AI agents need trustworthy external knowledge. Deep Water transforms hundreds of sources into verified evidence. Reports are generated from evidence, not from raw web pages.
Evidence first · Text last · The frontier model never reads raw sources
Provider input units
4.0M
immutable provider metering
Jobs this month
47
14 Deep, 28 Standard, 5 Light
URLs fetched
1,962
avg 41.7 per job
Provider actual cost
$1.18
centrally recorded
Token spend · last 8 weeks
Recent research
The Question
Confident isn't the same as correct. An AI answer that sounds certain is not the same as one you can verify against its source.
The Problem
Before an agent writes a line of code, it has to understand the world it's writing into. That knowledge is scattered across hundreds of sources — and a model alone cannot reach it reliably.
Documentation
Benchmarks
Incidents
Postmortems
RFCs
A larger model doesn't fix this. The bottleneck isn't reasoning — it's collecting and verifying the evidence the reasoning runs on.
Why Existing Solutions Fall Short
Report-first systems read the web, then write prose in one pass. The text reads well — but the evidence behind it is gone by the time you see it. Three problems follow from that.
A report-first model blends what it read with what it already believed. There is no record of which source supports which claim — so you cannot tell a sourced fact from a fluent guess.
A chat box hides every decision: which sources were read, how deep it searched, what it discarded. You get a finished answer and no way for software to inspect, gate, or reproduce how it was reached.
Pushing hundreds of raw pages through a frontier model on every run is expensive by construction. Run it across an agent fleet and the economics break before the quality does.
The report is the symptom, not the unit. When the report is generated directly from web content, there is nothing underneath it to audit. Deep Water inverts that order.
The Atomic Unit of Truth
Not a citation feature. The building block the entire report is assembled from. Every card is one fact, bound to the exact text that supports it.
Fact
MCP authentication failures often stem from improper token-refresh handling rather than malformed credentials.
Source span
… clients MUST refresh the access token before expiry; failing to handle token refresh is the most common cause of authorization errors in long-running sessions …
Citation
Anthropic Documentation · lines 182–201
Open any claim in the report and see the exact card, span, and source it was built from. Nothing is asserted without a receipt.
Every sentence in the report walks back through a card to a span on a real page. No orphaned statements, no untraceable summaries.
An evidence trail you can hand to a reviewer, an auditor, or a compliance team. Confidence scores and source provenance travel with the output.
When a downstream agent acts on a fact, you can answer why — the card carries the reasoning chain from raw text to verified claim.
Evidence-First Architecture
The report is generated from verified evidence — never directly from raw web pages. The frontier model is the last step in the pipeline, not the first.
Sources
Hundreds to thousands of URLs — docs, papers, filings, forums — searched and fetched.
Evidence Cards
Inexpensive models extract atomic facts, each bound to its exact source span.
Verification
Each card is checked against its span. Unsupported claims are dropped; conflicts are flagged.
Report
The frontier model writes from verified evidence only — one call, fully traceable.
The frontier model never reads raw sources.
It sees a clean, verified evidence set — not hundreds of noisy web pages. That's what makes the output auditable and the cost predictable.
Every claim traces back to a span.
Because text is assembled from evidence rather than generated over it, every sentence in the report walks back to the source it came from.
The Pipeline
A research pipeline built for auditability, traceability, and cost efficiency. The expensive frontier model is called once, and only ever on verified evidence — never on raw web pages.
The query is broken into targeted research tracks so the run stays focused, explainable, and free of wasted searches.
Hundreds of sources are searched, fetched, and reduced to atomic evidence cards — each fact bound to an exact source span for traceability.
The frontier model is called once, and only sees verified evidence cards — never raw pages. That keeps the reasoning layer auditable and the expensive work contained.
Every claim is checked against its source span so unsupported assertions are flagged and every conclusion keeps a clear evidence trail.
98%
Extraction Success
1000+
URLs per Job
99.4%
Verified Citations
4 min
Avg. Turnaround
Benchmarks
A report-first system optimises for readable prose. Deep Water optimises for evidence you can trace. The difference shows up on every dimension that matters.
And here's what that looks like on three real queries, run side by side.
Research query
“Latest approved treatments for triple-negative breast cancer”
Research query
“Compare REST vs GraphQL vs gRPC for high-throughput microservices”
Research query
“Market size and growth drivers for solid-state batteries 2024–2030”
For AI Product Builders
Stop building research infrastructure yourself. Wire one engine into your product and ship the feature — instead of spending two quarters on scope, gather, extract, and verify.
Feed your agents verified, sourced evidence instead of raw search dumps. Memory your agent can cite — and you can audit.
Continuous, structured tracking of competitors, pricing, and positioning — delivered as evidence your product can render or act on.
