Chapter 43: Capstone 2 — The Misinformation Tracker

"The correction almost never travels as far as the original lie. That's not a bug in the system. That's the feature misinformation exploits." — Sam Harding, ODA Lead Data Journalist, internal memo, October of the Garza-Whitfield race

A Note Before You Begin

This capstone is more uncomfortable than the first one. Building a polling audit — Capstone 1 — required you to evaluate other people's methodological choices. That's intellectually demanding, but the object of analysis is a technical artifact. A poll has a methodology. You assess the methodology.

Building a misinformation tracker requires you to make judgments about the truth of claims that real political actors have made. That's a different kind of responsibility. It touches questions that political scientists, journalists, and philosophers have debated for decades: Who decides what is false? What makes a claim "misleading" rather than merely "wrong"? Is exaggeration the same as misinformation? What happens when your tracker gets it wrong?

Sam Harding, ODA's lead data journalist, lost sleep over these questions when Adaeze Nwosu first tasked them with the project. "I've done investigative journalism for ten years," Sam told the ODA staff meeting where the project was announced. "I know how to verify a claim. What I'm not sure I know is how to do that at scale, in near-real-time, for a contested political race, without the system itself becoming a partisan weapon." That tension — speed versus rigor, scale versus judgment, accountability versus fairness — runs through every design decision in this capstone.

You will not resolve those tensions. But you will learn to work productively inside them.

Section 1: Introduction — The Misinformation Problem in the Race

Setting the Scene

It is six weeks before Election Day in the Garza-Whitfield Senate race. Adaeze Nwosu has called Sam Harding into her office at ODA's Atlanta headquarters — a converted loft space above a copy center, cluttered with dual monitors, annotated printouts of FEC filings, and a corkboard covered in index cards connected by red string that Adaeze insists is "visual thinking" and Sam insists is "the murder board."

"We're building the tracker," Adaeze says, handing Sam a two-page memo. "I've been putting it off because I was scared of getting it wrong. But every week we wait, the misinformation landscape gets worse, and we have no systematic way to document it. If we're going to be a credible accountability organization, we can't just write articles about individual lies. We need infrastructure."

Sam reads the memo. It outlines a public-facing dashboard — ODA's "Truth Tracker" for the race — that will catalog false and misleading claims from both candidates, rate them on an evidence-based rubric, track their spread across media platforms, and document whether corrections reached the same audiences as the original claims.

"This is going to make us enemies," Sam says.

"We already have enemies," Adaeze replies. "At least this way we'll have evidence."

That exchange captures the dual imperatives of the project you are about to build: rigor and courage. Rigor, because a misinformation tracker that makes errors destroys its own credibility and can itself become a vector for misinformation. Courage, because a tracker that only documents the obvious, the clear, the uncontested — a tracker designed to never make an enemy — is not actually tracking anything.

The Misinformation Landscape in the Race

To understand what ODA is trying to track, you need to understand what has actually circulated in the Garza-Whitfield race. Over the course of the campaign, ODA's preliminary monitoring (before the formal tracker existed) documented misinformation flowing in multiple directions.

From the Whitfield campaign and its allies, the most significant claims have included:

Claim W-1: "Maria Garza supports open borders." This claim appeared in Whitfield's television advertising, in statements by allied PACs, and was repeated by social media accounts associated with his supporter base. The factual record shows that Garza has explicitly called for comprehensive immigration reform including border security measures, opposed unlimited immigration, and — as Attorney General — prosecuted numerous cases involving undocumented persons for crimes. The claim relies on selective quotation of Garza's statement that she believes "everyone who comes to this country deserves to be treated with dignity," stripping the phrase from a speech that also emphasized the importance of legal immigration processes.

Claim W-2: "Garza's healthcare plan would eliminate private insurance." Garza has proposed a public option — a government-run insurance plan that would compete with private insurers. She has explicitly stated, in multiple forums, that private insurance would remain available. Whitfield's characterization is false as a description of her proposal, though advocates of universal single-payer systems have criticized Garza's plan precisely because it does not eliminate private insurance.

Claim W-3: "Violent crime rose 40 percent under Garza's tenure as AG." This claim appeared in a PAC television advertisement in the race's final stretch. It is based on a selective reading of crime statistics that cherry-picks a two-year period (2019-2021) that overlapped with the COVID-19 pandemic's documented effect on crime data across the country, ignores the longer trend under Garza's tenure, and compares numbers from incompatible measurement methodologies across those years.

From the Garza campaign and its allies, the misinformation has been different in character but present:

Claim G-1: "Tom Whitfield outsourced jobs to China." Garza's campaign has run this claim in both digital and television advertising. The factual record is mixed: Whitfield's hardware chain did source some products from Chinese manufacturers — as do virtually all hardware retailers — but there is no documented case of Whitfield's company closing a domestic manufacturing facility or laying off American workers to move production offshore. The claim conflates retail purchasing decisions (buying products made in China) with the specific practice of "outsourcing" (moving jobs from the U.S. to another country), which has a distinct and more serious connotation.

*Claim G-2: "Whitfield's plan would cut Medicare by $200 billion."* Whitfield has stated general support for "reforming entitlement programs" and endorsed a budget framework produced by a conservative think tank. That framework includes a Medicare savings target of $200 billion over ten years. Whether this constitutes a "cut" depends on baseline assumptions — the framework characterizes the savings as efficiency gains — and Whitfield has neither endorsed the specific number nor released his own Medicare plan. Garza's characterization is a reasonable inference from Whitfield's stated positions, but it presents as a committed plan something that is better characterized as a possibility consistent with his stated general principles.

Claim G-3: "Whitfield said immigrants 'don't belong here.'" This claim circulated on social media associated with Garza supporters. The documented statement by Whitfield, from a campaign event in a rural county, was: "Illegal immigrants who are here breaking the law don't belong here — they need to go back." The social media circulation dropped the qualifier "illegal immigrants who are here breaking the law," transforming a statement about undocumented persons who have committed crimes into a blanket statement about all immigrants.

Why This Landscape Matters

Notice several things about this misinformation landscape:

First, both sides engage in it. Not symmetrically — the claims differ in character, severity, and volume — but any credible misinformation tracker must document claims from across the political spectrum. ODA's credibility depends on its willingness to track Garza's misleading claims with exactly the same rigor it applies to Whitfield's.

Second, the claims exist on a spectrum. Some claims are simply false (W-1 as stated is flatly contradicted by Garza's documented record). Others are exaggerations of real tendencies (G-2 describes a real risk from real policy preferences). Others are misleading through selective context (W-3's selective time period). Others are distortions of real statements (G-3's quote truncation). A single binary "true/false" judgment cannot capture this complexity.

Third, framing and context are load-bearing. As we explored in Chapter 24, the way a claim is presented — the emphasis, the comparison class, the omissions — can make a true statement misleading. This means the tracker cannot simply verify atomic factual claims in isolation; it must analyze the communicative context.

Fourth, the claims have different origins and spread through different pathways. W-1 originated in Whitfield's own advertising. G-3 originated among supporters and was never stated by Garza herself. Tracking "campaign misinformation" must account for this ecosystem complexity.

The Wardle-Derakhshan Typology Applied

Claire Wardle and Hossein Derakhshan's foundational typology (developed for the Council of Europe in 2017) distinguishes among:

• Mis-information: False information shared without intent to harm (honest error)
• Dis-information: False information shared with intent to deceive
• Mal-information: True information shared with intent to harm

Within disinformation, they identify seven types of content: satire/parody, misleading content, imposter content, fabricated content, false context, manipulated content, and false connection.

For ODA's tracker, this typology provides the conceptual foundation but requires adaptation to the political context. The tracker will use a modified four-category system:

Category 1 — Fabricated/False: Claims that are straightforwardly contradicted by documented evidence. The claim makes a factual assertion; the evidence shows the assertion is wrong. (Example: W-3 on crime statistics.)

Category 2 — Misleading Context: Claims where the underlying facts have some basis but the framing, omissions, or context create a fundamentally distorted impression. (Example: G-1 on outsourcing.)

Category 3 — Contested Inference: Claims where a reasonable person could reach the stated conclusion from available evidence, but the claim presents a contested interpretation as established fact. (Example: G-2 on Medicare cuts.)

Category 4 — Out-of-Context Quote: Cases where an accurate statement is stripped of qualifying context in ways that materially change its meaning. (Example: G-3 on Whitfield's immigration statement.)

This four-category system is more nuanced than binary true/false but simple enough to be consistently applied across reviewers and explained to a general audience.

