What current tools do well
- Deliver filings and alerts quickly
- Search large document collections
- Summarize individual documents
- Support notes and collaboration
Institutional research workflow
Analyst attention—not information—is the scarce resource.
Filings, alerts, headlines, and internal notes arrive continuously. Analysts still have to decide what deserves review, rebuild history, and defend the conclusion.
Too much informationManual comparisonRepeated searchingFragmented memory
Useful—but incomplete
Search finds documents. Summaries compress them. The analyst still has to decide what changed and why it matters.
The opportunity is not another feed. It is a governed layer between incoming information and analyst judgment.
What remains manual
- Compare against prior language
- Judge review priority
- Check earlier disclosure
- Preserve a defensible evidence trail
Product wedge: rank changed disclosures, show the historical comparison, and keep the source evidence attached.
N = 6 completed interviews
Trust, fit, and prioritization mattered more than another summary.
Paraphrased findings only. Titles, employers, and direct quotations are not inferred.
| Interviewee · perspective | Workflow problem | Trust requirement | Product implication |
|---|---|---|---|
| Matthew G. Surprenant Search, Q&A, usability | Retrieval friction | Answers trace to evidence | Source-linked excerpts |
| Bilal Mustafa Systematic, cross-asset | Noisy prediction target | Claims match evidence | Separate product from alpha |
| Matt DiCenso Fundamental workflow | Manual prioritization | Historical context | Rank changes, show prior text |
| Joseph Marks Innovation framing | Problem too broad | Testable value | Narrow the beachhead |
| Redmond Costello Enterprise implementation | Workflow disruption | Integration and security | Complement existing systems |
| Krishna Valluru Data, governance, analytics | Fragile research memory | Provenance and control | Governed evidence record |
Users neededPrioritization · comparison · visible evidence · relevance · low friction
Buyers and gatekeepers neededProductivity · adoption · integration · security · auditability · governance
The product changedEvidence-first ranking · side-by-side history · institutional memory · analyst accountability
Still unvalidated: measured time savings, repeat use, willingness to pay, buyer approval, and enterprise integration.
The first experiment
Could news and filing language add incremental investment signal?
The original project treated document information as a potential input to return and risk prediction.
News + filingsPoint-in-time documents
Language featuresModel representations
Incremental signal?Locked empirical test
Investment outcomeReturn or downside risk
Claim under test
Added documents improve out-of-sample prediction.
Evidence standard
Locked comparisons, issuer-disjoint samples, and point-in-time inputs.
Decision rule
Do not turn a null or inconclusive result into a product claim.
Disciplined innovation
The market claims were not supported. The evidence-processing capability was.
Negative and null results were preserved rather than tuned away.
Incremental news or 10-K alpha
Adding 10-K information did not improve the existing 8-K and inexpensive-news model.
Better expected-language reconstruction
The prior 10-K alone predicted the next filing’s representation better than augmented models.
Disclosure-change downside risk
The product found evidence-backed changes; the bounded risk test did not establish a reliable signal.
Final research classificationINSUFFICIENT_EVIDENCE
Evidence changed the project
The research did not prove alpha. It revealed a better innovation opportunity.
Hypothesis tested. Result rejected. User problem refined. Solution redesigned. Next experiment defined.
Original ideaPredict returns
→
Empirical testLock the comparison
→
Null resultReject the claim
→
Customer discoveryListen for workflow pain
→
New problemFocus analyst attention
→
Working productEvidence-linked changes
→
Analyst pilotTest workflow value
Evidence changed the project rather than being tuned away.
The MVP
From new filing to review-ready evidence in six steps.
- 1IngestPreserve source and filing time
- 2CompareMatch prior and current sections
- 3ClassifyCategory, direction, materiality
- 4Check earlier disclosure8-Ks and pre-filing headlines
- 5ExplainWhy it may merit review
- 6PreserveExport an auditable record
KHCThe Kraft Heinz Company
High materialityPreviously disclosedConfidence 0.947Debt and refinancing language became more specific.
Added: acceleration, covenant breaches, waiver dependency, default, and access to the Senior Credit Facility.
Why review: financing access and cash-demand riskInspect evidence →
Kraft Heinz · 2023 10-K · Item 1A
A more specific debt warning—but not a wholly new issue.
Selected for evidence diversity and clarity, not later market performance.
Available at filing timePrior 10-K + current 10-K + strictly earlier 8-K and headline evidence
Prior filing
General refinancing risk
“Changes in financial and capital markets … may increase the cost of financing as well as the risks of refinancing maturing debt…”
Current filing
Specific covenant consequences
“The creditors who hold our debt could accelerate amounts due … [and we could be] unable to access our Senior Credit Facility.”
What changedDefault, acceleration, waivers, and facility access were added.
NoveltyPreviously disclosedAn intervening 8-K and earlier headline existed.
Analyst judgmentReconcile covenant headroom and waiver status.
From evidence to a research question
A verified change can become a falsifiable research agenda—after a full-prior check.
RH demonstrates a grounded hypothesis with explicit uncertainty, not a recommendation.
Verified change
Full prior-evidence check
Economic mechanism
Falsifiable hypothesis
Skeptical review
Analyst decision
Evidence coverage is 1.00. The result is a research question requiring analyst and market-data checks.
A second bounded case passed as a hypothesis only. Neither case produced an actionable trade view.
Independent evidence review and automated grounding checks are complete; formal blinded human ratings remain pending.
Open the frozen RH hypothesis · Open the frozen DVN hypothesis
EFX grounding correction
Full-prior retrieval stopped a plausible but unsupported story.
Safe rejection is a successful system outcome.
Missing context found
The selected prior excerpt omitted the already-disclosed $125M conditional top-up.
Roles corrected
$346.7M was a remaining balance; approximately $345M was a cash deposit. The quantities were not substitutes.
Mechanism pruned
Unsupported liquidity, reserve, and bondholder claims were removed.
Safe disposition
No actionable trade view was produced.
Two cases produced grounded research hypotheses. Six were safely rejected or held. None produced an actionable trade view.
Supported versus unsupported
A working product demonstration is not the same as a validated business or investment result.
Four labels keep the claims legible.
- Reliable document processing
- Historical comparison
- Evidence-linked change detection
- Point-in-time integrity
- Reproducible classifications
- Incremental news alpha
- Improved expected-language model
- Causal prediction
- Customer adoption
- Willingness to pay
- Downside-risk signal
- Economic usefulness of risk labels
- Analyst time savings
- Recall and trust
- Repeat use
- Institutional-memory value
Next experiment · 2–4 weeks
Can 3–6 analysts reach better-supported conclusions faster?
Compare the normal workflow with Pure News Intelligence-assisted review on a bounded set of 8-K and 10-K changes.
Time to identify and explain the most important change
Recall, agreement, correction burden, and evidence quality
Repeat use, exports, stored-context usefulness, and preference
Questions for the professor
- Is the customer problem focused enough?
- Which user or buyer is still missing?
- Which pilot metric should be primary?
- What evidence would justify a go decision?
Help test the next version