Designing for Traceability: Every Forecast Links to Evidence

TL;DR — Trust in AI forecasting requires separating text extraction from mathematical prediction. This decoupled architecture allows every forecast driver to point to an exact page in a medical record or pleading, enabling data-driven negotiations.

An adjuster stares at a screen displaying a $2.5 million settlement forecast. The current reserve sits at $800,000. If the adjuster escalates this file to the claims committee, they must explain the $1.7 million delta. The statement that the computer generated the number is not a defense. A naked prediction is a liability. In claims and litigation, an output is only as valuable as the evidence supporting it. If a system cannot point to the exact paragraph in a medical report or the specific plaintiff demand that drove the forecast, the number is useless.

We built Canotera around a single engineering constraint called absolute traceability. When dealing with social inflation, third-party litigation funding, and rising nuclear verdicts, the stakes are too high for black-box architecture. Users must negotiate settlements and allocate defense spend based on facts they can verify. Traceability is not a user interface skin applied after the fact. It requires a specific, decoupled pipeline architecture from the moment a document enters the system.

The Ingestion Pipeline and the Pointer Problem

Claims files are messy. A single case involves thousands of pages of unstructured data. We ingest pleadings, medical records, police reports, and attorney correspondence. These files arrive as poorly scanned PDFs, emails with nested attachments, and faxed medical charts. Our first technical hurdle is reading and structuring this chaos without losing the origin of the extracted information. We use generative AI for this specific task. Generative models excel at parsing language and extracting structured facts from unstructured text. They read the case file and isolate the claim drivers. Extraction is only half the ingestion battle. Every extracted fact carries metadata linking it back to its source. We engineer the ingestion pipeline to generate bounding boxes and text anchors for every structured data point. If the generative model identifies a traumatic brain injury or a specific jurisdiction, it also logs the document ID, the page number, and the exact text snippet. We store these pointers in a relational database alongside the structured facts. This guarantees no piece of data enters the forecasting layer without a verifiable origin. We discard untraceable data.

Processing thousands of pages per claim introduces significant latency and security challenges. We never pipe sensitive personal health information through public API endpoints. Our ingestion architecture runs in isolated, secure environments to protect sensitive records. We trade minor processing latency for strict data locality and compliance. The ingestion phase takes minutes to parse a massive case file. That time is necessary to build a rigorous, secure index of the facts. We monitor these ingestion queues constantly. If a document parsing job stalls or encounters an unrecognizable file format, our telemetry flags it for review. System reliability is a prerequisite for user trust. We cannot afford dropped documents when a single missing medical report alters the entire reserve delta.

Decoupling Reading from Forecasting

A common mistake in applied AI is treating large language models as prediction engines. If you ask a generative model to predict a settlement value based on a text prompt, it guesses. It produces a number based on statistical word proximity rather than mathematical rigor. You cannot calibrate a hallucination. We enforce a strict boundary between reading the file and forecasting the outcome. The generative AI reads. It does not predict.

A separate mathematical machine-learning model generates the actual forecast. This model operates entirely on the structured, traced data produced by the ingestion pipeline. We train these geometric models on large volumes of resolved cases with known outcomes. Because the inputs are discrete mathematical features rather than open-ended text, the outputs are highly calibrated. The model calculates a settlement range, an escalation probability, and a reserve delta. We output a range rather than a single point guess because litigation is inherently probabilistic. A single number implies false certainty. A range provides a realistic baseline for negotiation. The geometric model calculates the weight of each input feature. It knows exactly how much a specific injury code or a change in venue shifts the settlement range. It compares the current matter to comparable resolved cases in the historical training set. This mathematical rigor is possible because we feed the model clean, structured data extracted during the ingestion phase.

Exposing the Audit Trail

This decoupled architecture makes the audit trail possible. When the system outputs a settlement range, it also outputs the specific drivers pushing that number up or down. Because the mathematical model knows which features influenced the prediction, and the ingestion database knows exactly where those features originated in the raw text, we connect the two ends of the pipeline. The claims professional sees the forecast and the drivers. When they click on a driver, the interface retrieves the exact page of the medical record or pleading. The user reads the underlying text and verifies the system logic.

This level of transparency changes how claims organizations operate. Instead of reacting to adverse developments late in the litigation lifecycle, adjusters set realistic reserves on day one. They identify the specific documents indicating a high probability of escalation. They enter settlement negotiations armed with data grounded in comparable resolved cases and verified case facts. The plaintiff attorney presents a demand. The adjuster counters with a settlement range calibrated against hundreds of similar resolved outcomes. The system provides the mathematical baseline. The human retains the context and the authority. Traceable architecture bridges the gap between statistical probability and practical claims management. This shifts the negotiation dynamic from emotional arguments to verifiable historical data.

Building a forecasting platform for litigation means accepting that the user will scrutinize every output. We monitor API latency and system performance, but our primary metric for success is whether an adjuster can defend the reserve delta in a committee meeting. The technology must serve the workflow, and the workflow demands proof.

A forecast is just a hypothesis until you read the source material.