never leaves the page to ask whether the claims are true. The only control that works is one that leaves the page on every claim and checks it against the actual source, and that control is far cheaper to run before the prose has done its persuasive work than after. The show-your-sources pattern is the operational form of that insight: make the sources visible, check them first, read the prose last.

The Pattern: A Source for Every Factual Claim

The core instruction of the pattern is simple to state and demanding to enforce: every factual claim in the draft must carry an inline citation to the specific source document and locator it came from. A factual claim is any sentence or clause asserting something about the trial, the product, or the data: a number, a population, a direction, an event count, a comparison, a date, a coded term. The citation must point at something checkable, a named loaded source and a precise location within it, such as "(CSR Table 14.2.1.4)" or "(IB Section 5.3, p. 41)" or "(TLF 14.3.1.2, safety population)", not a vague "(per the CSR)" that sends the reviewer hunting through six hundred pages.

The precision of the locator is what makes the pattern work, because a citation is only useful to the degree it can be checked quickly. A claim tagged "(CSR Table 14.2.1.4)" can be reconciled in seconds by opening that one table; a claim tagged "(clinical data)" cannot be reconciled at all and is functionally uncited. So the pattern instructs the model not merely to cite, but to cite to a resolvable locator, and where the loaded sources were labeled in the prompt, as the prompt-anatomy lesson recommended, the model can cite to "Source C, Table 14.2.1.4," which the writer maps directly to the document in front of them. The citation is the handle the verification step grabs, and a citation without a precise locator is a handle that comes off in your hand.

The pattern also distinguishes between claims that need sources and the connective tissue that does not. Structural scaffolding, headings, transitions, and conventional phrasing carry no factual assertion and need no citation; demanding a source for "The following section summarizes the key efficacy findings" produces noise. The discipline is to require citations for the load-bearing claims, the facts a reviewer would challenge, and to leave the scaffolding uncited, so that the citations that appear are signal. A draft where every claim that matters carries a resolvable source, and nothing else does, is a draft a reviewer can verify systematically rather than impressionistically.

Reject Any Sentence Without a Citation

The second half of the pattern is the rejection rule, and it is what gives the citation requirement teeth. A factual sentence that arrives without a citation is not a minor lapse to be tolerated; it is treated as unverified and therefore unusable until a source is supplied or the claim is removed. The model is instructed that if it cannot ground a factual claim in a provided source, it must not state the claim as fact; it must instead flag it, with the NEEDS SOURCE marker from the prompt-anatomy lesson, so the absence becomes visible rather than silent. An uncited factual sentence is the model telling you, if you are listening, that it could not find a source, and the rejection rule makes that admission impossible to overlook.

This rule does real work against the most dangerous behavior, which is the model's tendency to manufacture a citation rather than admit it has none. Left to its defaults, a model asked to cite everything will sometimes invent a plausible locator to satisfy the instruction, which is worse than no citation because it looks resolved. The pattern counters this by pairing the citation requirement with the explicit permission, indeed the instruction, to flag rather than fabricate: it is always acceptable to write NEEDS SOURCE, and never acceptable to invent a citation. Framing the flag as the correct, sanctioned behavior removes the pressure that drives fabrication, and a draft peppered with honest NEEDS SOURCE flags is far safer than a draft with a clean-looking citation on every line, because the flags tell you exactly where to look and the fabricated citations tell you nothing while pretending to tell you everything.

The rejection rule extends to the reviewer's own acceptance criteria, not just the model's behavior. A sentence whose citation cannot be resolved, because the cited table does not exist or does not contain the claimed value, is rejected on the same principle: an uncheckable or false citation is treated as no citation, and the claim reverts to unverified. This is the discipline that catches the fabricated cross-reference at the desk rather than at the assessment, and it follows directly from the rule that a claim is only as good as a source you can actually open and confirm.

The Verification Cadence: Source-Link Check Before You Read

The pattern is incomplete without its cadence, the specific order in which the writer engages the draft, and the cadence is its most counterintuitive and most important element: the source-link check comes before the writer reads the paragraph for sense. The writer does not read the signal section as prose and then spot-check a few numbers. The writer first walks the citations, claim by claim, opening each cited source and confirming the value, the population, the direction, and the date against the actual table or section, and only after a claim's source is confirmed does the writer let themselves read the sentence as language. The verification precedes the comprehension, on purpose.

