Technical diligence from the record of failure

Every source you pay for indexes what worked.

A dissertation is the only document in science required to print the table showing the author's own method losing. The committee wants the full experimental record, so the failures go in chapter four and stay there, read by five people. We read chapter four.

Run a claim past us See what comes back Read a full memo

QUERY“Active learning cuts our wet-lab experiment count by 10×.”

Four institutions independently found plain random sampling matched or beat active learning.

01 Georgia Tech — adaptive sampling for active search lost to random selection.

02 Imperial — active-learning acquisition strategies underperformed the baseline.

03 MIT — random sampling, 15.9 mm against 17.39 mm MSE.

04 UT Austin — active learning with similarity cutoff, 0.75× the baseline.

7 theses · 4 institutions · none cited each other · 48 documented failures across 40 theses at six institutions. The comparison appears in every one of these dissertations and in none of the papers they produced. Read the memo →

What we read

69,718 doctoral dissertations from 22 institutions. 22,537 of them read end to end rather than by abstract, because the admissions we want are never in the abstract.

49,915 Documented failures. Canonicalised so the same dead end collides across theses that never met.

33,404 Rejected designs. An approach the author considered, chose against, and gave the reason for. Published nowhere else.

9,328 Baseline defeats. The thesis's own method losing, with the number, inside the success table.

73,688 Unfinished work. What the author said was worth doing and did not do — each naming a person, an institution and a year.

183 Corroborated findings. Agreed across independent committees. This is the part that survives cross-examination.

Who it answers for

Best fit

Deep-tech seed and Series A

An associate has a deck making a technical claim nobody on the investment team can personally evaluate. Today that means calling two academics and hoping one has an opinion.

Biggest prize

Accelerators and venture studios

Every unfinished-work item names a gap and the person who identified it. Cross-reference against who has since closed it and you have founders before they incorporate.

A different buyer

Funders and foundations

852 NSF awards, matched to the dissertations they paid for by the grant numbers printed in those dissertations. Every funder measures outputs by publications. None of them can say what their money ruled out. We can.

Engagements

Available now

Diligence memo

$7,500 per claim

Send the claim, or the deck
Memo in 48 hours, every finding cited to a named thesis
Includes what we could not find, stated plainly
First one is free if you are evaluating us on a live deal
Priced as work, not as queries. It is outside the plan.

Q4 — when the questions repeat

Standing access

$3,000 per month

5 seats, 3,000 queries a month, API included
Allowance resets monthly and does not roll over
$0.40 a query beyond it, never capped
Analyst tier at $400 for one seat and 300 queries

In build

Sourcing feed

By arrangement

Technical gaps, weekly
The person who identified each one
Whether anyone has since closed it

What this does not do

You are the one audience that will check, so here are our own numbers before you ask for them.

22 institutions

Elite-skewed, English-dominant, 2021–2026. Independent means different committees, not a sample of science.

77% agreement

Re-judging our prior-art calls on a different model agreed 77% of the time, and every disagreement found more prior art, never less. We lean optimistic about novelty.

100% agreement

The baseline-defeat findings held at 100% across models on identical evidence. That is the column to lean on.

Not an idea list

We generate startup ideas too. They reproduce across models only 53% of the time, so we do not sell them.

Send us the claim you can't check.

If there is nothing in the record, we will tell you that, and it is worth knowing. If four institutions already tried it, you will have their names by Thursday.

Send a claim