Investors

Evidence for an AI-shaped world.

This page is the website version of the pitch deck. A clear explanation of the problem, the product, the research behind it, and what the company is building next.

The problem

The major AI systems produce information that is unsupported, outdated, or difficult to verify.

AI is becoming part of how people research, write, decide, and communicate. The problem is not simply finding information. It is knowing which claims are supported and which sources deserve trust.

22-94%hallucination rates across 26 leading models in a 2026 accuracy benchmark. Stanford HAI

98.2% → 64.4%GPT-4o accuracy fell on a new benchmark. Stanford HAI

42% ↓factual citation accuracy fell as research tool calls scaled from 2 to 150. Citation-support study

Where current AI systems still fall shortSource links are attached to every row.

Sources	Problem	What users need	What current systems do	Why the problem remains	How Aletheia improves it
O1 A1 R1 R8	Unsupported factual claims in fluent prose	Know which sentences are factual and whether each is supported.	All four systems can produce fluent answers with uneven support.	Probabilistic generation can outrun retrieved support even as average accuracy improves.	Break answers into atomic claims and verify them independently.
O1 A1 R2	Fabricated or non-resolving references	Do not waste time on sources that do not exist or cannot be opened.	Systems provide citations, but citation quality varies.	Browsing reduces errors; it does not eliminate generated references or link decay.	Validate URLs, identifiers, archives, and source availability.
O2 A3 R1	A real citation does not support the attached claim	Know whether the cited passage actually entails the assertion.	A relevant-looking link can be returned without claim-level support.	Link presence is only a surface signal; support requires passage-level testing.	Compare each claim with the exact source passage.
U1 R1	Model inference is presented as a sourced fact	Separate what a record says from what a system inferred.	Synthesis often blends source statements and model reasoning.	Inference is unavoidable in synthesis; the boundary must be labeled.	Label fact, inference, allegation, and opinion separately.
U4 U14	A secondary source replaces the primary record	Reach the originating paper, filing, law, dataset, or statement.	Search ranking optimizes access and relevance, not always authority.	The open web makes derivative material easier to find than the original.	Resolve claims toward primary records and show the authority path.
U6 U8	Promotional material is treated as independent authority	Know who benefits from the claim and whether independent support exists.	Systems can retrieve promotional, self-interested, or SEO-driven material.	Availability and ranking do not equal independence.	Record source type, ownership, incentives, and independence.
U9 R7	Credible sources disagree but the answer hides the conflict	See competing accounts, definitions, dates, and authority levels.	Systems often converge on one clean answer.	Many disagreements are real questions of evidence and governance.	Surface conflicts and preserve an unresolved state when needed.
R6 R8	Only obvious claims are reviewed	Receive a complete inventory before verification begins.	Long-form answers contain nested and implied claims that are easy to miss.	Compound claims require decomposition and coverage accounting.	Extract atomic claims and show review coverage.
U1	Similar people, companies, places, or products are conflated	Confirm the exact person, organization, product, or jurisdiction.	Systems can attach a true fact to the wrong entity.	Real-world identifiers are ambiguous and context-dependent.	Resolve entities before attaching facts.
O2 A1 R5	Stale facts are returned without an effective-date warning	Know whether a fact is current for the requested date.	Retrieval can find current material without preserving prior truth state.	Facts can change after training and after a web retrieval.	Record effective date, last-checked date, volatility, and recheck policy.
R6 U6	No search result is mistaken for proof that something did not happen	Distinguish no accessible record from proof of a negative.	Search results are often presented without a clear coverage boundary.	No search system can prove exhaustiveness over inaccessible records.	Disclose search scope, access limits, and negative-search rules.
R7 U9	Confidence is poorly calibrated	Know what is known, uncertain, disputed, or unresolvable.	Systems may answer confidently when evidence is insufficient.	Refusal and uncertainty calibration remain separate from raw accuracy.	Attach claim-level uncertainty and insufficient-basis states.
U9 U12	Answers contradict earlier answers or other systems	Identify and reconcile material contradictions.	Different runs, routes, and models can produce different answers.	Non-determinism, changing context, and model updates persist.	Detect contradictions and link them to the same claim.
O6 P4 G4 R4	Long documents are only partly grounded	Verify every relevant section rather than trust a broad summary.	Large context windows can still miss material in the middle.	More context does not guarantee robust use of every passage.	Map claims to pages, passages, and completeness checks.
R4 U10	Conflicts across multiple documents are missed	Find inconsistent dates, figures, names, and representations.	Systems summarize files without always normalizing them against each other.	Cross-document reconciliation requires persistent comparable claims.	Maintain a cross-document claim register and inconsistency scan.
A1	Dates, arithmetic, units, or calculations are improvised incorrectly	Route deterministic questions to deterministic tools.	Systems can reason through calculations instead of validating inputs.	Tool routing and incorrect inputs remain failure points.	Use recorded calculators and parsers for deterministic facts.
P4 G4	Tables, charts, screenshots, scanned PDFs, or video scenes are misread	Verify factual content in non-text assets.	OCR, layout, image quality, and scene interpretation remain uneven.	Multimodal extraction is improving but is not uniformly reliable.	Track asset regions, page references, and timestamps.