Design Principles for a Credible, Nonpartisan Tracker

Before any line of code is written, before any claim is logged, ODA establishes four design principles that will govern every subsequent decision:

Principle 1 — Proportionality: The tracker rates claims by the severity of the distortion relative to documented evidence. It does not treat "an exaggeration" and "a fabrication" as equivalent.

Principle 2 — Symmetry: Claims are tracked from both campaigns and their allies. The tracker has no minimum threshold that applies to one side but not the other. Sam will flag for Adaeze's review any situation where the claim log shows a significant asymmetry across a two-week window — not because asymmetry is necessarily wrong, but because it requires additional scrutiny to ensure it reflects reality rather than selection bias.

Principle 3 — Transparency: Every rating includes a full explanation, links to all sources consulted, and a clear statement of what would change the rating. Readers can see exactly how ODA reached its conclusion.

Principle 4 — Humility: The tracker includes an explicit "correction" mechanism. If ODA gets a rating wrong — if new evidence emerges, if a campaign provides documentation that changes the picture — the correction is published prominently, not buried, and the original rating is updated with a visible notation of the change.

These principles will be tested repeatedly as the project unfolds. The tracker's credibility depends not on its infallibility but on its transparency about its own fallibility.

Section 2: System Architecture and Data Collection

What a Misinformation Tracker Needs to Track

A complete misinformation tracking system must capture several distinct types of information across the lifecycle of a false or misleading claim:

The claim itself: The exact wording of the claim, as it appeared in its original form. Not a paraphrase. Not a summary. The actual statement, with direct attribution, date, venue, and speaker/sponsor.

The claim's source: Who made the claim? A candidate directly? A campaign advertisement? An allied PAC? An unofficial social media account supportive of the campaign? Each source type has different implications for how the claim should be characterized and whether it can be attributed to the campaign.

The verification record: What evidence was consulted? What did it show? Who were the subject-matter experts contacted? When was the verification conducted? This is the core of the tracker's evidentiary chain.

The rating: The Category 1-4 classification plus a plain-language explanation suitable for a general audience.

The spread data: Where has this claim appeared? In what volume? At what velocity? On which platforms and media types? This is where the media monitoring component becomes essential.

The correction record: Has the claim been corrected by a major fact-checking outlet? By the campaign itself? By the original media source? If corrections exist, what is their documented reach relative to the original claim?

The resolution status: Is the claim still circulating? Has it become dormant? Has it been substantially corrected in the media ecosystem?

Tracking all of this for dozens of claims across a multi-week election campaign is a significant operational undertaking. This is why the system architecture matters: without automated data collection and structured data storage, the human effort required would exceed any realistic staffing capacity.

Data Sources: Building the Intake Pipeline

ODA's tracker draws on six primary data source categories, each with its own collection methodology:

Source Category A — Social Media Monitoring: The tracker uses platform APIs (where available) and ODA's existing social media scraping infrastructure to monitor keyword-flagged content across major platforms. The keyword lists are seeded with names (Garza, Whitfield), known claim phrases ("open borders," "outsourced jobs"), and related policy terms. The social media monitoring does not capture all misinformation — platforms' API restrictions and the volume of content make comprehensive coverage impossible — but it captures trending claims that reach threshold volumes.

The ODA media dataset (oda_media.csv) includes a subset of social media-adjacent content: articles from digital-native outlets and fact-checker publications. The full social media monitoring pipeline supplements this with direct platform data.

Source Category B — News Database Monitoring: ODA maintains API access to two major news aggregation services. The tracker runs daily queries against these services for all coverage mentioning Garza, Whitfield, their campaigns, and specific factual claims flagged for monitoring. Results are ingested into the claims database and tagged by source, date, and sentiment.

Source Category C — Fact-Checker RSS Feeds: ODA aggregates RSS feeds from PolitiFact, FactCheck.org, Washington Post Fact Checker, AP Fact Check, and two regional fact-checking outlets that cover the race's state. These feeds provide both a source of pre-verified claims (when external fact-checkers have already worked a claim) and an independent check on ODA's own ratings (disagreements between ODA and established fact-checkers are flagged for review).

Source Category D — Campaign Ad Monitoring: ODA accesses the FCC's Public Inspection File and supplementary sources to track television advertising buys. Digital advertising is monitored through platform transparency tools, where available, and through ODA's own ad archiving process for digital platforms. All campaign and PAC advertisements are logged and their factual claims extracted.

Source Category E — Candidate Statements and Official Communications: Press releases, campaign website content, social media posts by the official candidate accounts, and documented statements from press availabilities and campaign events. These are the primary source for candidate-attributed claims.

Source Category F — Expert and Document Sources: For claim verification, ODA maintains a network of subject-matter experts (immigration law, healthcare policy, criminal justice statistics, economic analysis) who serve as consultants for specific claims. Government data sources (Census, BLS, FBI UCR, CMS) are used for claims involving statistics. Legislative records, court documents, and company filings provide documentation for claims about candidates' records.

Building the Claim Taxonomy

Not every statement by a campaign qualifies as a "claim" for tracker purposes. Political rhetoric includes exaggeration, emphasis, argument, and advocacy — all legitimate features of democratic discourse. The tracker's first challenge is distinguishing verifiable factual claims from non-verifiable rhetorical moves.

ODA's claim taxonomy uses a decision tree:

Step 1 — Is the statement verifiable? A claim must make an assertion about a fact that can, in principle, be verified by reference to evidence. "Tom Whitfield is wrong for this state" is not verifiable — it's an evaluative judgment. "Tom Whitfield's company sourced products from China" is verifiable. "Maria Garza supports open borders" presents as verifiable (it's a claim about her policy position) but requires additional interpretation because "open borders" is both a specific policy concept and a political pejorative.

Step 2 — Is the statement material? The claim must be significant enough to matter to voter decision-making. Trivial errors (a candidate misstates a date by one year in passing) are noted but not tracked with full rigor.

Step 3 — Is the statement attributable? The tracker can only rate claims with a clear attribution chain. Rumors circulating anonymously on social media may be worth monitoring for spread data, but they cannot be "rated" in the same way as a claim in a television advertisement.

Step 4 — Is the statement already adjudicated? If a major fact-checker has already published a thorough analysis, ODA can incorporate that analysis with attribution rather than duplicating the full verification process. This preserves ODA's bandwidth for claims that have not yet received fact-checker attention.

Nonpartisanship Protocols

The most operationally demanding aspect of a nonpartisan tracker is ensuring that the threshold for claim entry is applied consistently. Political misinformation is not symmetric — different campaigns make different kinds of claims at different rates, and a nonpartisan tracker will, in fact, find asymmetric results if reality is asymmetric. But the process must be symmetric, or the tracker's integrity is compromised.

ODA's nonpartisanship protocols specify:

Blind review for threshold decisions: The analyst who makes the initial decision to enter a claim into the tracker's pipeline does so after redacting the candidate attribution. Sam reviews the claim text and source type without seeing the candidate's name. Once the claim is entered into the pipeline, the attribution is restored for the verification process.

Weekly balance review: Each Friday, Sam and Adaeze review the week's intake. If the week's new claims are significantly skewed toward one candidate (more than a 60/40 split), this triggers a documented review: Is the imbalance explained by one campaign having made more verifiable false claims? Are there comparable claims from the other campaign that did not meet the entry threshold but should be reconsidered? This is not a mandate for artificial balance — it's a mandate for conscious justification of imbalance.

External review panel: ODA convenes a three-person advisory panel (one conservative-leaning political scientist, one liberal-leaning political scientist, one independent journalist) who review the tracker's ratings on a biweekly basis. Their role is not to override Sam's ratings but to flag cases where the reasoning is unclear or the evidence is ambiguous.

Limitations and Scope Decisions

Honesty about what the tracker cannot do is essential to its credibility. ODA publishes a limitations statement prominently on the tracker's public page:

The tracker does not claim to capture all misinformation in the race. Social media volume alone makes comprehensive coverage impossible. The tracker focuses on claims that: (1) reach documented threshold audiences (defined as appearing in broadcast advertising, major news coverage, or achieving measurable viral spread on monitored platforms); (2) make verifiable factual assertions; and (3) can be attributed to a specific source with sufficient documentation.

The tracker does not rate opinion, political argument, or interpretation as misinformation, even where ODA might disagree with the analysis. The tracker is explicitly limited to factual claims — matters that have a correct answer by reference to evidence.

The tracker does not adjudicate disputes between credentialed experts. Where genuine expert disagreement exists on a technical question (for example, the economic effects of a proposed policy), the tracker notes the disagreement and does not issue a rating.

