Inside the SOAR-V Pipeline, Part 5: The Researcher Escalation Queue — When Machines Need Humans

This is the final post in our five-part series on the SOAR-V verification pipeline. We've covered how the Claim Extraction Engine finds every verifiable assertion, how the Citation Resolution Service retrieves source content, how the Content Grounding Analyzer evaluates claims through three stages, and how the Verification Scoring Module calculates transparent trust scores.

Every one of those modules is automated. They run on every report, on every claim, without human intervention. For the vast majority of claims — typically 85-90% — automation produces confident, reliable results.

But not always.

The Researcher Escalation Queue is the module that handles the remaining 10-15%: the claims where automated verification can't make a confident determination and human judgment is needed to resolve the question. It's the safety net that ensures nothing falls through the cracks, and it's the feedback mechanism that makes the entire system smarter over time.

What Is the Researcher Escalation Queue?

The REQ is a managed workflow that routes verification failures to human domain experts for resolution. It doesn't just dump unresolved claims into a queue and wait — it triages failures by type and severity, provides experts with all the context they need to make a fast decision, captures their reasoning in structured form, and feeds every decision back into the platform's self-improvement engine.

The REQ exists because we believe in a principle that too many AI companies ignore: knowing what you don't know is as valuable as knowing what you do know.

An AI system that's 90% accurate and acknowledges the remaining 10% is more trustworthy than an AI system that's 90% accurate and presents everything with 100% confidence. The REQ is how we operationalize that principle.

Why Automated Verification Has Limits

Even a sophisticated three-stage verification pipeline encounters situations where confidence is genuinely uncertain. These fall into predictable categories.

Contradictory sources. Two reputable sources make conflicting claims. One analyst firm says the market is $40B; another says $52B. Both are credible. Both used reasonable methodology. The CGA can verify that the report's claim matches one source, but it can't resolve which source is more correct. That requires domain expertise — understanding the methodological differences, knowing which firm's market definition is more appropriate for this context, recognizing which data is more recent.

Niche domain knowledge. The CGA evaluates whether a source supports a claim based on the text. But sometimes the text requires domain context to interpret correctly. A pharmaceutical report might claim a drug "showed significant efficacy in Phase III trials." The cited source confirms "statistically significant improvement (p < 0.05)." But a domain expert might note that while statistically significant, the clinical significance was marginal — the improvement was real but too small to change treatment guidelines. This distinction between statistical and clinical significance is domain knowledge that the CGA can't apply.

Rapidly evolving information. A claim might cite a source that was accurate when published but has been superseded by new information. An earnings report cited from October 2025 might have been restated in a January 2026 filing. The CGA sees that the claim matches the cited source. A domain expert who follows the company would know the data has been corrected.

Ambiguous fidelity. Sometimes the CGA's fidelity score lands in an ambiguous zone — above the failure threshold but below comfortable confidence. A claim that scores 0.83 fidelity in a general business context (threshold: 0.80) technically passes, but the margin is thin. Is this claim "good enough" or does the slight interpretive drift matter? That judgment call depends on the specific context and downstream use of the report.

Novel or unprecedented claims. For claims about emerging technologies, new regulations, or recent events, there may be genuinely limited source material. The CEE extracts the claim, the CRS retrieves the available sources, but the CGA can only find weak support because comprehensive, authoritative coverage doesn't exist yet. A human expert can assess whether the available evidence is sufficient for the claim's context or whether the claim should be softened.

How the REQ Works

Step 1: Triage and Routing

Not every verification failure gets the same treatment. The REQ applies a decision tree based on the failure type:

Reference failure (source unavailable): Before escalating to a human, the REQ triggers automatic re-research. Our specialist research agents receive a targeted query: find an alternative source that can verify this specific claim. This works for cases where the original source went offline, moved to a new URL, or was behind a temporary paywall. If re-research finds a valid alternative source within three attempts, the claim is re-verified through the CGA without human involvement.