U6 U14	Paywalled, blocked, dynamic, proprietary, or unindexed records are absent	Know what could not be accessed and add authorized material.	Systems disclose some access limits but cannot see every record.	Licensing, authentication, robots, and indexing boundaries persist.	Report access gaps and support authorized uploads or archives.
R2	A cited page changes or disappears after the answer	Reproduce what the source showed when the claim was reviewed.	Links point to a mutable live web.	Link validity decays and pages can change without notice.	Preserve permitted snapshots, hashes, and retrieval dates.
R3 R6 U14	Low-resource languages, local records, and non-digital archives receive uneven treatment	Avoid blind spots outside well-indexed English material.	Coverage varies by language, geography, and digitization.	Availability, OCR, language resources, and local indexing remain unequal.	Record coverage and expand the source base deliberately.
U5	The same prompt produces different answers after reruns or updates	Reproduce the basis of a business decision.	Outputs can change with routing, model versions, and context.	Stochastic generation and silent updates are product features.	Freeze inputs, method version, model/tool metadata, and source set.
O3 R6 U5	Users cannot tell what was searched or whether the search was exhaustive	Know corpus boundaries, exclusions, dates, and stopping conditions.	Research reports describe activity without proving exhaustiveness.	Exhaustive web search is generally impossible and must be disclosed.	Provide a search-scope manifest and stopping rationale.
R1 R3 U5	More sources and tool calls can make synthesis less accurate	Control breadth, depth, and information overload.	Deep research can trade speed and volume for synthesis quality.	Retrieval volume and accuracy have a persistent tradeoff.	Use compact curated retrieval and conflict-aware stopping.
A3 U2 U5	Reliable results require elaborate prompts and manual workflows	Get a repeatable professional verification workflow.	Interfaces improve, but reliable use still depends on user technique.	Organization-specific rules still need explicit workflow design.	Offer fixed, repeatable verification workflows.
A2 P2 G3	Deep research is slow, quota-heavy, and unpredictable	Resolve known facts efficiently before open-ended research.	Deeper retrieval and longer synthesis consume time and compute.	That tradeoff persists even as hardware improves.	Reuse verified facts and escalate only unresolved or stale claims.
O1 A1 R8	The same model generates and then self-checks the answer	Obtain verification independent from the originating assistant.	Self-review remains inside the same model and context.	A system cannot create institutional independence from itself by adding a prompt.	Provide a platform-neutral verification layer.
O5 A4 P3	Each assistant’s memory is vendor-scoped and not authoritative	Carry verified facts between ChatGPT, Claude, Perplexity, Gemini, and Aletheia.	Project memory stays inside each platform.	Commercial platforms maintain separate accounts, permissions, and histories.	Maintain a portable, cross-platform fact record.
U7 U12	No documented native record preserves status, support, uncertainty, and last review	Reuse verified work instead of starting research over.	Conversation history is prose, not a normalized verification record.	Reusable verification requires persistent structured state.	Maintain versioned records with support links and status.
O4 P2 G3	Scheduled search lacks prior claim state and semantic change rules	Know when a previously reviewed claim becomes stale or disputed.	Systems can schedule a search without defining factual change.	Scheduling solves when to run; it does not define what changed.	Monitor claim state, materiality, and before/after history.
U7	No formal correction and revision workflow	Correct a conclusion without losing its original basis.	Systems can generate a new answer without preserving the revision record.	Governed revisions need identity, permissions, reasons, and history.	Record corrections, supersession, reviewer, and reason.
O3 U5 U7	No universal claim-source-passage-time-method audit record	Hand a client, editor, lawyer, or executive an inspectable verification file.	Research histories show activity but not one normalized audit record.	Accountability is a recordkeeping and governance function.	Produce exportable reports and audit history.
U7	An assistant cannot accept professional responsibility for a consequential conclusion	Know when human review is required and who approved the result.	Systems can assist but cannot assume institutional responsibility.	Legal and organizational responsibility remains with people.	Use risk-based escalation and named approval where appropriate.
R9	Web pages and connected content can manipulate agent behavior	Prevent untrusted retrieved text from changing the workflow.	Connected content is data, but it can also contain adversarial instructions.	Attackers adapt as defenses improve.	Use provenance-aware isolation and injection detection.
A5	Research may move sensitive business material across models and connectors	Control what data is sent, stored, and returned.	Connected workflows depend on permissions and provider policies.	Cross-service retention and access obligations remain organization-specific.	Use minimal inputs, account controls, retention policy, and access logs.
O3 O5 A4 P3	Reports and histories are not normalized across assistants	Move verified work without losing identity or context.	Each vendor optimizes its own formats and account state.	Portability requires a neutral representation of claims and evidence.	Use a neutral export and import structure.
R5 O5 A4 P3	Known facts are researched repeatedly instead of reused	Reduce redundant browsing while preserving freshness.	General memory is not a governed truth store.	Volatile facts still need revalidation; stable facts should not be rebuilt every time.	Use curated retrieval first, then escalate when freshness requires it.