Section 3: The Claim Identification Pipeline

How False and Misleading Claims Enter the Pipeline

Sam's claim identification pipeline operates in three parallel streams that feed a central review queue.

Stream 1 — Automated Flagging: A keyword-monitoring script (described in detail in Section 5) runs against ODA's incoming media and social media data. It flags content that contains pre-seeded claim phrases, names combined with known misinformation keywords, and statistical claims that can be cross-referenced against ODA's data. Automated flags are not claims — they are candidates for human review.

Stream 2 — Proactive Monitoring: Sam and ODA's two research associates monitor candidate events, press releases, and advertising in real time. They use a standardized claim intake form to log any potential claim as they encounter it. The form captures: claim text, source, date, initial category assessment, and supporting/contradicting evidence links.

Stream 3 — External Referrals: ODA's public email address and a dedicated tracker tip line receive referrals from journalists, researchers, campaign workers (anonymously), and members of the public. These are reviewed but handled more carefully: anonymous tips require independent corroboration before any claim is entered.

All three streams feed into the same review queue. Sam triages the queue daily, typically in the morning. Claims that clear the taxonomy decision tree (Section 2) are moved to the verification stage. Claims that don't clear it are archived with a brief notation.

Manual vs. Automated Identification

The appropriate balance between manual and automated identification depends on what you're trying to catch and what you can afford to miss.

Automated identification excels at breadth. A well-designed script can monitor millions of pieces of content that no human team could read. It will catch high-volume spread of flagged claims and identify emerging claims that are gaining traction before they break into mainstream coverage. But automated identification is weak on context. It cannot distinguish between a news report accurately quoting a false claim (for the purpose of fact-checking it) and an original propagation of the claim. It cannot catch novel claims that don't match pre-seeded keywords. And it will generate significant false positives — content that triggers the keyword filter but is not actually a claim that requires tracking.

Manual identification excels at precision. An experienced journalist who watches a campaign event can identify a misleading claim even when its phrasing is entirely novel, because they have context about the candidate's record, the factual literature, and the campaign's communications strategy. But manual identification has severe bandwidth constraints. Sam cannot watch every campaign event, read every press release, and monitor every social media account.

ODA's solution is to use automated identification for breadth and volume monitoring, and manual identification for the initial entry decision. The automated system is designed to cast a wide net and generate false positives — that's acceptable, because the false positives are caught in human review. The manual system is reserved for the judgment calls that require context.

A practical implication: the tracker's claim database will not contain every false statement that circulates in the race. It will contain every false statement that ODA has evaluated using this process. Those are different things, and the tracker says so explicitly.

Claim Verification Process

Once a claim is in the review queue and cleared the taxonomy filter, Sam begins the verification process. This process has four stages:

Stage 1 — Primary Source Compilation: Identify all primary sources directly relevant to the claim. If the claim is about crime statistics, the primary sources are the actual crime statistics from the relevant agency. If the claim is about a candidate's record, the primary sources are legislative votes, court documents, official statements, and contemporaneous news coverage. Sam does not begin with secondary analysis; secondary sources (including other fact-checkers' work) are consulted after primary sources are established.

Stage 2 — Expert Consultation: For claims involving specialized knowledge — healthcare policy, immigration law, criminal justice statistics — Sam contacts at least one subject-matter expert. These consultations are documented: who was contacted, when, and what they said. Expert opinions are quoted with attribution (or anonymously, if the expert requests it, with a notation explaining why attribution is withheld).

Stage 3 — Campaign Response: Standard journalistic practice requires giving the subject of a fact-check an opportunity to respond. ODA emails the relevant campaign (or PAC) with a specific request: "We are preparing a fact-check of the following claim. Please provide any documentation or context you believe is relevant." Responses (or non-responses, noted as such) are incorporated into the published rating.

Stage 4 — Collegial Review: Before a rating is published, a second ODA analyst or Adaeze herself reviews the evidence file and the proposed rating. This is not to override Sam's judgment — Sam has final authority on the rating — but to catch logical errors, missing sources, or ambiguous reasoning.

The verification process takes, on average, three to five days for complex claims. Simpler claims (where primary source evidence is unambiguous) can be turned around in 24-48 hours. This timeline creates a genuine tension with news cycles, which Sam and Adaeze discuss in detail in Section 7.

Rating Systems: Truth-O-Meter, Pinocchios, and ODA's Custom Rubric

Three major fact-checking rating systems exist in the American political landscape, and ODA must decide how to position the tracker relative to them.

PolitiFact's Truth-O-Meter uses six ratings: True, Mostly True, Half True, Mostly False, False, and "Pants on Fire" (for egregiously absurd false claims). The system is widely recognized but has faced criticism for imposing a misleading symmetry — "Half True" suggests a claim is half accurate and half false, when in reality many claims are true in a narrow literal sense but misleading in their communicative intent.

The Washington Post Fact Checker's Pinocchio Scale uses one to four Pinocchios for false or misleading claims, with the Geppetto Checkmark for true statements and a Kessler Collection for repeated false claims. The Pinocchio metaphor is culturally resonant but implies intent (Pinocchio lies) in a way that may not accurately characterize all misleading claims.

AP Fact Check does not use a rating scale at all — it simply reports what the claim says and what the evidence shows, leaving readers to draw their own conclusions. This approach maximizes transparency but sacrifices the at-a-glance communication value of a rating.

ODA's Custom Rubric is designed to address the limitations of each system while remaining simple enough for general audience comprehension. The rubric uses two dimensions:

Dimension 1 — Accuracy (A1 through A4): - A1 — Documented False: The claim is directly contradicted by primary source evidence with high confidence. - A2 — Materially Misleading: The claim contains true elements but omits or distorts context such that the overall impression conveyed is significantly inaccurate. - A3 — Contested/Ambiguous: The claim is a reasonable interpretation of ambiguous evidence, but is presented as established fact rather than interpretation. - A4 — Largely Accurate: The claim is supported by available evidence with minor errors or imprecision.

Dimension 2 — Impact (I1 through I3): - I1 — High Impact: The claim appeared in paid advertising, major news coverage, or achieved documented viral spread (threshold: 100,000+ estimated impressions). - I2 — Medium Impact: The claim appeared in candidate statements, press releases, or earned media coverage with significant but not massive reach. - I3 — Low Impact: The claim appeared in a limited-reach context (a small campaign event, a minor social media post) and has not spread significantly.

A claim rated A1-I1 (documented false, high impact) receives the tracker's highest attention and most prominent display. A claim rated A3-I3 (contested interpretation, low reach) is logged but does not receive feature treatment.

This two-dimensional system allows the tracker to communicate both the severity of the distortion and the magnitude of the harm, which are distinct questions. A documented falsehood that reaches five people is less damaging to the information environment than an ambiguous claim that reaches five million.

Designing for Reproducibility and Transparency

Sam is a data journalist with a background in computer science, and they bring to the tracker the standards of reproducible research: every rating should be reproducible by an independent analyst working from the same evidence. This doesn't mean every analyst will reach exactly the same rating — judgment is involved — but it means the evidence chain is transparent enough that another analyst can understand exactly how ODA reached its conclusion and identify the specific judgment calls where they might disagree.

In practice, this means: all sources linked (not just cited), all expert consultation documented, all campaign responses included verbatim, all data used in statistical fact-checks published as downloadable files, and a published methodology document that explains the rating system in enough detail that a reader could apply it themselves.

The tracker's publicly visible pages show the final rating and a plain-language explanation. A "Methodology Details" tab shows the full evidence chain. This layered presentation serves both general readers (who need the plain-language summary) and specialist readers (journalists, researchers, campaigns) who want to interrogate the underlying evidence.

Section 4: Spread Analysis

Measuring Claim Spread: Reach, Velocity, and Amplification

A false claim that reaches one hundred people and a false claim that reaches ten million people are both false claims, but they are not equally damaging to the political information environment. Spread analysis — measuring the reach, velocity, and amplification of claims — is what allows the tracker to distinguish between these cases and prioritize accordingly.

Reach is the total estimated audience exposed to a claim across all channels. Reach is imprecise — we cannot know exactly who has seen a piece of content — but it can be estimated using platform-provided metrics (where available), advertising buy data, and news audience research. For broadcast television, Nielsen ratings data provides reach estimates. For digital content, platform analytics and third-party tools provide engagement estimates. For social media, retweet/share counts provide a proxy for spread (though they undercount organic reach).

Velocity is the rate at which a claim is spreading at a given point in time. A claim that went from 1,000 to 1,000,000 impressions in 48 hours has a very different velocity profile than one that accumulated the same total reach over three weeks. High velocity is an early warning indicator: a rapidly spreading claim may need an expedited fact-check response even if the total reach hasn't yet reached the high-impact threshold.