Outputs ship with a confidence-scored evidence trail, so regulated workflows have something concrete to review and retain.
Process hundreds of primary sources per target in minutes, with conflicting claims surfaced rather than averaged away.
Market sizing, growth drivers, and segment analysis assembled from filings and primary reports — every number traceable to its source.
REST API, webhook callbacks, MCP, and an embedded runtime. One integration, and research becomes a backend you call — not a system you maintain.
Research-Heavy Teams
Beyond AI products, any team whose decisions depend on a defensible evidence trail — not just a readable summary — runs on Deep Water.
Series A–C startups, 5–10 person teams
Running 50–200 deep research queries a month — competitor analysis, market sizing, user sentiment. Currently burning $5K–20K/mo in analyst hours or hitting ChatGPT Pro caps.
You replace
2–5 analyst seats + $200/seat ChatGPT Pro
Small funds, 3–5 analysts, deals on 48-hour cycles
300–800 due-diligence queries per month. Speed is the edge: your analyst processes 20 URLs in 3 hours, Deep Water processes 500 in 4 minutes. Incognito mode keeps deal flow private.
You replace
60% of junior analyst research time
Agencies billing $5K–50K per engagement
Cut per-report research cost by 80–90%. Deep Water produces a sourced first draft — your analysts refine and client-face. Same billable hours, higher margin.
You replace
$500–2,000 research cost per report
Fortune 1000 market intelligence groups
Gartner and Forrester take 3–4 weeks and $100K–500K/year. Deep Water delivers the first-pass competitive landscape in 4 minutes. Use them together, not instead.
You replace
3-week research cycles
Founders shipping research features inside their own apps
Need a deep-research backend you can call from your own product? Forget building your own scope/gather/verify pipeline. REST API, webhook callbacks, bring-your-own keys in minutes.
You replace
3–6 months of infra engineering
Researchers, think tanks, policy analysts
A 40-hour literature review takes 4. Triage 500 sources, find the 20 that matter, then read deeply. Every claim cited so you can walk the bibliography yourself.
You replace
Weeks of manual search & synthesis
The Foundation
Deep Water is the research layer that agents call when they need information they can trust.
API
REST + webhooks
CLI
terminal + scripts
MCP
agent-native
Embedded
in-process runtime
Best for personal use
The fastest way to run deep research from your terminal. Zero install ceremony — npm i -g @unlikeotherai/deepwater, paste your key, done. Great for ad-hoc questions, side projects, research before a purchase, editor integrations, cron-driven briefings.
Best for production & heavy volume
The same engine, exposed as a REST API built for scale. Webhooks instead of polling, per-project billing, bring-your-own-model-keys, higher rate limits, SLA. This is the path if research is a product feature, a backend workflow, or a recurring internal process.
Both paths hit the same engine, the same pricing, the same cost caps. The only difference is how the work enters and how the result comes out.
Pricing
Every provider call is recorded with exact usage, cost, and attribution. Your centrally managed workspace plan owns commercial rating, statements, subscriptions, invoices, credits, and collection.
1 · Metering
Original provider counts, cost, and product attribution remain immutable.
2 · Plan
Default, organization, and team terms are assigned in one control plane.
3 · Statement
One billing record contains rated usage, added value, subscription lines, and totals.
4 · Payment
Checkout, invoices, credits, Portal, and collection stay centralized.
A default, free, at-cost, or custom tariff can be assigned at organization or team level.
Members see their own usage. Team owners and admins can review the same breakdown per user.
Every call is metered once; the central billing authority applies commercial terms and owns the customer record.
Open Billing to see current plan configuration and immutable provider metering. Deep Water never derives a customer total locally.
View plan and meteringCustom monthly and usage terms, higher rate limits, an SLA, and embedded deployment for teams putting research on a critical path.
Talk to usIf you're running non-commercial research at a medical or pharma university, an ecological programme, or any initiative aimed squarely at human benefit, we'll run your jobs at cost during launch. No margin. No upsell. Just infrastructure you couldn't otherwise afford.
Medical and pharma university research, ecological and climate studies, public-health work, and other non-profit programmes with a clear human-benefit thesis. Must be non-commercial.
Reach out with a short description of the programme, the institution, and what you're trying to answer. We review everything individually and assign custom terms — there's no self-serve path for at-cost access.
This is a launch-only programme. If it works we'll bring it back in later cohorts, but we can't commit to it indefinitely. If you need predictable access, apply now.
research@deepwater.live · Individual review · Terms assigned on approval
FAQ
Give your agents and your team an auditable evidence trail instead of plausible text. One engine, exposed as API, CLI, MCP, and an embedded runtime.
Evidence-first architecture · Pay for compute, not seats · MCP-compatible