The reason for this ordering is the fluency trap from the opening. Once the writer has read the fluent paragraph, the prose has already done its persuasive work, and the confirmed-feeling of a well-written sentence contaminates the subsequent source check, making the writer more likely to glance at the table and see what they expect to see. By checking the source first, while the claim is still a naked fact with a locator attached and no surrounding rhetoric, the writer confronts the table without the prose's thumb on the scale. The cadence is a debiasing procedure: it puts the cold, checkable fact in front of the eye before the warm, persuasive sentence gets a chance to bias the judgment. Verify the source, then read the prose, never the reverse.

Operationally, the cadence turns the claim-to-source table from the prompt-anatomy lesson into a worklist. The writer takes each row, resolves its locator, marks the claim verified or flags a discrepancy, and only when every row is verified or resolved does the section move to the language-and-style read. A discrepancy, a cited value that does not match the source, a population swap, a citation to a table that does not exist, halts acceptance of that claim and routes it back for correction, exactly as the rejection rule requires. The cadence is what converts the pattern from a formatting convention into a verification protocol, and it is the cadence, not the citations alone, that catches the 1.2 million patient-years before the assessor does.

Why This Survives an Inspection and Rereading Does Not

The show-your-sources pattern produces a draft whose verification is visible, systematic, and reproducible, which is what makes it defensible in a way that a carefully reread draft is not. When every factual claim carries a resolvable locator and the writer has walked each one against its source, the resulting artifact, the prose plus the claim-to-source table with each row marked verified, is itself a record of the verification performed. An assessor or an auditor can see not only the conclusion but the reconciliation behind it, and a sponsor can show that the human verification step was real and structured rather than an impressionistic skim. This is the operational embodiment of the transparency and human-led-decision principles the system prompt encodes.

Contrast the reread. A writer who reread the PSUR section three times and missed the 1.2 million patient-years has no record of having checked anything, because rereading leaves no trace and verifies no claim against a source. If the error reaches the assessor, the sponsor cannot demonstrate that any meaningful verification occurred, and the defense collapses into "our writer is experienced," which is not a control. The show-your-sources pattern replaces that with a defensible artifact: claims grounded, locators resolved, discrepancies routed, the whole thing legible to a reviewer who reads for conformance. The pattern is not extra work bolted onto drafting; it is the verification work made visible, ordered correctly, and turned into the record the regulation expects, and once a writer adopts the cadence of checking the source before reading the sentence, the fluency trap that defeated the careful reader loses its power.

Building the Pattern Into the Prompt and the Output Format

The show-your-sources pattern is not a habit a writer applies after the fact; it is an instruction set written into the prompt and the output format so the model produces a self-verifying artifact rather than a wall of prose the writer must instrument by hand. At the prompt layer, the instruction has three parts that work together. First, every factual claim must carry an inline citation to a named loaded source and a resolvable locator, in a fixed format the writer can scan, such as the source label and table number the prompt-anatomy lesson established. Second, any claim the model cannot ground in a provided source must be written with the NEEDS SOURCE flag rather than stated as fact or supported by an invented citation. Third, the model must produce, after the prose, a claim-to-source table that lists each factual claim, its source, its locator, and a column for the writer's verification status. These three instructions turn the draft into a document that arrives with its own verification harness attached, and the writer's job becomes walking the harness rather than constructing one.

The claim-to-source table is the operational heart of the pattern, and its design determines whether the verification cadence is fast or painful. A well-formed table has one row per load-bearing claim, with the claim text in the writer's own scannable form, the named source, the precise locator, and an empty status column the writer fills with verified, discrepancy, or NEEDS SOURCE as they walk each row against the actual document. Because the table is generated by the same model that wrote the prose, it inherits the same risk: the model can populate a row with a locator it invented, exactly as it can invent an inline citation, so the table is not a substitute for checking the source but a worklist that organizes the checking. The value is that it makes the verification finite and ordered. Instead of reading six hundred pages of prose and hoping to notice the claims, the writer has an enumerated list of every claim that needs a source, and the cadence becomes a matter of resolving rows rather than scanning text, which is both faster and far harder to do incompletely.