Amplification refers to who is spreading the claim and what kind of attention they're bringing to it. A claim amplified by major television networks, even as part of a fact-check, reaches a larger and more politically heterogeneous audience than the same claim spreading virally in a partisan social media ecosystem. The nature of the amplification affects the corrective response strategy.

ODA's spread analysis for the tracker pulls from three data sources: the oda_media.csv dataset (tracking news coverage), supplementary social media monitoring data (tracking platform spread), and advertising buy records from public FCC files and platform transparency tools (tracking paid amplification).

Network Pathways: From Origin to Mainstream

One of the most important insights from misinformation research is that false claims rarely travel directly from their origin point to mass audiences. They move through networks, gaining legitimacy (or at least familiarity) at each step. Understanding these pathways is essential to the tracker's spread analysis.

The typical pathway for a major political misinformation claim in the Garza-Whitfield race looks something like this:

Stage 1 — Origin: The claim is made in a primary source: a candidate speech, a campaign advertisement, an allied PAC communication, or an organic post by a campaign-adjacent social media account.

Stage 2 — Partisan Amplification: Partisan media outlets, supportive social media accounts, and ideologically aligned email lists amplify the claim within a like-minded audience. At this stage, the claim is spreading but largely within an audience that was already inclined to believe it.

Stage 3 — Mainstream Pickup: Either the claim's volume becomes large enough to attract mainstream media coverage, or the campaign promotes it aggressively enough to force mainstream coverage. This is the pivotal moment — mainstream pickup dramatically expands reach to audiences that would not have encountered the claim in partisan channels.

Stage 4 — Fact-Checker Response: Major fact-checkers publish analyses. This generates its own coverage cycle, which — paradoxically — can further amplify the original claim even while correcting it. Studies consistently show that corrections reach smaller audiences than the original claims, but the claim-plus-correction cycle reaches larger audiences than the original claim alone.

Stage 5 — Background Radiation: After the initial cycle, the claim may settle into what researchers call "background radiation" — a persistent low-level presence in the information environment, where it continues to circulate among those who believe it and will emerge in response to new relevant events.

Sam's spread analysis for each tracked claim attempts to document where in this pathway cycle the claim currently sits and to estimate the cumulative reach at each stage.

The Correction Gap: How Often Corrections Reach the Same Audience

The correction gap is the most disheartening finding in misinformation research, and it is central to the tracker's public communication strategy.

Research across multiple studies consistently shows that corrections reach a fraction of the audience that the original claim reaches. A 2022 meta-analysis of correction studies found that corrections typically reach approximately 20-30 percent of the original claim's audience when the correction comes from the same source and platform. Cross-platform corrections (fact-checker on one platform correcting a claim that originated on another) reach 5-15 percent of the original audience. This is not a technological problem that better design can solve; it reflects the basic dynamics of attention and sharing behavior: misinformation tends to be more emotionally arousing and shareable than correction, and people are less motivated to share information that challenges beliefs they hold.

This finding has profound implications for the tracker's design. If the correction gap cannot be closed by corrections alone, the tracker's value is not primarily in issuing corrections — other organizations already do that. The tracker's value is in documenting the gap itself: showing the public that Claim W-3 (the crime statistics claim) was exposed in a detailed fact-check that reached an estimated 180,000 people, while the original advertising claim had estimated reach of 1.4 million in the state. Making the correction gap visible is a form of accountability that goes beyond any individual fact-check.

Applied to Five Specific Claims in the Garza-Whitfield Race

Let's trace the spread of five claims through ODA's analysis framework.

Claim W-1 — "Maria Garza Supports Open Borders"

Origin: Whitfield television advertisement, aired beginning 42 days before Election Day. Also appeared as a Facebook/Instagram advertisement with an estimated $85,000 digital buy across the state. Campaign estimates suggest approximately 2.1 million television impressions across the state market and 340,000 digital impressions.

Partisan amplification: Three conservative talk radio stations in the state aired extended discussions of the claim in the 72 hours following the ad's launch. A PAC allied with Whitfield sent an email to its 45,000-subscriber list featuring the claim. Conservative Twitter/X accounts with an aggregate following of approximately 210,000 retweeted campaign-produced content featuring the claim.

Mainstream pickup: The claim was covered in a 90-second segment on the state's largest newspaper's political podcast (Tuesday following launch), in a print article in the metro daily that mentioned the claim while noting its disputed nature, and in a brief segment on one of the major network's local affiliates that characterized it as "a claim the Garza campaign disputes."

Fact-checker response: PolitiFact rated the claim "False" and published a 1,200-word analysis eight days after the ad first aired. The Washington Post Fact Checker declined to write a separate analysis, deferring to PolitiFact's prior work. ODA published its own rating (A1-I1) the same day as PolitiFact, with additional detail on the legal cases Garza prosecuted as AG.

Correction reach estimate: PolitiFact's rating received approximately 28,000 page views. ODA's rating received approximately 12,000 page views. The combined correction audience represents approximately 1.9 percent of the total estimated exposure to the original claim.

Current status: The claim continues to circulate on social media at a low level. The television advertisement is still running at reduced frequency as of the tracker's most recent update.

Claim W-3 — "Violent Crime Rose 40% Under Garza's Tenure as AG"

Origin: PAC television advertisement, aired beginning 28 days before Election Day. Estimated reach: 980,000 television impressions statewide. Digital version of the advertisement estimated at 195,000 impressions.

The underlying data: Violent crime in the state did increase from 2019 to 2021, the period the advertisement selects. However, the increase from 2018 (the year before Garza's tenure began) through 2022 (the last full year of her tenure) shows a net increase of 11 percent — not 40 percent. The 40 percent figure is derived by selecting 2019 as the baseline (a year in which crime was unusually low) and 2021 as the endpoint (a year in which crime was elevated nationwide due to pandemic-related disruptions). The FBI changed its crime data collection methodology in 2021, making direct 2019-2021 comparisons additionally unreliable. No criminologist consulted by ODA considered the 40 percent figure to be a valid characterization of crime trends under Garza's tenure.

Partisan amplification: The advertisement's claims were repeated by several Whitfield-allied social media accounts and one conservative statewide news aggregator. A viral social media post citing the "40 percent" figure received approximately 4,200 shares.

Mainstream pickup: The state's two major newspapers both ran fact-checks within 10 days of the advertisement's launch. The metro daily's fact-checker gave the claim "Three Pinocchios" (on the WaPo scale). The state's largest paper called the claim "misleading and statistically dubious."

ODA rating: A1-I1. The methodological manipulations (baseline selection, methodology change, cherry-picked period) make this a documented false characterization, not a contested interpretation.

Correction reach: Combined fact-checker readership estimated at 65,000. Original ad's estimated reach: 1.175 million. Correction coverage ratio: approximately 5.5 percent.

Claim G-1 — "Tom Whitfield Outsourced Jobs to China"

Origin: Garza campaign television advertisement, aired beginning 38 days before Election Day. Digital advertising component with an estimated $62,000 buy. Estimated reach: 1.6 million television impressions, 280,000 digital.

The underlying record: Whitfield's Hardware employs approximately 340 people across its eight-location regional chain. The company purchases approximately 35-40 percent of its products from manufacturers in China, Vietnam, and other Asian countries — standard practice across the retail hardware industry. ODA found no evidence that Whitfield's company ever closed a domestic manufacturing operation or explicitly displaced American manufacturing workers by shifting production offshore. The company does not manufacture products; it retails them. The "outsourcing" framing implies a manufacturing context that does not apply to a retail chain.

ODA rating: A2-I1. The claim is materially misleading: the underlying practice (purchasing from foreign manufacturers) is real, but the "outsourced jobs" framing implies a job-displacement dynamic not supported by the evidence.

Garza campaign response: "Tom Whitfield made the choice to stock his stores with goods made in China rather than American-made products. That's outsourcing." ODA notes this response in the rating, along with its assessment that the retail/manufacturing distinction is material to the "outsourcing" characterization.

Correction reach: Two fact-checks, combined estimated audience 48,000. Original ad estimated reach: 1.88 million. Ratio: approximately 2.6 percent.

Claim G-3 — "Whitfield Said Immigrants 'Don't Belong Here'"

Origin: This claim did not originate with the Garza campaign itself. It originated with a Twitter/X account that presented itself as a Garza supporter and posted a video clip edited to remove the qualifying language before "don't belong here." The edited clip received approximately 8,200 shares and 2.3 million impressions before the original unedited clip was widely circulated.

