How can you trust these results?
Four steps sit under everything Claira does. They are the base you can rely on, and knowing them lets you explain any result built on top.
The bottom course is the base you can trust and rely on. What stands on it is the information you need from your data.
You are handed a set of relevant documents and a chronology that were produced with AI. Sooner or later someone asks how you got them: a partner, a client, opposing counsel, a judge. “The AI found them” is not an answer you can stand behind.
So the question to put to any AI review tool, ours included, is this one: how can you trust these results over something else?
Our answer is the way a result is built. Everything in Claira rests on four steps, laid in order like bricks. Each one is written down and each one can be checked. Together they are the base you can rely on, and the audit trail for the review.
Brick 1. Define the task.
Before a single document is read, you say what you are looking for, in plain words, the way you’d brief a review team. That description becomes the prompt: your review protocol in writing, the instruction every document is coded against.
What you can show later: the protocol, word for word, and how it changed.
- Judgment calls and bright-line rulesA prompt can hold both. For example: if this lawyer is on the email, code it privileged no matter what.
- Calibrate on a sample firstClaira codes a few documents with you before the full run. You QC the sample, tighten the instruction, and release the full set only when the sample comes back the way you need.
- Version historyEach version of the prompt is kept, with who changed it, and each bulk scan is linked to the version it ran. The history downloads as a PDF or a spreadsheet.
Brick 2. Define the context.
A first-pass reviewer who doesn’t know the matter makes poor calls. The same is true here, so the matter is written down too.
What you can show later: what the AI was told about the matter, and who signed off on it.
- A case brief, drafted from the documentsGive Claira the pleadings and a few key emails. It drafts the case brief, called the Case Context, then asks only what the documents can’t answer.
- You sign off on itYou read it, correct it and approve it, and the approval is recorded under your name, with the date. It downloads as Word or PDF, so your team and the AI are working from the same brief.
- Prompts start from itOnce it is approved, the prompt for a task can be generated from the Case Context, so the issues in dispute you signed off on are the issues the review codes for.
Brick 3. Review every document on its own.
This is the brick the first two exist for. Every document is its own unit. Claira puts the same question to each one in turn: is this document relevant? Yes or no, and why.
What you can show later: for any document, the coding decision, with the reason the prompt asked for.
- One document at a timeEach document is read against the prompt, by itself. Nothing is carried over from the document before it.
- The same review at any scaleA bulk scan runs that review across the whole review set, on one version of the prompt. It decides only what the single-document review decides.
- Coded into your fieldsThe result is written to a field on that document in Nuix Discover, next to your reviewers’ coding. You can search, sort and export it like any other field.
- More than yes or noYou can ask for issue codes, the type of privilege, a confidence score, or anything Claira sees as a weakness in its own call.
- Nothing is silently excludedWhen the prompt asks for a reason, a document coded not relevant gets one too, so the null set can be sampled like the rest. A document that could not be reviewed is counted as failed or skipped, with the cause, so you know what still needs a reviewer’s eyes.
Brick 4. Build outputs only from those reviews.
Then ask for the work product: a chronology, a privilege log, a memo.
What you can show later: for any cited line in the output, the document it came from and the coding it rests on.
- Written from the reviewsAn Insight is written from the results already recorded on the documents you select. Claira reads the field as it stands when you generate, so your reviewers’ corrections count.
- Cited by document IDEach review was made on one document, so a statement in the Insight can point to the one it came from. Claira asks for a citation on each factual claim.
- Citations are checkedBefore the Insight reaches you, Claira checks that each citation names a document the Insight read, and removes any that doesn’t.
- Word, Excel or PDFDownload the Insight in the format you need, with its sources.
What stands on the bricks.
The four bricks are the base. What stands on them is the reason for the review: the information you need from your data, and the confidence that you have it.
What you are holding, before anyone reads a word. The early case assessment counts the case from its metadata: custodians, date ranges, file types, and who writes to whom.
What the documents say. A chronology, a privilege log or a memo, written from the coding on the documents, with its sources.
What to do next. Common jobs written out step by step, such as a privilege log or a chronology, so the process is the same from one matter to the next.
Ask for what you need in plain words. It prepares the Case Context edit, the prompt, the scan or the Insight, and waits for your go-ahead.
Each layer is as trustworthy as the bricks under it.
We built Claira in this order. First it could review one document against one prompt and write down its reasons, and everything since has been laid on top of that.
Eventually agents will run all of this. That is why the order matters. A layer adds speed or structure, and what you can rely on in it is the base under it. An answer that skipped the base would be one you couldn’t trace.
Put this test to any automated workflow, ours included: take one line of the work product and ask to see the document, the coding decision, the reason, the protocol and the context behind it.
What this means for you.
If the process is ever challenged, you have something to defend it with.
You make the calls.
It all sits in your Nuix Discover case: the same case, the same fields and the same permissions, with nothing to migrate.
The judgment stays with you. You still need a person in the loop who knows the matter, and that can be one trained reviewer or a small team. Whoever it is has to be able to explain the process, and the four bricks are what they explain it with.
See the four bricks on one of your own matters.
Bring a matter you know well. We’ll go through the four steps with you. Or sign up now and try it on your next case.