Discovery triage register for review priority

Nothing here is legal advice. A draft is a starting point for a qualified person, not a substitute for one.

Use this workflow to turn an initial document collection into a review queue with stated reasons, known gaps and duplicate groups. It is for litigation lawyers and e-discovery teams who need to decide what reviewers should see first.

Nothing here is legal advice. The register is a working draft for a qualified person to check against the issues, disclosure obligations, privilege process and case protocol.

1. Set the triage scope

Before loading documents into the model you are using, write a one-page triage brief. Do not ask for relevance decisions from a bare folder of files.

Include:

• Matter name or internal reference.
• The questions or pleaded issues that the team has authorised for this stage.
• The relevant date range, people, entities, projects and contract names.
• Known key documents, such as the agreement, termination notice, board minutes or correspondence chain.
• Categories requiring separate handling, including potentially privileged material, personal data, confidential material and non-English documents.
• The labels the review team will use.
• The person who can resolve uncertain classifications.

Use a short, fixed label set. For an initial pass, use:

Label	Use it when	Review priority
likely relevant	The document appears to address an issue, key person, event or agreed date range	High
possibly relevant	It has a credible connection, but the connection is not yet clear	Medium
background	It explains context but does not appear to prove or disprove an issue	Low
likely irrelevant	No apparent connection to the triage brief	Low
needs escalation	Privilege, confidentiality, language, corruption, missing attachment or another issue prevents safe triage	Hold

2. Prepare the collection and inventory

Create an inventory before asking for classification. Keep the original collection unchanged. Work from copies or from the approved review environment, following the matter's access controls and preservation process.

Give each item a stable document ID. Record, where available:

• File name and file path or source location.
• File type and size.
• Custodian or source system.
• Created, modified, sent and received dates.
• Author, sender, recipient and copied recipients.
• Family relationship, such as email with attachments.
• Hash or other duplicate identifier supplied by your review platform.
• Extracted text status, including whether the file was scanned or unreadable.

Split the collection into manageable batches. Keep a batch manifest showing the document IDs in each batch. Batch size and supported file handling depend on the version and interface you use. Check the current xAI documentation overview before processing confidential matter material.

For scanned PDFs and image files, note whether text extraction is present. A document with no extracted text is not evidence of irrelevance.

3. Give the model a constrained triage task

Provide the triage brief first. Then provide one batch with its document IDs and available metadata. Ask for a row for every document, including items that cannot be read.

Use this instruction as a starting point:

You are assisting with an initial discovery triage. This is not a legal conclusion.

Use only the triage brief and the documents supplied. Do not infer facts that are not stated. For every document ID, return one register row with:
- document_id
- proposed_label
- priority
- issue_or_topic
- date_or_date_range
- people_or_entities
- concise_reason, quoting or identifying the relevant text where possible
- key_terms
- duplicate_or_near_duplicate_group
- missing_attachment_or_family_note
- extraction_or_readability_issue
- escalation_reason
- confidence: high, medium or low

If a document may contain legal advice, a request for legal advice, settlement material, sensitive personal data, or another restricted category identified in the brief, use needs escalation. Do not decide privilege or disclosure status.

Flag contradictions, missing attachments, referenced documents not present, and near duplicates. Do not mark a document likely irrelevant solely because text extraction is poor.

Ask for a structured table or CSV-style output, not a narrative summary. Preserve the original document ID exactly. If a document contains more than one topic, allow more than one topic in the issue_or_topic field, but assign one proposed label for the initial queue.

4. Build the triage register

Combine batch outputs into one register. Add fields that the review team, rather than the model, controls:

Field	Owner	Purpose
document_id	Collection owner	Links the row to the source item
proposed_label	Triage output	Gives the initial group
reviewer_decision	Review lawyer	Records the checked outcome
priority	Review lead	Sets queue order
reason	Triage output, checked by reviewer	Explains why the item is in the queue
duplicate_group	Review platform or analyst	Groups exact and near duplicates
gap_flag	Analyst	Records missing or referred-to material
escalation_owner	Review lead	Names who resolves the hold
status	Review team	Tracks unreviewed, reviewed or resolved

Sort the working queue in this order:

1. needs escalation items with potential restricted material or unreadable content.
2. likely relevant items tied to key dates, actors and events.
3. Documents in families with missing attachments or missing referenced material.
4. possibly relevant items with medium or low confidence.
5. Background and likely irrelevant material, sampled for quality control.

Keep duplicates in the register even when the review platform suppresses them from the queue. State the representative document ID and the size of each group. A duplicate group can matter when one copy has different metadata, a different custodian or a complete attachment family.

5. Check for wrong output before handover

Review a sample from every label, not only the high-priority group. Start with documents the output called likely irrelevant, low-confidence documents, files with weak text extraction and the largest duplicate groups.

The output is likely wrong if you see any of these patterns:

What you see	Likely cause	What to do
Key names or dates appear in low-priority rows	The triage brief was incomplete or terms were missed	Update the brief and rerun the affected batch
An email is relevant but its attachment is not listed	Family links or attachment extraction failed	Rebuild the family list and escalate the missing item
Many unrelated files sit in one duplicate group	Near-duplicate matching was too broad	Check the representative files and split the group
Scanned documents are mostly labelled irrelevant	Text was unavailable or poor	Route them for OCR or manual inspection
Reasons make claims not found in the file	The output inferred facts	Mark the rows for re-triage using the source text only

6. Hand over the register and exceptions list

Deliver three items to the review lead:

• The triage register, with filters enabled and the batch manifest retained.
• An exceptions list for needs escalation, unreadable files, missing families, unknown dates and unresolved duplicate groups.
• A short handover note stating the triage brief version, batches processed, counts by label, sampling completed and open questions.

Do not describe a triage label as a final relevance, privilege or disclosure decision. The qualified reviewer records that decision in reviewer_decision and updates the reason where needed.

When the workflow does not work

Stop expanding the collection if document IDs do not reconcile, source files cannot be traced, or the output mixes up document families. Return to the inventory and repair the manifest first.

If classifications are broad or inconsistent, narrow the triage brief. Add the disputed issue, names, event dates and examples of documents that belong in each label. Re-run only the affected batch, preserve the earlier register version and compare changed rows. If a restricted-material question arises, place the item in needs escalation and send it through the matter's approved process rather than seeking a final decision from the draft.