What Whitfield actually said: At a rural county campaign event, Whitfield's full statement was: "Look, my family came here the right way, and I respect everyone who does it the right way. But illegal immigrants who are here breaking the law don't belong here — they need to go back and get in line like everyone else did." The edited clip retained only "don't belong here — they need to go back."

Attribution complexity: This claim presents a tracking challenge: the Garza campaign did not make this claim. The edited clip circulated among Garza supporters and was not disavowed by the campaign for approximately 36 hours after it went viral. The Garza campaign eventually tweeted the full quote with a statement: "Whatever you think about immigration, voters deserve the full context of what candidates say."

ODA rating: A1-I1, attributed to "social media accounts associated with Garza supporters" (not the Garza campaign itself), with a notation that the campaign eventually published the corrective context.

Lessons for the tracker: This case illustrates why attribution precision matters. "Garza campaign disinformation" and "disinformation circulating in Garza's support base with initial campaign inaction" are meaningfully different characterizations.

Claim W-2 — "Garza's Healthcare Plan Would Eliminate Private Insurance"

Origin: Whitfield campaign press release and associated social media content. Repeated in a Whitfield stump speech that received significant state media coverage.

ODA rating: A1-I2. The claim is documented false — Garza's published healthcare plan explicitly preserves private insurance — but its impact is classified I2 rather than I1 because it appeared in earned media (press release, speech) rather than paid advertising, and its documented spread is more limited.

Complication: Several progressive advocacy groups have criticized Garza's plan precisely because it does not eliminate private insurance. A press release from one such group ran the headline "Garza's Half-Measure: Why a Public Option Isn't Enough." The Whitfield campaign cited this as evidence that "even Garza's own side knows this plan changes everything." This is a sophisticated rhetorical move: using progressive criticism of Garza's plan as evidence for a right-wing characterization of the plan. The tracker notes this context explicitly in the rating explanation.

Section 5: Text Analysis for Automated Detection

Using VADER and Topic Modeling to Flag Potential Misinformation

The claim identification process described in Section 3 relies on human judgment for final decisions, but automated tools can dramatically expand the volume of content that receives initial screening. Sam's automated detection pipeline uses two primary natural language processing approaches: VADER sentiment analysis and Latent Dirichlet Allocation (LDA) topic modeling.

VADER for Sentiment-Based Flagging: VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-based sentiment analysis tool specifically designed for social media content. In Sam's pipeline, VADER is used not to classify claims as true or false — it cannot do that — but to identify content that exhibits patterns associated with misinformation: extreme sentiment (very high compound scores in either direction), high emotional activation combined with specific named entities (Garza or Whitfield), and claim-like grammatical structures (subject + "said" + quoted content, or subject + verb + statistical figure).

The VADER-based flagging uses a compound score threshold of |0.6| combined with the presence of at least one named entity from the watchlist. This threshold generates approximately 200-300 potential flags per day across ODA's monitored content. Of these, roughly 15-20 percent will be forwarded to Sam's human review queue after additional automated filtering.

LDA Topic Modeling: LDA identifies latent topics in a corpus by finding patterns of co-occurring words. Sam uses LDA on the accumulated corpus of ODA media data to identify topic clusters associated with tracked claims. Each claim in the active tracker is represented as a "seed" topic — a set of keywords characterizing the claim's semantic content. Incoming content is analyzed against these seed topics, and content with high similarity to active claim topics is flagged for review.

Topic modeling also serves a discovery function: it identifies emerging topic clusters that may not yet be associated with a tracked claim but show the semantic signature of a developing narrative. This "early warning" function helped ODA identify the "outsourcing" narrative about Whitfield several days before the Garza campaign released its television advertisement, giving the tracker time to prepare the verification process in advance.

Training a Classifier on Labeled Misinformation Examples

Beyond VADER and LDA, Sam has built a supervised classifier trained on labeled examples from ODA's claim database. This is a more powerful tool, but it requires training data — a labeled corpus of content that has been previously rated by human reviewers.

Sam uses the ODA media dataset's factcheck_rating column as a source of labeled examples. Articles with factcheck_rating of "false" or "misleading" (as rated by major fact-checkers) provide positive examples for the misinformation class; articles with ratings of "true" or "accurate" provide negative examples. Sam supplements this with ODA's own claim ratings for the current race.

The training corpus consists of approximately 2,400 labeled examples (from oda_media.csv plus historical ODA claim ratings). Sam uses a logistic regression classifier with TF-IDF features as the primary model, with a random forest classifier as a secondary model for ensemble averaging. The choice of logistic regression as the primary model reflects Sam's commitment to explainability: "I need to be able to explain to Adaeze — and to anyone who challenges our ratings — exactly what linguistic features drove the classifier's score for a given piece of content. A black-box neural network gives me better accuracy and worse accountability. In this context, I'll take worse accuracy."

Here is a representative version of Sam's classification pipeline:

import pandas as pd
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score, StratifiedKFold
from sklearn.pipeline import Pipeline
from sklearn.metrics import classification_report, confusion_matrix
from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
import re
import warnings
warnings.filterwarnings('ignore')

# Load ODA media dataset
df = pd.read_csv('oda_media.csv')

# Define binary label: 1 = potential misinformation, 0 = accurate
# Based on factcheck_rating column from oda_media.csv
df['mislabel'] = df['factcheck_rating'].apply(
    lambda x: 1 if str(x).lower() in ['false', 'mostly false', 'misleading', 'four pinocchios', 'three pinocchios'] else 0
)

# Filter to labeled examples only
labeled = df[df['factcheck_rating'].notna()].copy()
print(f"Labeled examples: {len(labeled)}")
print(f"Misinformation positives: {labeled['mislabel'].sum()}")
print(f"Accurate negatives: {(labeled['mislabel'] == 0).sum()}")

# Combine headline and excerpt for feature text
labeled['text_combined'] = labeled['headline'].fillna('') + ' ' + labeled['excerpt'].fillna('')

# VADER sentiment features
analyzer = SentimentIntensityAnalyzer()

def get_vader_features(text):
    scores = analyzer.polarity_scores(str(text))
    return pd.Series({
        'vader_compound': scores['compound'],
        'vader_pos': scores['pos'],
        'vader_neg': scores['neg'],
        'vader_neu': scores['neu'],
        'high_sentiment': 1 if abs(scores['compound']) > 0.6 else 0
    })

# Add entity-based features
WATCHLIST = ['garza', 'whitfield', 'open borders', 'outsourc',
             'medicare', 'crime', 'illegal immigrant', 'don\'t belong']

def entity_features(text):
    text_lower = str(text).lower()
    return pd.Series({
        'contains_watchlist': int(any(term in text_lower for term in WATCHLIST)),
        'watchlist_count': sum(1 for term in WATCHLIST if term in text_lower),
        'contains_statistic': int(bool(re.search(r'\d+\s*percent|\d+%|\$\d+\s*(billion|million)', text_lower))),
        'contains_quote': int('"' in str(text) or "'" in str(text))
    })

vader_features = labeled['text_combined'].apply(get_vader_features)
entity_feat = labeled['text_combined'].apply(entity_features)
labeled = pd.concat([labeled, vader_features, entity_feat], axis=1)

# Feature sets for the model
numeric_features = ['vader_compound', 'vader_pos', 'vader_neg',
                    'high_sentiment', 'contains_watchlist',
                    'watchlist_count', 'contains_statistic', 'contains_quote']

X_text = labeled['text_combined']
X_numeric = labeled[numeric_features].values
y = labeled['mislabel'].values

# TF-IDF pipeline with logistic regression (primary model)
text_pipeline = Pipeline([
    ('tfidf', TfidfVectorizer(
        max_features=5000,
        ngram_range=(1, 2),
        min_df=3,
        stop_words='english',
        sublinear_tf=True
    )),
    ('clf', LogisticRegression(
        C=1.0,
        class_weight='balanced',  # addresses class imbalance
        max_iter=500,
        random_state=42
    ))
])

# Cross-validation
cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
cv_scores = cross_val_score(text_pipeline, X_text, y,
                             cv=cv, scoring='f1', n_jobs=-1)
print(f"\nText-only classifier F1 (5-fold CV): {cv_scores.mean():.3f} ± {cv_scores.std():.3f}")

# Fit and report
text_pipeline.fit(X_text, y)

# Get top predictive features
tfidf = text_pipeline.named_steps['tfidf']
lr = text_pipeline.named_steps['clf']
feature_names = tfidf.get_feature_names_out()
coef = lr.coef_[0]

print("\nTop 15 features predicting MISINFORMATION:")
top_mis_idx = np.argsort(coef)[-15:]
for idx in reversed(top_mis_idx):
    print(f"  {feature_names[idx]}: {coef[idx]:.3f}")

print("\nTop 15 features predicting ACCURATE:")
top_acc_idx = np.argsort(coef)[:15]
for idx in top_acc_idx:
    print(f"  {feature_names[idx]}: {coef[idx]:.3f}")

When Sam runs this pipeline on the ODA media data, the cross-validation F1 score for the misinformation class is approximately 0.71 — meaning the model correctly identifies about 71 percent of misinformation cases while keeping false positives manageable. The most predictive features for the misinformation class include bigrams like "open borders," "crime rate," "record shows," and "contrary claims," along with unigrams like "surge," "crisis," and "eliminate." These are not surprising — they track closely with the claim vocabulary Sam has documented in the tracker.

