§ How It Works The process, step by step Technical walkthrough

How MIRA actually works.

The process runs six steps, every time. Reading the brief is not the work. The work is the process. Here is what happens when you hand a matter to MIRA.

01

Read the brief.

Parse the instruction, extract the legal question, map it to the regime (direct tax, GST, corporate, litigation, regulatory), the forum (which tribunal or court), the timeline, the exposure. Understand what success looks like before doing any work.

Intake parsing • Issue extraction • Forum detection • Stakes modelling
02

Pick the right skills.

Select from the library of 215. Informed by matter type, segment, deployment mode, and the firm playbook that is loaded. A GST matter in a tax boutique picks differently from a GST matter inside a listed manufacturer.

215 skills • 22 groups • Contextual selection • Playbook aware
03

Order them into a plan.

Assemble chosen skills into an execution graph. Dependencies first, parallel branches where possible, approval gates where required. The plan is visible before execution, so senior counsel can sanction, amend, or override.

Execution graph • Dependencies • Parallel branches • Approval gates
04

Reason through each step.

Execute each skill. Cite authorities from the verified corpus. Extract facts from matter files. Compose arguments in firm voice. Capture every interim output alongside its sources. No step runs blind, no claim is unsupported.

Skill execution • Source citation • Fact capture • Firm voice
05

Verify against the meta layer.

The output passes through the meta reasoning layer before release. Six checks: evidence mapping, counter argument construction, argument comparison, cross validation, weakness detection, rebuttal building. A failing check blocks release until the issue is addressed.

Six checks • Independent verification • Contested outputs flagged • Never silently released
06

Produce output with chain.

Deliver the final output alongside its reasoning chain. A senior lawyer can trace every decision, inspect every skill firing, interrogate any branch, drop into any source. The chain is exportable, archivable, and court defendable.

Output • Reasoning chain • Source trail • Audit ready
FIG. 05 6 STEPS REASONING AUDITABLE FIG. 05REASONING CHAIN TRANSPARENT & AUDITABLE 01. Read the brief INTAKE PARSING 02. Pick the right skills FROM 215 IN LIBRARY 03. Order them into a plan EXECUTION GRAPH 04. Reason through each step WITH CITATIONS 05. Verify against meta layer SIX CHECKS GATE 06. Produce output + chain DELIVERED WITH TRACE
§ Technical Notes

What is different about this process.

Note 01

The plan is visible before execution.

Most AI systems hide the plan inside the model. MIRA surfaces it. You can inspect the chosen skills, the ordering, the dependencies, before anything runs. Change any of them before execution. Or let MIRA run and review afterwards.

Note 02

Meta layer is a separate system.

The verification layer is not the same model that produced the output. It is a separate system with different training, different objectives, different failure modes. That separation is what makes the check meaningful.

Note 03

Citations are verified in the corpus.

Every citation in every output is checked against the live 14 million document corpus before release. Citations that cannot be grounded are flagged. No invented cases, no phantom paragraphs, no imaginary statutes.

Note 04

Chain is exportable.

The reasoning chain exports as a structured document you can file, archive, or produce in discovery. Every step, every source, every check. This is what legal AI that can stand up in court looks like.

Start today

Legal AID is live. MIRA is open.

Start a free trial of Ask, Interact and Draft, no credit card required. MIRA early access is open now, request a slot and a specialist will reach out within one working day.