IPAK-EDU DEVELOPS SEARCH ENGINE THAT WILL CHANGE THE WAY THE WORLD SEARCHES THE WEB
We built the search engine we needed 20 years ago. No more narrative. See both sides. It begins now.
Every search engine you have ever used is built on a hidden assumption: that you already know what to ask.
You type a question. The machine races through an index of the web and returns the documents that best match your words—ranked increasingly by what other people clicked, what the system expects you to prefer, and what someone paid to place in front of you.
It is a magnificent piece of engineering. It is also structurally incapable of telling you the one thing you may most need to know: what did you not think to ask?
That is not a bug in conventional search. It is what conventional search is. A query-response system can return only a function of the query it receives. If your question carries a premise, the answer inherits it. If your question is one-sided, the results tend to follow it. That leads to a restricted view on a topic. If thousands of people have already searched with the same premise, the ranking machinery may reinforce it before you ever see the page.
We built the correciton..
DeepViathan—pronounced /deep-VY-uh-thun/, from deep and leviathan—is now live on the IPAK-EDU platform ipakedu.online. It is a research engine built for science questions that require more than a ranked list. It searches beyond the wording of your question, surfaces related questions you did not think to ask, verifies the citations it retrieves, and—when the evidence genuinely conflicts—builds the strongest evidence-backed case - and this is important - on each side.
A Search Engine that Steelmans Deeply By Default
We can barely call each search as “search”. A “sounding” is the right word for what it does. A sounding is what you perform when you cannot see the bottom: you lower a line at a particular point and measure the actual depth.
Google casts a net across the whole ocean. Deepviathan dives where you are focused.
You can run a sounding immediately or assign a question to your research frontier and have Deepviathan work it overnight. Each sounding uses one credit, whether it runs on demand or as a scheduled frontier briefing.
IN ONE SOUNDING, DEEPVIATHAN CAN
• Make genuine evidentiary disagreement visible instead of averaging it away.
• Verify the sources it presents and flag uncertain citation records.
• Find the adjacent questions your original query left out.
• Preserve the result in a permanent, personal research library.
When the evidence genuinely conflicts, you see both cases.
The central design decision behind Deepviathan took the longest to get right.
When the system retrieves the literature on a question and finds that the evidence genuinely conflicts—not merely that an unpopular opinion exists, but that the retrieved sources themselves disagree—Deepviathan does not average the disagreement away. It does not hide it beneath a ranked list. It does not assign one side the status of “the answer” and the other the status of an objection.
Instead, it runs the question independently in two directions:
FOR
The strongest evidence-backed cases supporting the proposition.
AGAINST
The strongest evidence-backed cases opposing the proposition.
The two cases appear side by side. Neither column is permitted to undercut the other. Neither is labeled “consensus,” “fringe,” “approved,” or “discredited.” Each receives an independent, good-faith, fully resourced analysis.
Deepviathan presents the evidence. You retain the judgment.
This does not mean that Deepviathan manufactures two equal sides for every question. They already exist. The two-column treatment is activated only when the evidence retrieved by the system shows genuine disagreement. A controversy is not created merely because someone disputes a finding, and a question is not declared settled merely because one position is socially dominant.
Below the two evidence columns is a panel titled “Things You Might Want to Know.” There, each side identifies—in its own voice—the strongest question its own case has not resolved. The case for names what the case for still leaves open. The case against does the same. Neither side is allowed to turn the exercise into an attack on the other.
The bug that taught us what we were building
Early in development, I ran Deepviathan on my own name. What came back was not a balanced research result. It was a one-sided credibility audit. The analysis informed me that my work had not been independently replicated by groups with no stake in the outcome, then began looking for a credibility baseline against which to judge it.
Nobody had explicitly written that instruction. The asymmetry emerged on its own: The bias exists in the corpus. The challenge system had absorbed a default epistemology from the material on which it depended: the consensus position was treated as the baseline, while the dissenting position was treated as a claim on trial, carrying the burden of proof.
It was the right test. The replacement is the system described above. When the retrieved evidence genuinely conflicts, each side receives its own independent analysis. Neither is assigned the burden by default. Contention is detected from disagreement within the evidence—not from whether a position is mainstream, heterodox, popular, or institutionally favored.
A question is not settled because the consensus says so. It is not contested merely because a dissenter exists.
Discovering that failure inside my own build was one of the most useful things that happened during the project. It forced us to define, precisely, what fairness in an evidence engine should mean.
Every citation must pass through a gate
There is an obvious objection to any AI-assisted research tool, and it is the correct objection: these systems can fabricate citations. They can fabricate them confidently, in the correct format, with plausible authors, titles, journals, dates, and identifiers.
Deepviathan therefore places every source it surfaces through a citation-verification gate before presenting it to the user. After searching for three confirming sources of citation details, each citation receives a visible status: Verified, Plausible, Unverified, or Unconfirmed.
Those grades describe the reliability of the citation record—not the truth of the scientific claim. Deepviathan does not tell you that a finding is correct because a paper exists. It tells you whether the paper it presented can be found and whether the cited source appears to support the representation made about it.