False Positive Management

A classifier with an F1 score of 0.71 will generate meaningful false positives — content flagged as potential misinformation that is actually accurate. In Sam's pipeline, false positives flow through to human review, where they are caught. But the cost of false positives is not zero: they consume human review time, and in a system under time pressure, there is always risk that some false positives receive insufficient scrutiny before entering the tracker.

Sam's false positive management strategy operates on three levels:

Level 1 — Threshold calibration: The classifier outputs a probability score (0 to 1) rather than a binary flag. Sam has calibrated the threshold for human review referral at 0.55 (rather than the conventional 0.50), slightly reducing the false positive rate at the cost of missing some true positives. The logic: it is worse to flag accurate content as misinformation than to miss some misinformation that will be caught through other pipeline streams.

Level 2 — Source credibility weighting: Flags from high-credibility sources (established newspapers, academic research, official government documents) are automatically downgraded one severity level in the review queue. A high-probability flag from the New York Times website is almost certainly a false positive — it's reporting on misinformation, not propagating it. Source credibility metadata from oda_media.csv (source, source_type) is used for this weighting.

Level 3 — Claim-context matching: Before any classifier flag reaches human review, an additional filter checks whether the content actually makes a claim (rather than reporting on or fact-checking a claim). Content that contains phrases like "fact-check," "claims that," "we examined whether," or "the evidence does not support" is tagged as meta-coverage and routed to a separate lower-priority queue.

Validation Approach

Sam validates the automated system against the ground truth of the manual tracker in two ways:

Retrospective validation: ODA's manually identified claims (already in the tracker database) serve as ground truth. Sam checks what percentage of these manually identified claims the automated system would have flagged, and at what point in their lifecycle. The goal is for the automated system to have flagged at least 70 percent of high-impact claims (A1-I1 and A2-I1 ratings) before they received mainstream media coverage.

Prospective validation: Weekly, Sam logs all high-volume automated flags that did not make it to human review (below the referral threshold). At the end of the campaign, Sam will review these logs against the complete claim record to assess what the automated system missed and whether the misses show a systematic pattern.

Early validation results: The automated system successfully pre-flagged three of the five major claims described in Section 4. It missed Claim G-3 (the edited video clip) because the claim originated in a video format and ODA's text-based pipeline does not yet process video content. It missed the early stages of Claim W-1 because the "open borders" phrase alone wasn't in the initial keyword seed list — it was added after the first instance was manually identified.

Section 6: The Public-Facing Dashboard

What to Show the Public vs. What to Keep Internal

Sam's tracker produces more information than should appear on the public dashboard. The full claim database, including claims under active investigation, expert consultation records, campaign response correspondence, and claims that entered the pipeline but were rejected, contains operationally sensitive information that could undermine the tracker's effectiveness if made fully public.

The distinction ODA draws is between the publication layer (what the public sees) and the operational layer (what ODA uses to do its work).

The publication layer includes: - All claims that have received a final rating, with the full rating, explanation, and source citations - The spread data for each rated claim, updated on the publication cadence (described below) - The correction gap data for each rated claim - The methodology document explaining the rating system and process - A summary dashboard showing aggregate statistics: total claims rated, breakdown by category (A1-A4) and impact (I1-I3), breakdown by attributed source (Garza campaign, Whitfield campaign, allied groups, organic social media) - The tracker's correction record: cases where ODA has revised a rating and why

The operational layer — kept internal — includes: - Claims under active investigation (publishing these before verification is complete would constitute spreading the claim) - Identity of confidential sources - Expert consultation records not approved for public attribution - Campaign response correspondence (the correspondence itself; the substance is incorporated into published ratings) - The automated flag queue and the volume of rejected flags

This distinction matters for reasons beyond operational security. Publishing unverified claims in the context of a misinformation tracker — even as "under investigation" — gives them a kind of semi-official attention that can itself spread the claim. Sam has a rule: "Nothing goes on the public dashboard that hasn't cleared full verification. We are not a rumor tracker."

Dashboard Design for a General Audience

The public dashboard is designed for a reader who is: interested in the race but not a political scientist; willing to click one layer deeper but not three; and coming to the site because they've seen a specific claim and want to know what ODA thinks of it.

The dashboard's primary entry points are:

1. The Claim Search: A search bar that allows users to enter a claim phrase and receive matching rated claims. This is the most-used feature — users arriving after seeing a specific claim want to find ODA's rating for it quickly.

2. The Recent Claims Feed: Rated claims in reverse chronological order, with a brief summary card showing the claim text, attribution, rating (A1-A4 with color coding: red for A1, orange for A2, yellow for A3, green for A4), and impact level.

3. The By-Candidate View: Separate views for claims attributed to the Garza side and claims attributed to the Whitfield side, with the same rating summaries. This is where the symmetry principle is most visible to users.

4. The Spread Monitor: A section showing spread data for the highest-impact claims, with visual indicators of claim reach vs. correction reach. Sam considers this the most important section from a public accountability standpoint — it makes the correction gap visible.

5. The Methodology Page: A detailed explanation of ODA's rating system, process, limitations, and advisory panel composition.

Color coding in the dashboard uses a traffic light system: A1 ratings appear with a red border, A2 with orange, A3 with yellow, A4 with green. Sam initially resisted this design — "Color coding is reductive" — but Adaeze overruled them on user experience grounds. "If someone can't tell in three seconds whether this is bad or not bad, they'll leave. The color gets them in; the explanation gives them the substance."

Updating Cadence: Daily vs. Weekly

The tracker publishes new ratings on a rolling basis as they are completed — typically two to four new ratings per week at the current pace of claim intake and verification. Spread data for existing claims is updated twice weekly (Tuesday and Friday mornings) to capture the most recent dissemination data.

The twice-weekly spread update cadence reflects a deliberate tradeoff: more frequent updates (daily) would be more accurate but would create a publication rhythm that might cause ODA to rush verifications to fill the update. Less frequent updates (weekly) would lose timeliness on fast-moving claims. Twice weekly allows the tracker to catch major developments without being driven by the update schedule.

The tracker includes a clear timestamp on every rating ("Rating issued: [date]. Last updated: [date]") so users can see how recent the information is.

Communication Strategy: Presenting Findings Without Amplifying False Claims

The amplification paradox — the risk that fact-checking itself spreads claims to audiences who would not otherwise have seen them — is real but often overstated. Research suggests that the amplification effect of fact-checking is small compared to the direct reach of the original misinformation, and that the net effect of fact-checking is almost always to reduce belief in false claims. But there are specific scenarios where amplification concerns are legitimate: obscure claims that are genuinely unknown to most people, where a prominent fact-check would be the first exposure many audiences have to the claim itself.

ODA's communication strategy addresses this through two practices:

Minimum reach threshold for full-featured ratings: Claims below ODA's I3 threshold (limited reach, no documented viral spread) are logged in the database but are not featured prominently on the public dashboard. They appear in the comprehensive claim log but do not receive the prominent summary card treatment that drives most public engagement. This prevents the tracker from being the primary vehicle by which obscure claims reach large audiences.

Inoculation framing in rating text: Research on psychological inoculation suggests that presenting false claims in a context that explicitly identifies them as false — and explains the techniques used to mislead — is more effective at reducing belief than simply asserting the claim is false. Sam's rating explanations are written to do exactly this: "This claim uses a statistical technique called baseline manipulation. Here's how it works, and here's why the numbers actually show a different picture." Explaining the manipulation technique rather than just asserting the conclusion helps readers recognize similar techniques in future claims.

Section 7: Ethics and Accountability

Who Judges the Judges? Oversight and Accountability for the Tracker

A misinformation tracker is an accountability institution. But accountability institutions need accountability too. The question "who judges the judges?" is not rhetorical — it is a genuine design requirement.