The full source-backed problem record is designed to remain scrollable on the page so investors can go as deep as they want.

The product

Aletheia shows what an AI answer can actually support.

It takes an AI answer apart, tests the claims against source material, and returns a clear record of the claim, the evidence, and what remains uncertain.

01Start with the factual claim.

02Separate a complex answer into individual claims.

03Trace each claim to appropriate source material.

04Distinguish evidence, inference, disagreement, and uncertainty.

05Return the claim, the supporting evidence, and what remains uncertain.

Why now

AI is moving faster than we can verify.

Scale of harm

362

documented AI incidents in 2025, up from 233 in 2024. Stanford HAI ↗

Citation failure

39-77%

factual accuracy across 14 models in cited deep-research reports. Research study ↗

Accuracy collapse

>90% → 14.4%

DeepSeek R1 accuracy fell on the same new benchmark. Stanford HAI ↗

Initial customers & distribution

Start with the people who cannot afford to be wrong.

Initial customers

People and teams who publish, research, communicate, or make decisions using consequential facts.

Early customer groups include journalists and editors, researchers, legal and policy teams, communications teams, and businesses whose public descriptions or recommendations need to remain accurate over time.

Distribution

Founder-led trust first. Product-led expansion next.

Build attention through public fact-checks, clear explanations of the problem, direct outreach, early pilots, social media, and relationships with people who already feel the cost of unreliable information.

Business model

A product people use whenever accuracy matters.

Individual

Verify AI answers.

For people who need a reliable way to verify the information they use.

Teams

Shared workflows

For organizations that need consistent review, reusable records, and a clear basis for published work.

Institutional

Higher-value infrastructure

For organizations that need governed access, monitoring, portability, reporting, and integrations around factual reliability.

What funding unlocks

Build the verification layer AI now needs.

Initial funding supports the product, testing, and operating runway required to build it properly.

Product

Make the verification workflows more sophisticated: reusable records, monitoring, file review, and clear customer-facing outputs.

Testing

Test claim coverage, citation support, source quality, freshness, latency, cost, and verification time against current AI systems.

Company

Support founder runway, an appropriate operating base, core infrastructure, and the first carefully chosen contributors.

Founder & vision

Built by someone who has spent her career finding the record.

Shelby Benkard Rose founded Aletheia after working across journalism, law, research, and the art world—fields united by the discipline of finding the record and getting the facts right.

Direct investor CTA

Interested in building the verification layer for AI?

Start a conversation about Aletheia, the research, and the next stage of the company.

Start a conversation Meet the founder ↗