NEXUS LEGAL · FORENSIC ENGINE

Deterministic Forensic Audit — architecture in the open.

A technical walk-through of how the engine opens a session, loads its modules, challenges its own conclusions, and seals the output. No black box. No probabilistic shortcuts.

1 · HOW A SESSION OPENS

A typed session header, not a free-form prompt.

Every engagement begins with a structured header that pins the scope before any model sees a document. This is what removes the "Copilot" ambiguity.

SESSION: Unique deterministic ID per case (audit trail anchor).
JURISDICTION: Country + legal family (e.g. ES · Civil Code / UK · Common Law).
BRANCH: Domain: Tax · Corporate · Procedural · Regulatory · Commodities · AML…
MODE: Perspective: Defensive · Offensive · Neutral Due Diligence.
DOCUMENTS: Ordered list of inputs with page count and classification.
TIMESTAMP: ISO 8601 session start, pre-committed before analysis runs.
INITIAL STATE: ANALYSIS READY · DOCUMENTATION PENDING · CONDITIONED ANALYSIS.

2 · MODULES LOADED

Obligatory safeguards + adaptive modules per branch.

Three modules run always. The rest are activated deterministically based on the session header. The model never chooses on its own what to load.

Obligatory (always on)

MODULE-AUDITTRAILACTIVE
Records every step for forensic reconstruction.
Always on
MODULE-OUTPUTACTIVE
Enforces the typed output schema and citations.
Always on
MODULE-SAFEGUARDSACTIVE
Blocks known hallucination patterns at source.
Always on

Adaptive (activated by header)

MODULE-FISCALACTIVE
Tax-law rules, procedural deadlines, doctrine.
BRANCH = Tax
MODULE-LEGALACTIVE
Statutory and case-law cross-check.
Any legal branch
MODULE-PROCESALACTIVE
Administrative and judicial timelines.
Procedural branches
MODULE-DDACTIVE
Due-diligence scoring against supplied documents.
Second document or prior report supplied

3 · ADVERSARIAL FLOW

Layer A → Node A → Node B → Node C

The Presumption of Inaccuracy is not a marketing slogan; it is a pipeline. Every conclusion has to survive an adversarial review before it ever reaches the client.

1. Layer A
Literal source substrate
Extracted text from the input documents, page- and line-anchored. Everything above must be traceable here.
2. Node A
Analyst
Produces the draft opinion with a certainty tag on every claim (L1–L5). Must anchor to Layer A or self-declare the gap.
3. Node B
Validator (adversarial)
Applies the Presumption of Inaccuracy. Issues typed findings (H-NNN) with severity. If any [R] ERROR is raised, the verdict flips to NON-CONFORMING → CORRECT AND RESUBMIT.
4. Node C
Red team (opposing counsel)
Re-examines the surviving opinion as hostile opposing counsel: attacks each conclusion, probes the weakest links and surfaces the counter-arguments a court or adversary would raise. Each node is swappable for any AI provider. Findings that hold are reported back to the client alongside the residual risk — not silenced.

Isolation & Continuation Rule

When Layer A comes back empty (documents with no extractable text or unreadable scans), the affected section is flagged as [L5-V] CRITICAL and isolated. The rest of the analysis continues against whatever is actually readable. The engine never infers content from an empty source.

4 · CERTAINTY SCALE

L1–L5: every statement carries its own grade.

Instead of a free-form summary, the engine tags every claim with a certainty code. The reader knows exactly where a statement stands — and where to push back.

Code	Label	Meaning
[L1]	Literal	Verbatim text of the supplied document (Layer A).
[L2-J]	Case law	Case-law citation verified against the official source.
[L3]	Applicable norm	Generic applicable statute, reference cited.
[L3-NV]	Verified, not literally quoted	Norm verified against official source but not quoted word-for-word in session.
[L4]	Inference	Reasoning step derived from prior [L1]/[L3], not verified independently.
[L5-V]	Documentary void	Critical gap: expected source content is missing or unreadable.
[L5-P]	Preventive deadline	Possible procedural/statutory deadline not yet verified — alert raised.
[L5-C]	Configuration gap	Part of the session header could not be determined — module scope narrowed.

5 · VERDICT PROTOCOL

Three canonical outcomes.

Node B closes the session with one of three typed verdicts. The client receives the verdict first and the reasoning second, not a narrative with a hidden conclusion.

TOTAL CONSENSUS

VERDICT: CONFORMING · RECOMMENDATION: PROCEED

Node A and Node B agree on every finding. Claims are anchored to Layer A within the declared certainty levels. Safe to use.

DISCREPANCY DETECTED

VERDICT: NON-CONFORMING · RECOMMENDATION: CORRECT AND RESUBMIT

Node B raises at least one [R] ERROR finding (for example, unverified claim or contaminated citation). Node A must re-run under stricter instructions before release.

CONDITIONED ANALYSIS

VERDICT: CONDITIONED · RECOMMENDATION: COMPLETE DOCUMENTATION

The substantive opinion cannot advance without additional input (empty Layer A, missing document, configuration gap). The engine returns exactly what is missing.

Finding format (H-NNN)

Every discrepancy raised by Node B follows the same schema. No buried caveats, no prose hedging.

ID: H-[NNN]
Severity: [R] ERROR · [!] DISCREPANCY · [i] OBSERVATION
Description: Short, verifiable against Layer A.
Impact: Critical · Operational · Informative.
Recommendation: CORRECT AND RESUBMIT · CONTINUE · REVIEW.

6 · REFERENCE FRAMEWORKS

Risk, contract and legal-tech frameworks the engine aligns against.

The engine is not self-referential. Each module is built against an external, publicly auditable framework. Compliance officers can map our output to the control they already know.

ISO 31000:2018
Risk management — principles and guidelines.
ISDA Master Agreement
Derivatives & financial contracts.
UNIDROIT Principles
International commercial contracts.
National Security Frameworks (ISO 27001, NIST CSF and equivalents)
Applicable national security baselines by jurisdiction.

7 · INTEGRITY SEAL

Every session closes with a cryptographic integrity seal.

After Node B issues the verdict, the whole artefact is sealed with a keyed HMAC-SHA256 signature. Any post-issuance change — even a single character — breaks the seal and is detectable; the seal can be re-verified on demand by the issuing system. (On the roadmap: eIDAS qualified timestamps via an accredited QTSP.)

Algorithm: HMAC-SHA256 (keyed)
Timestamp format: ISO 8601 with ms (YYYY-MM-DDThh:mm:ss.sssZ)
Hash length: 64 hexadecimal characters
Re-verification: On demand by the issuing system

NEXUS LEGAL

Map this architecture onto your case load.

Book a 30-minute session with our engineering team. Bring one document (or describe it in abstract) and we will walk you through how the engine would open the session, load the modules, and close the verdict.

Request technical demo Back to Legal

Documentation is generated by AI and does not constitute legal advice. Quantum Nexus Ventures FZCO.