ODA's accountability structure for the tracker has three layers:

Internal accountability: Adaeze serves as editorial supervisor for all tracker ratings. Sam cannot publish a rating without Adaeze's final review. This is not because Adaeze second-guesses Sam's research — she rarely does — but because the executive director's sign-off creates a clear organizational accountability chain. If a rating is wrong, ODA owns it at the organizational level, not just at the individual journalist level.

External advisory panel: The three-person advisory panel (described in Section 2) reviews all ratings biweekly. Panel members are empowered to formally flag a rating as "disputed" in the public record, even if they cannot override the rating. ODA commits to publishing dissents from the advisory panel — if a panel member believes a rating is incorrect, that disagreement appears in the methodology details tab alongside ODA's response. This transparency about internal disagreement is unusual but intentional: it models the kind of intellectual humility the tracker demands of the campaigns it evaluates.

Public corrections mechanism: ODA's published corrections policy commits to: issuing a correction within 72 hours of discovering a rating error; publishing the correction with the same prominence as the original rating; explaining what was wrong and why; and maintaining a public log of all corrections with dates and descriptions. No correction is ever deleted from the record — only updated. The historical version of every rating remains accessible.

This three-layer structure cannot guarantee that the tracker never makes mistakes. It can only guarantee that when mistakes are made, they are acknowledged promptly and transparently, and that a process exists to catch errors before they're published. That is the honest limit of what an accountability institution can promise.

The Asymmetric Amplification Problem

The amplification paradox (touched on in Section 6) takes a specific and more troubling form in the context of a misinformation tracker: the tracker's own publications may spread claims to new audiences even as they debunk them.

The research on this question is genuinely mixed. Some studies find that fact-checks have a net negative effect on belief — they reduce acceptance of false claims, and the amplification effect (false claim reaches new people through the fact-check) is more than offset by the debunking effect (people who see the fact-check are more likely to correctly reject the claim). Other studies find that under conditions of motivated reasoning — when people strongly want to believe a claim — corrections can produce a "backfire effect" in which exposure to the correction actually strengthens belief in the original claim. More recent research has questioned whether the backfire effect is as robust as originally reported. The current scholarly consensus is that corrections are generally effective but that their effectiveness varies by individual, claim type, and correction presentation.

For ODA's tracker, the asymmetric amplification concern is most acute in two scenarios:

Scenario A — Obscure claim amplification: A false claim is circulating in a narrow partisan ecosystem. Fewer than 50,000 people have seen it. ODA publishes a prominent fact-check reaching 200,000 people. Net effect: more people know about the claim than would have otherwise, and the debunking effect may not be sufficient to offset the amplification. ODA's response is the minimum reach threshold policy described in Section 6 — this scenario is precisely what that policy is designed to prevent.

Scenario B — Motivated reasoning amplification: A false claim is circulating primarily among voters who are strongly committed to the candidate who made the claim. They are the most likely people to encounter the fact-check, and they are the least likely to update their beliefs in response. The fact-check may drive engagement (shares, comments, counter-arguments) that actually increases the claim's visibility. ODA's response here is to focus its rating explanations on the technique (how the misleading was done) rather than the conclusion (this is false), on the grounds that explaining manipulation techniques is more effective than asserting falsity for highly motivated reasoners.

Free Speech Concerns and Editorial Responsibility

The First Amendment implications of a misinformation tracker require careful navigation, particularly as the legal landscape around "misinformation" has become politically charged.

ODA is not a government entity. It has no power to remove content, restrict speech, or impose legal penalties. It is a journalistic organization that publishes its assessments of the accuracy of public political claims. This is constitutionally protected activity — rating the accuracy of political statements is itself protected speech.

But there are subtler concerns. When ODA rates a claim as A1-I1 (Documented False, High Impact), it is issuing a public statement that may affect the campaign's reputation, its fundraising, its earned media coverage, and voter perception. This is not legally problematic — it is exactly what accountability journalism is supposed to do — but it creates editorial obligations that Sam takes seriously.

Obligation 1 — Proportionality: The rating must be proportionate to the actual evidence. Calling something "documented false" when it is better characterized as "misleading context" is both analytically wrong and editorially irresponsible. The graduated A1-A4 scale exists precisely to enforce this proportionality.

Obligation 2 — Due process: Every rated campaign receives advance notice and an opportunity to respond before the rating is published. This is standard journalistic practice. It is not giving the campaign veto power — Sam and Adaeze have final authority. But it ensures that ODA has heard the campaign's perspective before publishing.

Obligation 3 — Consistency: The standards applied to Garza's claims are identical to the standards applied to Whitfield's claims. This cannot be stated often enough, because the accusation that a fact-checker is applying different standards to different sides is the most common and most damaging criticism a fact-checking organization can face.

Obligation 4 — Accuracy about accuracy: ODA does not rate opinions, predictions, or values claims. It does not rate rhetoric. It does not rate characterizations unless those characterizations make verifiable factual assertions. Sam has a folder of draft fact-checks that were never published because, on closer examination, the "claim" was actually an argument — a contestable interpretation of complex evidence — rather than a verifiable factual assertion.

Handling Pushback from Campaigns: Realistic Scenarios

When ODA publishes its rating of Claim W-1 ("open borders"), the Whitfield campaign's response is swift. A campaign spokesperson emails ODA within four hours: "Your so-called 'fact-check' is partisan hackery. You're carrying water for the Garza campaign. Garza's own record shows she opposed the border security bill in 2019. We demand you retract your rating."

Sam and Adaeze's response protocol:

First, they read the email carefully. The Whitfield campaign has raised a factual point — Garza did oppose a specific border security bill in 2019. This is worth checking.

Second, Sam verifies the claim. Garza did, in fact, oppose the 2019 state-level border security bill — which would have increased mandatory minimum sentences for people caught crossing the border illegally. She opposed it because, as she publicly stated at the time, she believed its sentencing provisions were disproportionate, not because she opposed border enforcement per se. Her public statements consistently called for increased resources for border processing capacity, legal immigration pathways, and coordination with federal authorities.

Third, ODA reviews whether this new information changes the rating. It doesn't — the evidence still overwhelmingly contradicts the "open borders" characterization — but Sam adds a paragraph to the rating's methodology section noting the 2019 bill opposition and explaining why it does not support the "open borders" characterization in context.

Fourth, ODA responds to the campaign: "We have reviewed the information you provided. We have updated the methodology section of our rating to address the 2019 border security bill. Our rating remains unchanged. The full evidence is available at the link below." ODA does not respond to "partisan hackery" characterizations.

The Garza campaign, meanwhile, objects to ODA's rating of Claim G-1 on outsourcing: "We have documented evidence that Whitfield's stores sell Chinese products while claiming to support American workers. The outsourcing characterization is accurate." ODA's response follows the same protocol: review the new documentation provided, assess whether it changes the rating, update the methodology notes with the new information, communicate the decision to the campaign.

In both cases, ODA applies the same process. In neither case does campaign pressure change a rating. The only thing that changes ratings is evidence.

Adaeze later reflects on these exchanges in an internal document: "The campaigns will always push back. Some pushback will be legitimate — they'll have evidence we didn't have. Some will be rhetorical pressure. The only way to tell the difference is to have a process rigorous enough that you're confident in it when pressure comes. If you built the process right, you can explain every decision. If you can't explain a decision, that's the signal you need to look again."

Adaeze's Editorial Standards Document

Adaeze drafts a formal editorial standards document for the tracker. It covers:

On neutrality: "We are not neutral between truth and falsehood. We are neutral between political parties and candidates. These are different things. We apply rigorous standards equally to all sides; we do not pretend that the evidence is ambiguous when it is clear."

On speed: "We will not sacrifice accuracy for timeliness. A wrong fact-check published quickly is worse than a correct fact-check published two days later. If we are not ready to publish a rating, we do not publish it."

On corrections: "We will correct our errors promptly, prominently, and without defensiveness. Corrections are not failures — they are demonstrations that our process works. An organization that never corrects is either making no errors (impossible) or not acknowledging them (unacceptable)."

On self-disclosure: "If ODA has any relationship — financial, personal, or organizational — with a person involved in a claim we are rating, that relationship must be disclosed before the rating is issued. If the relationship is disqualifying, a different analyst reviews the claim."

On scope: "We rate factual claims. We do not rate opinions, predictions, values claims, or characterizations that do not contain verifiable factual assertions. When a claim is on the boundary, we err toward not rating rather than stretching our rubric."

This document is published on ODA's website as part of the tracker's methodology materials.

Section 8: Equity Audit of the Tracker

Are Communities of Color's Concerns Tracked with Equal Rigor?