Grounding failure with marginal score (CGA score > 0.50): The source exists and partially supports the claim but not strongly enough. The REQ triggers re-research with a refined query — seeking a source that more specifically addresses the exact assertion. Often, the original research found a general source and a more specific one exists. If found, the claim is re-verified with the better source.

Grounding failure with low score (CGA score < 0.50): The source exists but provides minimal support for the claim. This indicates a likely content-grounding failure — the claim probably doesn't accurately represent its evidence. Escalated to human review immediately.

Contradiction detected (CGA Support score = 1): The cited source directly contradicts the claim. This is always escalated to human review because contradictions require understanding which side is correct.

Three re-research attempts failed: If automated re-research couldn't find a source that verifies the claim after three attempts, the claim is escalated with a note: "Unable to find supporting evidence. Consider removing or softening this claim."

Step 2: Expert Assignment

Escalated claims are routed to human experts based on domain matching. A healthcare claim goes to a healthcare researcher. A financial claim goes to a financial analyst. A legal claim goes to someone with legal expertise.

Each expert receives a complete context package:

• The full claim text and its location in the report

• The original cited source with the relevant excerpt

• The CGA's three-stage evaluation results with scores

• The specific failure mode and the CGA's reasoning

• Any alternative sources found during re-research

• The broader report context (what section the claim appears in, what the report is about)

This context package ensures experts can make fast, informed decisions without having to reconstruct the verification themselves. A well-packaged escalation should take an expert 2-5 minutes to resolve.

Step 3: Resolution Options

Experts have four resolution options, each with a required reasoning field:

Verify: The expert confirms the claim is correct despite the CGA's uncertainty. This happens when the expert has domain knowledge that resolves the ambiguity — they know the source is credible, or they can interpret the fidelity question in context, or they have additional knowledge that supports the claim. The claim's verification status updates to "Human Verified" with the expert's reasoning attached.

Reject: The expert confirms the claim is incorrect. The claim is flagged in the report with the expert's explanation of what's wrong. The report may be revised with a corrected claim, or the claim may be removed and the affected section rewritten.

Edit: The expert provides a corrected version of the claim that accurately represents the available evidence. For example, changing "grew 23%" to "grew approximately 15-20%" to match the source's actual language. The corrected claim replaces the original and is re-verified through the CGA.

Dismiss: The expert determines that the claim isn't significant enough to warrant verification concern — it's a minor detail that doesn't affect the report's conclusions even if slightly imprecise. The claim is marked as "Reviewed — Low Impact" and excluded from the report's verification statistics.

Step 4: Learning Feedback

This is where the REQ becomes more than a safety net — it becomes a training engine.

Every expert decision generates a learning signal that feeds into the platform's Recursive Self-Improvement Engine:

• Verify override (expert approves what CGA flagged): Signals that the CGA threshold may be too strict for this domain or claim type. Over time, accumulated overrides in a specific domain can trigger threshold recalibration.

• Reject confirmation (expert confirms CGA failure): Validates the CGA's calibration. This reinforcement signal strengthens confidence in the pipeline's accuracy.

• Edit (expert provides correction): Creates a training example of what "correct" looks like for this type of claim. These examples improve the Writer Agent's accuracy in future reports.

• Dismiss (expert deprioritizes): Signals that the CEE's priority classification might be too aggressive for this claim type. Over time, accumulated dismissals for a claim type can adjust extraction priority.

Every human decision, aggregated across hundreds of escalations, continuously calibrates the automated pipeline's thresholds, extraction rules, and verification standards. The system doesn't just use humans as a fallback — it learns from them.

How This Exceeds Expectations

No silent failures. In every other AI research tool, if a claim is wrong, it's silently included in the output and the user has to find it themselves. In PromptReports, claims that can't be confidently verified are visibly flagged, not hidden. Users always know which claims have full automated verification, which were verified by human experts, and which remain unresolved.

Expert reasoning is preserved. When a human expert verifies, rejects, or edits a claim, their reasoning is attached to the claim and visible to the user. This means users don't just see "Human Verified" — they see why the human made that determination. This transparency builds even more trust than automated verification because users can evaluate the expert's reasoning.

