When software is free,
your knowledge is the moat.
Every competitor will have the same models, the same compute and the same literature. What they will not have is every test your teams have run, every result that did not work, and every decision and why it was made.
- 2.5hrs/day
Time a knowledge worker spends hunting for internal answers, about 30% of the working day.
IDC ↗ - 60%
Search four or more separate data sources to find what they need; nearly a fifth juggle seven or more.
Coveo, 2022, 4,000 employees ↗ - 81%
Have to interrupt a colleague for help when their own search fails.
Slite, 2026, 100+ knowledge workers ↗
The asset is already paid for. It sits in notebooks, instruments, drives and threads, and the work of re-deriving it is a salary line your organization pays every year.
What leaving it latent costs you
Not a software cost. A salary cost, paid every year, in the most expensive headcount you have. Move the sliders to your own numbers.
Every figure above is your input, not our claim. We measure your real repeat rate from your logs before anyone signs anything.
An answer you can’t audit isn’t an answer
Reuse only pays if the reused answer is right. We benchmarked our claim layer against standard retrieval on 150 contested biomedical questions, counting how often a cited record says nothing about the question asked. Each method cites a different number of records, so the rate is per 1,000 citations.
- Omic claim layer0per 1,000from 0 of 1,148 citations
- Standard retrieval7.7per 1,000from 32 of 4,156 citations
- Reading the whole corpus117.3per 1,000from 323 of 2,754 citations
- Sources behind a contested answer7.5claim layervs2.6vector index
- Minority position recalled0.99of contested pairs
150 opposed-predication pairs over SemMedDB, median pool 2,960 statements. Preliminary, and we publish where it loses: on a separate 50-paper fixture a vector index retrieves more raw figures, faster and cheaper.Read the report ↗
Your IP should not be someone else’s training data
Whatever the policy says, unpublished methods and test results are going into consumer chatbots today. The risk is not that people use AI. It is that nobody can say what left the building.
Ungoverned today
- Proprietary data pasted into personal accounts
- No record of what was asked or disclosed
- Consumer terms that may permit training on your input
- Departing staff take the working context with them
- Nothing to hand an auditor or a partner's diligence team
Governed deployment
- Your tenancy, and inputs never train a shared model
- Every query and answer logged and attributable
- Access inherits the identity and permissions you already run
- Knowledge accrues to the organization, not the individual
- An evidence trail that survives audit and diligence
What we can tell your security team
Where data lives
Your cloud or ours. You choose.
Training
Never on your data. Any plan.
Access
SSO and SAML on enterprise.
Provenance
Every claim carries who asserted it, when, and whether it still stands.
Open source
AttestDB, the store under the fabric, is public. Read it before you buy.
Need a specific control for your review? Ask and we will answer it straight.
You own the agent code
We build alongside your teams and you keep everything underneath. No black box, and no renegotiating later for access to something you already paid to build. End the engagement and the capability stays, and keeps running.
How an engagement starts
Thirty minutes, your documents, your questions, and the measured repeat rate from your own logs, whether or not it makes the case for buying.
Proof
Before the form: what we have published, including the parts that did not work.
- technical reportAttest: a claim-native data layer for shared knowledge
- technical reportA survey of 3,974 archaeal biosynthetic gene clusters, and a substrate predictor for the fraction tools can see
- technical reportPRESTO: one model for binding affinity, and where it stops working