There is a subtle but important interaction between the table and the rejection rule. A claim that appears in the prose but not in the claim-to-source table is itself a red flag, because it means the model asserted a fact it did not even attempt to source, and a claim that appears in the table with a NEEDS SOURCE marker is the model honestly telling the writer where the evidence ran out. Both are signals the writer acts on, and both are invisible in a draft without the table. The discipline, then, is to require the table in the output format of every drafting prompt for regulated content, so that the verification harness is never optional and never something the writer has to remember to build, because the failure mode of the careful reader is precisely that, under time pressure, they skip the construction step and fall back on rereading. Baking the table into the output format removes the temptation, because the harness is already there when the draft arrives.

When the Pattern Meets the Rest of the Workflow

The show-your-sources pattern does not live in isolation; it is the verification layer that the rest of the program's workflows plug into, and seeing how it connects clarifies why it is taught early and reused everywhere. The prompt-anatomy lesson built the five-component prompt that loads and labels sources and requests a claim-to-source table; this pattern is what the writer does with that table once it arrives. The system-prompt lesson encoded citation discipline as standing behavior across every draft; this pattern is the per-document cadence that cashes that behavior out into an actual reconciliation. Downstream, the Module 2.5 drafting lesson, the IND form lessons, the IB update, and the health-authority correspondence lessons each generate factual claims that must reconcile to a source, and each one applies this same cadence: cite every claim, reject the uncited, check the source before reading the prose, and keep the marked-up table as the record. The pattern is the connective tissue that makes the whole program one discipline rather than a series of disconnected tasks.

The cadence also scales differently than rereading does, which matters when the document is long or the deadline is short. Rereading a long section gets slower and less reliable as fatigue sets in, because every pass is a full read and the reader's attention degrades. Walking a claim-to-source table degrades more gracefully, because the work is chunked into independent rows that can be paused, resumed, divided among reviewers, or audited after the fact, and a half-finished table visibly shows which claims remain unverified rather than leaving the writer to wonder whether they checked everything. Under deadline pressure, the temptation is always to skim, and the table is what makes skimming visible: an unfinished status column is a record that the verification was incomplete, which a writer cannot honestly sign over. The pattern thus protects the writer from their own time pressure by making the shortcut leave a trace.

Finally, the pattern changes the economics of AI-assisted drafting in a way that is easy to miss. The naive view is that the model's speed is the whole benefit and verification is a tax on it; the show-your-sources view is that the model's real contribution is producing a structured, claim-enumerated draft that makes verification tractable, and that a draft without a verification harness is not actually faster, because the time the model saved in drafting is spent, and often overspent, reconstructing what the claims are and where they came from. A draft that arrives with every claim cited to a resolvable locator and a table that enumerates them is a draft a writer can verify in a fraction of the time it would take to verify an unstructured one, so the pattern is not a tax on the acceleration but the thing that makes the acceleration real. The model that shows its sources is doing more of the work than the model that merely writes well, because it is doing the part that makes the writing safe to use.

Key Takeaways

• Fluency defeats the careful reader because the model states true and fabricated claims in the same even tone, with no linguistic tell. Rereading is a weak control because it checks the prose against itself; a self-consistent paragraph can be entirely fabricated and still pass. The only control that works leaves the page on every claim and checks it against the actual source.
• The pattern requires an inline citation with a resolvable locator for every factual claim: a named source and a precise location such as "(CSR Table 14.2.1.4)", not a vague "(per the CSR)". The locator is the handle the verification step grabs, and a citation without a precise locator is a handle that comes off in your hand. Scaffolding and transitions need no citation, so the citations that appear are signal.
• The rejection rule treats any factual sentence without a resolvable citation as unverified and unusable. It is always acceptable to write NEEDS SOURCE and never acceptable to invent a citation, which removes the pressure that drives a model to fabricate a plausible locator. An uncheckable or false citation is treated as no citation, catching the fabricated cross-reference at the desk.
• The verification cadence puts the source-link check before the writer reads the paragraph for sense. Verifying the cold, checkable fact before the warm, persuasive sentence is a debiasing procedure that keeps the prose's confidence from contaminating the source check. Verify the source, then read the prose, never the reverse.
• The pattern produces a defensible, visible verification record that rereading cannot. The prose plus a claim-to-source table with each row marked verified shows an assessor the reconciliation behind the conclusion, embodying the transparency and human-led-decision principles, where a reread leaves no trace and no control to point to.