A verification gate would be worthless if it merely displayed its own confidence. We therefore tested it by deliberately planting a fabricated citation in the pipeline and requiring the gate to detect it. It did. The citation was flagged as a hallucination risk.
The gate is deliberately tuned to over-flag rather than under-flag. A false alarm may cost the user thirty seconds of checking. A fabricated citation that passes unnoticed can cost the user credibility, publication integrity, and trust. That is the correct direction in which to make the error.
What a sounding gives you
Each sounding gives you an evidence table. When a question is contested, Deepviathan produces a compact table showing the types of evidence found on each side—meta-analyses, randomized trials, prospective and retrospective cohorts, case-control and cross-sectional studies, case series, mechanistic research, case reports, and opinion. It also shows how many sources support the proposition, oppose it, or produce indeterminate findings. You can see at a glance whether “the evidence against” consists of five randomized trials or five editorials.
It also gives you Adjacents. Every sounding surfaces a concentric ring of related questions that the original query did not contain. These are the assumptions, mechanisms, comparisons, consequences, and unresolved issues that become visible only after the system begins exploring the subject. In practice, this is often where the greatest value appears: the engine does not merely retrieve material about the question you asked; it helps reveal the research program surrounding it.
You can teach it a document. Upload a paper, preprint, report, manuscript, or other document and ask Deepviathan to sound it. The system extracts the document’s substantive conclusions, tags them, and examines those conclusions against the surrounding literature using the same evidence-sensitive architecture.
Create and store your own research library. Soundings you retain are archived on your personal Deepviathan page. A sequence of individual searches can accumulate into a persistent body of work rather than disappearing into browser history. Storage is free and uncapped.
How access works
Deepviathan is available inside the IPAK-EDU platform. Paid members receive two sounding credits each month as part of their paid membership.
Each sounding uses one credit, whether you run it immediately or schedule it as an overnight frontier briefing.
Members can currently schedule one frontier question per night, provided that a sounding credit is available. Additional sounding-credit bundles begin at $4.99. Every sounding you retain is archived in your personal library, with free and uncapped storage.
Membership also includes the broader IPAK-EDU platform: the course catalog, live webinars, the community, the educational archive, and the expanding IPAK-EDU research toolset.
EXPLORE DEEPVIATHAN (account required)
JOIN IPAK-EDU
I want to be plain about the economics because I am asking people to pay for the service, and they are entitled to understand what they are purchasing. A sounding is not a database lookup. Each sounding commissions live research across the available literature. When a question is contested, the system may perform the retrieval and verification process multiple times: once to detect the disagreement, once to construct the evidence-backed case for, and once to construct the evidence-backed case against.
That work carries a real per-search cost. The credit model allows the people who use the system most heavily to carry the corresponding cost rather than distributing it across members who use it less often.
No paid placement or ranking. The results are not for sale.
The moment a search system sells placement, it stops being solely an instrument and becomes a channel. The entire purpose of building the inverse of conventional search was to create something with no commercial incentive to decide what evidence you should encounter first.
See Deepviathan work live
Dr. Lyons-Weiler will demonstrate Deepviathan live on Monday, July 27, 2026, at 7:00 PM Eastern, as part of the Director’s Science Webinar series: “IPAK-EDU Development Labs: Use the Web Like a PhD.”
I will walk through what four months of development taught us, run Deepviathan against live contested questions, and demonstrate the broader set of research tools we are placing in members’ hands.
The two-column contested view has to be seen working on a real question before its significance fully lands. In prose, it can sound like a feature. In operation, it is a different way of encountering evidence.
REGISTER FOR THE LIVE DEMONSTRATION
Monday, July 27, 2026 • 7:00 PM Eastern
What comes next
The frontier will eventually be able to manage and prioritize a larger queue of research questions, allowing members to direct a continuing program of investigation rather than initiating every search manually. Each completed sounding will continue to draw from the member’s available credit balance.
The document library is also being calibrated so that a conclusion extracted from an uploaded paper can carry a persistent evidence status: Corroborated, Contested, Partially Tested, or Not Yet Tested. That status can remain attached to the claim as it moves through the research library.
At that point, the question is no longer merely: What does the web say about this? It becomes: How well has anyone actually tested it?
That is the instrument I have wanted for twenty years—and the reason I stopped waiting for someone else to build it.
The web taught a generation of people to ask questions and receive answers. The harder skill—the one that separates a researcher from a consumer of research—is recognizing which questions remain unsettled, where the evidence actually conflicts, which assumptions shaped the search, and what neither side has yet explained.
That skill has never been available as a product.
Now it is.
TRY IT OUT FOR YOURSELF (account required)
Deepviathan is an IPAK-EDU Development Labs project. Members support this work through their subscriptions and use of the search engine. Supporters who want to accelerate development can contribute through the IPAKEDU store by signing up for courses, webinars and gift programs. IPAK-EDU LLC is not a tax-exempt organization, and contributions are not tax-deductible.
Sounds intriguing.
Thank you SO much❤️