Continuous improvement is real. Users generating reports in a domain that has had many previous escalations benefit from all the learning those escalations produced. If 50 previous healthcare report escalations taught the system that clinical trial claims need tighter fidelity thresholds, the 51st healthcare report will have better automated verification because of those 50 human decisions. The REQ isn't just handling edge cases — it's systematically teaching the pipeline to have fewer edge cases.

Speed of resolution. Escalated claims don't sit in a queue for days. The context packaging ensures experts can resolve most escalations in under 5 minutes. For standard reports, the entire escalation process — from failure detection to expert resolution — typically adds 15-30 minutes to the report generation timeline, not hours or days.

Real-World Use Cases

Use Case 1: Resolving Conflicting Market Data. A technology market report cites two analyst firms with different market size estimates. The CGA verifies the claim against one source but detects that another source in the corpus contradicts it. The REQ escalates to a technology domain expert who examines both methodologies and determines that the discrepancy is due to different market definitions (one includes managed services, the other doesn't). The expert edits the claim to specify the market definition and notes the discrepancy for the reader. The report is more nuanced and more accurate than either source alone.

Use Case 2: Clinical Significance vs. Statistical Significance. A healthcare report claims a drug "demonstrated superior efficacy" based on a clinical trial publication. The CGA rates fidelity at 0.84 (just above the 0.80 general threshold but below the 0.90 healthcare threshold), triggering escalation. A healthcare expert reviews the trial data and notes that while the improvement was statistically significant (p=0.04), the clinical effect size was modest (3% absolute improvement) and unlikely to change treatment guidelines. The expert edits the claim to "demonstrated statistically significant but clinically modest improvement" — a distinction that matters enormously for a pharmaceutical company making strategic decisions.

Use Case 3: Superseded Financial Data. A financial report cites Q3 2025 earnings data for a company. The CGA verifies the claim against the cited earnings report. But a financial domain expert in the REQ recognizes that the company issued a restatement in January 2026 that revised the Q3 figures. The expert rejects the original claim, provides the updated figures from the restatement, and the claim is corrected. Without domain expertise, this error would have been invisible to the automated pipeline — the cited source was accurate when published but had been superseded.

Use Case 4: Emerging Technology Assessment. A report on quantum computing capabilities includes claims about a startup's recent benchmark results. The CGA finds weak support (Score 3/5) because the only source is the company's own press release, and independent validation doesn't exist yet. The REQ escalates to a technical expert who evaluates the claim in context: the benchmark is plausible given the company's known technology approach, but should be presented as "company-reported" rather than independently verified. The expert edits the claim to include attribution and adds a caveat about independent validation pending.

Use Case 5: Regulatory Interpretation. A fintech compliance report claims that a new regulation "requires real-time transaction monitoring." The CGA detects ambiguous fidelity — the cited regulatory text uses language that could be interpreted as either a requirement or a recommendation. A legal domain expert reviews the actual regulatory language and determines that the requirement applies only to firms above a specific transaction volume threshold. The expert edits the claim to include the threshold qualifier, preventing the fintech company from implementing unnecessary compliance infrastructure based on an overgeneralized interpretation.

The Complete Pipeline

With the REQ, the SOAR-V pipeline is complete:

1. CEE extracts every verifiable claim → creating the verification map

2. CRS retrieves actual source content → bridging citations to evidence

3. CGA evaluates claims through three stages → detecting content-grounding failures

4. VSM calculates transparent, decomposable scores → quantifying trust

5. REQ routes uncertain claims to human experts → ensuring nothing falls through the cracks, and every decision improves the system

Together, these five modules deliver something no other AI research platform offers: reports where every factual claim has been checked, scored, and made transparent — with human expertise as the safety net and the continuous learning engine.

This is what Verified Intelligence means. Not a marketing phrase. A five-module pipeline that runs on every report, on every claim, every time.

Five modules. Every claim. Every report. [Experience Verified Intelligence →](/register)