The Versance Engine.
Versance is built on a proprietary AI engine designed for public-company work. It does not rely on a single prompt, a generic model response, or a simple document search layer. Before an answer, draft, review, or analysis reaches the user, the engine retrieves the relevant record, verifies the evidence, applies disclosure rules, cites its sources, and checks the quality of the output.
Four systems work together: evidence-first reasoning, a continuously updated disclosure foundation, embedded securities-law review, and production-grade evaluation.
Together, they power every Versance solution from the same infrastructure.
Reasoning
Versance does not answer from the first result it finds. The engine searches the disclosure record through multiple paths, refines the search when evidence is incomplete, and writes only from verified source material.
Multi-path retrieval.
Searches across meaning, keywords, dates, archives, and source types.
Iterative refinement.
Searches again when the first pass is not strong enough.
Verified synthesis.
Builds answers from checked evidence, not model assumptions.
Source attribution.
Attaches the documents, dates, and sections used.
Conservative output.
When the record does not support an answer, Versance does not guess.
Model routing.
Uses different models for different tasks, so the platform is not dependent on a single AI provider.
Disclosure
Versance starts with the public-company record: filings, news releases, and earnings call transcripts across SEDAR+ and EDGAR, continuously ingested and prepared for AI use.
The platform has 8,615+ issuers pre-indexed across Canadian and U.S. markets. For client deployments, that record expands to include the company’s investor presentations, website content, approved Q&A, and selected corporate materials.
Every document is converted into a normalized, time-aware source the engine can retrieve, cite, and compare — giving every Versance solution a reliable disclosure foundation from day one.
Coverage
Versance tracks the core disclosure record across Canadian and U.S. public companies:
-
SEC and SEDAR+ filings
-
News releases
-
Earnings call transcripts, where available
-
Client investor presentations, website content, approved Q&A, and selected corporate materials
This gives Versance both breadth across the public markets and depth inside each client’s own disclosure record.
Processing
Every document is processed before it reaches the AI layer. Versance classifies the document, extracts key content, corrects text where needed, applies dates and metadata, and prepares the material for accurate retrieval and citation.
That means the platform is not searching a loose folder of documents. It is working from a structured disclosure record built for AI.
Architecture
The disclosure pipeline runs separately from the core platform, so Versance can continuously ingest new filings, add new sources, and expand coverage without disrupting product performance.
The result is a disclosure foundation that keeps the platform current, scalable, and consistent across investor-facing AI, research, communications, compliance review, and workflow tools.
Specific data partners, infrastructure choices, and processing methods remain proprietary.
The Compliance Layer
Versance applies securities-law awareness inside every workflow, not as an after-the-fact review of generic AI output.
It is built to recognize the disclosure issues that matter in public-company communications: selective disclosure, forward-looking statements, non-GAAP measures, technical disclosure, disclosure consistency, and unsupported claims.
Canadian and U.S. coverage includes NI 51-102, NI 43-101, NI 52-109, NI 52-112, NP 51-201, Reg FD, Reg S-K, Reg G, S-K 1300, Section 21E, and Rule 10b-5.
The Compliance Layer runs across IR Agent answers, investor email drafts, news releases, social posts, compliance reviews, and internal materials — flagging issues, tying them to the relevant rule or principle, and giving the user a practical path to revision.
Evaluation Discipline
A production-grade evaluation framework that gates every release and continuously monitors the platform against a permanent test harness.
Five-dimensional scoring
Every response is evaluated across five dimensions: factuality, completeness, relevance, tone compliance, and confidence. The system checks whether claims are grounded in source documents, the answer is complete, the scope is controlled, required language is present, and certainty matches the evidence.
Scenario suite
The evaluation harness covers standard production queries, query variants, and high-stress designed-to-fail scenarios. These tests measure how the platform behaves across common questions, difficult edge cases, and prompts built specifically to provoke unsupported answer
Release gating
Every model, prompt, retrieval, or routing change runs through the evaluation suite before production. Changes that improve or maintain performance can ship. Changes that regress are held back.
Continuous monitoring
The framework also evaluates live deployments against real company data, generating company-specific quality reports and surfacing issues as they emerge. Versance currently measures 99% answer appropriateness on its internal evaluation suite.
The full methodology, including scenario structure and scoring examples, lives on the Methodology page. Specific test prompts, scoring rubrics, and scenario sets remain internal.
In regulated communications, "sounds right" isn't good enough.
Five failure modes that disqualify generic AI from public-company work:
Partial retrieval becomes confident error.
A missed paragraph flips the answer. RAG systems that stop at "find top chunks, then summarize" can't meet a public-company standard.
Staleness.
Vector stores decay. Without real-time retrieval and temporal ordering, answers cite superseded facts as if they were current.
No provenance.
Summaries blur sources — which filing, which line, which date. An audit problem before it's a content problem.
One-shot thinking.
If the first retrieval misses, most systems guess. Versance reformulates and searches again, iterating until evidence holds.
No regression controls.
Behavior changes silently between releases. Without a permanent eval harness, the chatbot that worked yesterday gives a different answer today.
Independent benchmarks place finance Q&A failure rates at 81%, legal RAG at 19-65% accuracy, and one in three complex queries returns wrong. The SEC has begun prosecuting AI misstatements. For a public company, those numbers are not acceptable, and the regulatory consequences fall on the issuer, not on the AI vendor.