Briefing thesis
Large Language Models are effective at language-heavy legal tasks — summarizing depositions, drafting motions, structuring narrative arguments. They are a poor primary engine for sequential state calculation on commingled ledgers. LIBR tracing requires every row, every balance, and every strategy assumption to be reproducible. That is a code problem, not a prompt problem.
Counsel and forensic teams increasingly ask whether uploaded bank statements can be handed to a general-purpose AI with a tracing instruction. The temptation is understandable: the interface feels conversational, and the output arrives quickly.
The risk is structural. Generative models predict likely tokens; they do not guarantee that a $47 service fee on March 2 was incorporated before the LIBR floor was evaluated on March 3. One omitted row can change the modeled separate-property claim for years downstream.
Why probability breaks sequential tracing
When you ask an AI a simple arithmetic question, it often answers correctly — because the pattern is common in training data. Forensic tracing is not a single arithmetic step. It is a state machine:
- The task: Apply LIBR to a normalized chronology. If the balance drops below a traced floor on a given date, cap the separate-property claim from that point forward under the selected strategy.
- The failure mode: A model may skip a visually small fee, misread a scanned date, or round inconsistently — producing a plausible narrative with non-reproducible math.
The hybrid model: perception vs. reasoning
Exit Protocol is built as a hybrid workflow: use machine perception where format variability is high; use deterministic code where financial state must be exact.
Perception layer
Document intelligence and OCR normalize messy PDF statements into structured rows — dates, amounts, and descriptions reviewers can inspect.
Reasoning layer
Once structured, LIBR tracing runs in deterministic Python logic with a visible strategy — separate from generative text models.
Source provenance
Selected uploads retain identity and hash references tied to the matter record for downstream verification.
Snapshot integrity
Exports include SHA-256 references at generation time so reviewers can confirm the workpaper matches stored inputs.
The reviewability standard
Opposing counsel and the court will ask how a number was derived. A defensible answer points to:
- Which source documents were selected and how they were identified.
- Which LIBR strategy was applied and where it appears on the workpaper face.
- Which transaction caused a balance floor event in the modeled chronology.
- Whether the same inputs reproduce the same outputs on re-run.
A generative chat session cannot reliably supply that chain. A structured workpaper path can — when counsel and a qualified reviewer confirm characterization and final use.
Deployment and data handling
Sensitive financial records should follow firm policy on storage, subprocessors, and access control. Exit Protocol supports controlled deployment patterns — including containerized environments for teams that require local processing boundaries. Data handling choices remain a firm and client decision; the platform documents what is stored, hashed, and exported for reviewer inspection.
Practical division of labor
Use generative tools where language throughput helps. Use deterministic tracing where commingled dollars must reconcile to a visible ledger. For the V1 path, Exit Protocol focuses on the latter: selected financial records → single-claim LIBR workpaper → attorney-reviewable export with source provenance and snapshot integrity references.