The defensible investigations engine

Investigate every document.Prove every word.

An investigation engine, where every claim traces to its source.

Brought in byLitigation, regulatory and investigations teams.
Why Tidebreak exists

The answer is already in the documents. Getting to it, and proving it, is the hard part.

A serious investigation can bury the answer in tens of thousands of files: contracts, ledgers, whole mailboxes, scans in three languages. Finding it by hand takes weeks of senior time, and the work is hard to retrace once it is done.

General-purpose AI will read it all and answer with total confidence, but it cannot show you where the answer came from, and it will invent a source rather than admit a gap. So someone still has to re-verify every line. Either way, the risk lands on you.

How it works

From a question to a cited finding, without asking you to trust the model.

1

Your documents become evidence

Around twenty file types, from mailboxes to audio and video, are read, transcribed and OCR’d into a searchable, entity-linked corpus. Privileged and sealed material is excluded from retrieval.
PDFreport.pdf
XLSXledger.xlsx
MAILinbox.pst
A/Vcall.mp4
OCRscan.tiff
+673
212 entities
698 files read
AR↔EN aliases
2

A plan you approve before it runs

Ask in plain language and Tidebreak returns a short, costed plan. Nothing runs and nothing is spent until you approve it.
01screen_sanctions$0.18
02ask_userpauses
03trace_ownership$0.52
est. $1.40Approve & run →
3

It gathers evidence, and asks when it should

You watch it run. It pauses to ask when it needs to, and can extend the investigation within your caps — a bounded re-plan that never runs away.
Running5 of 7 · $1.42
✓ screen_sanctions
✓ trace_ownership
↻ Re-plan · cycle 1 · +5 steps
… resolve_nominee_directors
4

A finding you can follow

A confidence-banded finding, entities ranked by evidence, every claim cited. Whatever the evidence could not support was removed on the way.
◆ FindingConfidence: High
Three entities route through a BVI shell held by a serving state official.
Khazarvan Maritime on the OFAC SDN list, rerouted via Al-Mawjan.§ C1§ C2
✕ 1 uncited claim removed before you saw it
See it work

A real Investigate run, start to finish.

Below is a walkthrough over a fictional set of evidence. Every step is the product's actual behaviour, including where it stops to ask you a question, and where it admits what the evidence does not show.

Real vs staged for this replay
REALThe plan, the approval gate, the citations, the confidence band, and the re-plan behaviour.
STAGEDThe documents and entities are fictional, and the run is replayed on a fixed script with no live model calls.
tidebreak.app / engagements / falcon / investigate
GUIDED DEMO
The intelligence layer

More than an answer: the surfaces behind it.

Because the corpus is structured, not just indexed, a run leaves you with surfaces to keep working in, not only a written finding. Each preview below is drawn from the same investigation the demo runs.

Network graph

Trace ownership and control across entities, following the relationships that no keyword search would surface. Here: three traders converging on a BVI shell, and the person behind it.
Al-Mawjan General Trading LLCOMAN
الموجان للتجارة العامةAl Mawjan Gen. Trad.+2
Documents44
SanctionsEU CFSP
RolePayment intermediary
Evidence0.78

Entity explorer

Every person and organisation resolved and deduplicated across aliases and transliterations, with its documents, sanctions status and role attached.
1
Mar 24
4
Nov 24
7
Feb 25
11
May 25
9
Jun 25
Wire instructions / month · peak May 2025

Timeline

Reconstruct the sequence of events across the corpus and see how a pattern escalates, here from one wire instruction a month to eleven.
Khazarvan Maritime Logistics PJSCOFAC SDN0.91
Al-Mawjan General TradingEU CFSP0.78
شركة الموجان للتجارة العامةEU CFSP0.71

Sanctions screening

Screen every entity and alias against the OFAC, EU, UN and HMT consolidated lists as part of the run.
The guarantee

Three things the engine will not let the model do.

No. 01

Make an uncited claim

A finding row with no citation, or one whose citation does not resolve to a retrieved passage, is removed before it reaches you. Not flagged. Removed.
Enforced by the engine
No. 02

Invent a source

A reference to a document that was not actually retrieved is stripped, so you never follow a citation to nowhere.
Enforced by the engine
No. 03

Run or spend without you

The plan waits for human approval, and the approver is recorded. Every run has a hard spending cap the engine will not exceed.
Enforced by the engine

And when the evidence runs out, it says so. Findings carry a confidence band; where the support is thin, the finding states the gap. It will not fill a hole to look complete.

HighCorroborated, cited
MediumSupported, thinner
LowStates the gap
Why not just use a chatbot?

Against a chatbot, and against doing it by hand.

A chatbot is fast but cannot show its work. Doing it by hand can, but not at the scale of a real matter. Line by line:

Conventional AI chatbotA human research assistantTidebreak
Traces every claim to its sourceNo, and it sounds just as sure when it is wrongYes, but slowly and by handYes, enforced, or the claim is removed
Reads the entire corpusLimited by its context windowNot realistically, at scaleYes, every file in the set
Will not invent a sourceHallucinates referencesNoFabricated citations are stripped
Consistent and fully loggedVaries run to runVaries by person and dayEvery step in a hash-chained record
Speed on a large matterInstant but unverifiableWeeks of senior timeHours, inside a set budget
Your data stays under your controlSent to a third-party modelStays in-houseRuns in your environment
Where it is used

One engine, across the matters that turn on documents.

Asset tracing

Follow the money to the person

Trace ownership and payment chains through nominees and shells to the party who actually controls them, with every hop cited.
Breach & incident

Triage a disclosed data set fast

Turn a dumped corpus into a ranked picture of who is named, what is sensitive, and where the exposure sits, in one run.
Sanctions & AML

Screen the whole set for exposure

Check every entity and alias against the OFAC, EU, UN and HMT lists, and surface the evasion typologies behind the matches.
Disputes & litigation

Build the account the disclosure supports

Reconstruct events and reconcile the documents behind a claim, with a defensible chain from each finding to its source.
Internal investigations

Answer the board’s question, provably

Investigate conduct across mail, chat and files, and hand back a finding a committee can act on and stand behind.
Corporate intelligence

Map a network before you engage

Resolve entities across languages and aliases, and map the relationships that no keyword search would surface.
Data security

Where your data lives is your decision.

Investigations run on the most sensitive material there is. So Tidebreak bends to your security posture, not the other way around. Choose where it runs.

Fully cloud-hosted

We run it for you in a managed, EU-region environment. The fastest way to start.

Hybrid

Keep the sensitive corpus inside your perimeter while the workspace runs managed. Split the line where your matter needs it.

On-premise, air-gapped

Deployed entirely inside your own environment with no outbound connection. Availability depends on your hardware and the model you run.
  • Your data is used only for your matter, retained only for the life of the engagement, and deleted when it closes.
  • Never used to train models. Never sold or shared for others to use.
  • Privileged and sealed documents are excluded from retrieval.
  • Every export is signed, so you can prove the record is intact.
Trust

Our product cites its sources. So do we.

Every guarantee on this page is enforced in the product, not promised on a slide. A signed export proves the record behind a finding is intact; whether the conclusion is right stays your judgement.

Contact

Book a demo.

It is a guided walkthrough that we run, on our environment, over synthetic data. Tell us the kind of matter you work on and we will shape the session around it.

We are onboarding a small number of design partners through 2026, and early matters steer the roadmap.

We reply within two working days.
Your details are handled as described in our Privacy notice.