The tracker's equity audit is motivated by a question that Adaeze raises explicitly in the project's planning phase: "We're tracking misinformation. But whose misinformation concerns get tracked? If false claims circulate about immigration and are primarily believed by white voters, we track them. If false claims circulate about voting rights or voter suppression and are primarily of concern to Black and Latino voters, are we tracking those with equal rigor?"

The equity audit examines the tracker's claim database along three dimensions.

Dimension 1 — Issue coverage: Does the tracker's claim database include claims about issues that disproportionately affect communities of color? The audit maps each tracked claim to the policy issue it involves. In the Garza-Whitfield race, the policy areas that disproportionately affect communities of color include immigration enforcement, voting rights, criminal justice policy, and access to healthcare. The audit finds that immigration claims are well-represented in the tracker (reflecting the race's geographic context), but voting rights claims — particularly claims about voting procedures, eligibility, or polling place changes — are underrepresented. The tracker has no current ratings involving voting access, despite ODA's media monitoring identifying at least two circulating claims about voting hour changes and ID requirements that were factually inaccurate.

This is a gap. ODA adds voting rights claims to the tracker's active claim intake list.

Dimension 2 — Source coverage: Are misinformation claims affecting communities of color reaching the tracker through the same pathways as other claims? The audit examines the geographic distribution of the tracker's sources and flags an underrepresentation of Spanish-language media outlets in ODA's monitoring infrastructure. The oda_media.csv dataset includes sources tagged with source_type, and filtering for outlets that serve primarily Latino audiences reveals a significant gap: very few such outlets appear in ODA's media monitoring feeds.

This is also a gap. ODA adds four Spanish-language regional media outlets to its monitoring infrastructure.

Dimension 3 — Impact classification: The tracker's impact classification (I1-I3) is based on estimated reach among the general population. A claim that reaches 200,000 people in a geographically or demographically concentrated community may receive a lower impact rating than a claim with the same reach spread across the general population, even if its effect on the specific community is proportionally much larger. The audit recommends adding a supplementary "community impact" flag for claims whose reach, while not reaching the I1 threshold in aggregate, is concentrated in a community with heightened vulnerability to the specific claim.

This recommendation is partially implemented: ODA adds a "community impact: [community]" notation to claims meeting this criterion, while maintaining the I1-I3 scale for primary classification.

Language Access: Tracking Spanish-Language Misinformation

The state in which the Garza-Whitfield race is occurring has a substantial Spanish-speaking population — approximately 28 percent of adults report Spanish as their primary home language. Yet before the equity audit, ODA's tracker monitored English-language media almost exclusively.

Spanish-language misinformation presents distinct challenges:

Translation asymmetry: False claims sometimes translate differently from English to Spanish. A claim that is patently absurd to an English-speaking, politically informed audience may be plausible in a Spanish-language context where the listener has less prior exposure to the specific candidates or political system.

Source ecosystem differences: The Spanish-language media ecosystem in the state includes a mix of national Spanish-language networks (Telemundo, Univision affiliates), regional Spanish-language newspapers and radio, WhatsApp group networks, and social media influencers with primarily Spanish-speaking audiences. The WhatsApp component is particularly significant — WhatsApp is the dominant messaging platform in many Latino communities, and its end-to-end encryption makes monitoring essentially impossible through standard tools.

Credibility dynamics: Political misinformation sometimes uses different legitimacy anchors in Spanish-language contexts — claims attributed to "the government" or "doctors" or religious authorities may carry different persuasive weight than in English-language media.

After the equity audit, Sam develops a partnership with a Spanish-language political journalism nonprofit that conducts its own fact-checking in Spanish. ODA and this partner organization share claim identification data (with appropriate confidentiality protections) and coordinate on ratings for claims that appear in both English and Spanish ecosystems. Spanish-language ratings appear on ODA's tracker with the partner organization's attribution, with a link to the Spanish-language rating.

The WhatsApp monitoring problem is not solved — it may not be solvable with available tools — and ODA's public methodology statement acknowledges this limitation explicitly.

ODA's Equity Checklist Applied to the Tracker System

Adaeze maintains an equity checklist that ODA applies to all its major projects. Applied to the misinformation tracker, the checklist asks:

Question 1 — Who benefits and who bears risk?: The tracker's ratings are public information that can affect campaigns' reputations. Are there asymmetric risks — cases where communities of color bear disproportionate reputational risk from the tracker's methodology choices? The audit finds one potential asymmetry: claims attributed to "social media accounts associated with [campaign's] supporters" (as in Claim G-3) may disproportionately track claims originating in communities of color, where organic grassroots content is more likely to be characterized as "supporter activity" rather than campaign activity. ODA addresses this by adding a notation to all such attributions: "This claim originated in social media activity that was not produced or authorized by the candidate's official campaign."

Question 2 — Are data sources equitable?: Addressed in the language access section above.

Question 3 — Does the methodology systematically disadvantage any group?: The minimum reach threshold for full-featured ratings (below which claims are logged but not prominently featured) is reviewed for equity implications. Because Spanish-language media typically reaches smaller aggregate audiences than English-language media, claims circulating only in Spanish-language ecosystems may systematically fail to reach the full-feature threshold even when their community impact is significant. The community impact flag (added after the equity audit) is designed to address this.

Question 4 — Is the tracker's own language accessible?: ODA's English-language rating explanations are written for approximately an eighth-grade reading level. Spanish-language ratings (via the partner organization) are also written at a comparable level. Technical terminology is defined when first used and included in the tracker's glossary.

Question 5 — Does the tracker's governance reflect diversity?: ODA's advisory panel, as constituted, is all-white. Adaeze identifies this as a problem and recruits two additional panel members — a Latino political scientist and a Black journalist — to join the panel for the tracker's second month. This is a direct outcome of the equity audit.

Section 9: Conclusions and Lessons

The ODA misinformation tracker is, at the end of the campaign cycle, an imperfect but valuable public accountability tool. It rated 23 claims in the Garza-Whitfield race before Election Day: 9 attributed to the Whitfield side (6 A1 or A2, 3 A3 or A4), 8 attributed to the Garza side (3 A1 or A2, 5 A3 or A4), and 6 attributed to third-party actors (PACs, unofficial social media networks). The correction gap data — published for all high-impact claims — received significant media coverage and was cited in three news articles and one academic working paper on political misinformation in the state.

What did Sam and ODA learn?

Lesson 1 — Process integrity is the only defense against pressure. Every time the tracker was challenged — by campaigns, by partisan critics, by social media attacks — the only defensible response was to point to the process. "Here is the evidence. Here is how we evaluated it. Here is what we would need to see to change the rating." A tracker built on judgment calls without documented rationale collapses under pressure.

Lesson 2 — The correction gap is the real story. Individual ratings matter less than the systemic evidence that corrections don't reach the same audiences as the original claims. The tracker's most impactful contribution was not any single rating but the aggregate documentation of how misinformation operates in the race's media ecosystem.

Lesson 3 — Equity cannot be retrofitted. The language access gaps and source coverage gaps that the equity audit identified would not have been caught without the audit — and by the time the audit was done, ODA was several weeks into the tracker's operation. The next iteration of the tracker will build equity auditing into the design process, not add it afterward.

Lesson 4 — Automated and human systems need each other. Neither the automated pipeline nor the human review process alone would have produced a tracker that was both comprehensive and accurate. The combination — with automation providing breadth and humans providing judgment — is what made the tracker viable.

Lesson 5 — Humility is a methodology. Every decision about what to rate, how to rate it, what to show the public, and how to handle pushback involved genuine uncertainty. The tracker's credibility came not from projecting certainty it didn't have but from being transparent about what it knew, how it knew it, and what could change its conclusions. Humility, systematically applied, is not weakness — it is the foundation of credible accountability work.

These are not just lessons for misinformation trackers. They are lessons for every form of political analytics that aspires to hold power accountable rather than simply describe it.

Your Assignment

You are joining ODA as a research associate at the beginning of the tracker's second campaign cycle — a competitive gubernatorial race in a state with significant demographic diversity. Sam Harding will be your supervisor. Adaeze Nwosu is the executive director. You have six weeks before Election Day.

Using the framework and tools developed in this capstone, your task is to design and partially build the tracker for this new race. The complete deliverables are specified in the Student Guide. What you should know before you begin:

The goal is not a perfect tracker. The goal is a rigorous, honest, transparent tracker — one that you could defend in public, to the campaigns, to skeptics, and to the communities it serves. Build that, and you'll have built something that matters.

Next: Chapter 44 — Capstone 3: The Campaign Analytics Plan, where you'll move from the accountability side of political analytics to the practitioner side, building a complete analytics plan for a political campaign under the mentorship of Nadia Osei.