← Back to Quintarthai
AI Transparency Statement
Last reviewed: 2026-04-24 · Operating entity: Quintessentia Network Inc. (operating as Quintarthai)
Summary. Quintarthai uses AI in well-bounded, clearly labeled ways. No AI output is presented as a buy/sell recommendation. Every engine below is documented with its (1) model architecture, (2) training/grounding data, (3) evaluation methodology, and (4) known limitations. We update this page whenever an engine's architecture, model provider, or training data changes materially.
Version history
| 2026-04-24 | v2: per-engine sections rewritten with architecture / training / evaluation / limitations. Three additional sections added (Composite Intelligence Score · Technical Score · Sentiment Analysis). Cross-border processing notice for Quebec users added to /legal/disclosures.html. |
| 2026-04-23 | Quinn corpus v1 frozen at 404 chunks across 13 jurisdictions. 21-test E2E regression suite passing 21/21. |
| 2026-04-16 | Initial AI transparency statement (v1). |
1. Systems in Use
Quintarthai operates six distinct AI systems. Each is described in detail below.
| Quinn | Conversational RAG assistant trained on 404 regulatory chunks across 13 jurisdictions. |
| Geo-LLM | Mining-geology question answering grounded in NI 43-101 and S-K 1300 technical reports. |
| Delisting Risk Composite | Deterministic 12-factor scoring model (NOT an LLM) — every input is a declared, auditable variable. |
| VPIN Microstructure | Volume-synchronized probability of informed trading, a published academic metric (Easley & O'Hara, 2012). |
| Cross-Border FX-Adjusted Spread Engine | Deterministic finance formula — no ML component. |
| RLHF Feedback Loop | User “thumbs up/down” feedback used to tune Quinn retrieval rankings. No fine-tuning of base model weights. |
| Composite Intelligence Score | Meta-score that aggregates Risk, Fundamental, Momentum, Volume, and Insider sub-scores. Deterministic aggregation; sub-score sources vary. |
| Technical Score | Standard technical indicators (SMA / EMA / RSI / MACD / Bollinger / OBV). Deterministic formulas, no AI. |
| Sentiment Analysis | Financial-domain NLP classifier (transformer, ~110M params), Canadian-hosted. Outputs per-ticker sentiment score with time-decay. |
2. Quinn — Conversational RAG Assistant
Architecture
| Base LLM | Groq Llama 3.3 70B (primary) · Claude Haiku 4.5 (fallback for tool calls) · local Ollama Llama 3 8B (air-gap mode). |
| Retrieval | Qdrant 1.9 vector DB · quintarth-regulations-v1 collection · Ollama nomic-embed-text 768-dim embeddings. |
| Tool-calling | ReAct loop (max 3 iterations) over 8 deterministic tools: watchlist quote, insider filings lookup, delisting risk scorer, Blue Sky checker, OCR parser, regulatory search, chart endpoint, vision upload. |
| Prompt | System-prompt v5 — enforces MUST-call on search_regulations for any jurisdictional question, refuses direct buy/sell advice, cites retrieved passages with jurisdiction + section. |
Training & Grounding Data
- Retrieval corpus: 404 chunks across 13 jurisdictions — NI 43-101 (CA mining), S-K 1300 (US mining), NI 45-106/62-103/62-104/51-102 (CA securities), OSC/AMF/ASC/BCSC guidance, TSX/TSX-V listing manuals, NASDAQ rules, FINTRAC, MI 61-101, NI 33-109/31-103, CBCA, NI 51-101 (oil & gas), CIRO, Reg S-K/S-X/D/A+, Rules 144/144A, Forms 8-K/10-K/10-Q/20-F, SOX, Volcker, FCPA, OFAC, PIPEDA, CASL, MJDS, mining operations, deposit types, commodities, DeFi primitives, IFRS 9/15/16, ASC 606/740/805, VWAP/TWAP/Almgren-Chriss.
- Base model training: We do not control the pre-training corpus of Llama 3.3 70B or Claude Haiku 4.5. See Meta and Anthropic published documentation for their model cards.
- No user data is used to train any model. User queries and uploaded documents are processed in-session and not used for weight updates.
Evaluation Methodology
- 21-test E2E regression suite covering core retrieval, answer grounding, refusal behavior on advice-seeking questions, and tool-call correctness. Run on every deploy. Current pass rate: 21/21 on 2026-04-23.
- Retrieval quality sampled on 8 representative queries with similarity thresholds (e.g. “NI 43-101 resource classification” → CP §2.1 @ 0.785; “EigenLayer restaking” → 0.771).
- Hallucination mitigation: Quinn is instructed to refuse answers when retrieval similarity falls below 0.65 or when the question implies a specific investment recommendation.
Known Limitations
- Base LLMs may produce plausible-but-incorrect statements. Always verify any factual claim against the original filing linked in the citation.
- Regulatory corpus is current as of 2026-04-23 and does not include rule changes since that ingestion date.
- Groq and Anthropic are third-party inference providers. Outages or rate-limiting on their side degrade Quinn availability; see status.
- Quinn does not know your personal circumstances and will not give personalized financial advice. Refusal is intentional.
3. Geo-LLM — Mining-Geology Specialist
Architecture
| Model | Same base LLM as Quinn (Groq Llama 3.3 70B) but with a specialized retrieval namespace and system prompt tuned for mining terminology (orogenic, porphyry, VMS, IOCG, SedEx, Carlin deposit types; SAG/flotation/autoclave processing; NI 43-101 resource classification). |
| Retrieval | Subset of the Quinn corpus: ~160 chunks specific to mining geology, technical reports, and commodity markets (gold, silver, copper, lithium, uranium, iron ore). |
Evaluation
- Domain-specific test set: resource-classification distinctions (inferred vs indicated vs measured), cut-off-grade sensitivity questions, QA/QC protocol descriptions.
- Retrieval quality target: ≥ 0.75 similarity on domain queries.
Limitations
- Not a substitute for a Qualified Person's review. Any number quoted from a 43-101 must be verified against the original technical report.
- Deposit-type classification may be ambiguous at property boundaries; Geo-LLM flags low-confidence answers explicitly.
4. Delisting Risk Composite Scorer
Architecture
Not an LLM. Deterministic 0-100 composite score computed from 12 declared sub-factors, each with a published weight. Output is reproducible given the same inputs.
| Sub-factors | float concentration, insider-share ratio, institutional ownership, restricted-share overhang, promotional language NLP score (text classifier), jurisdiction risk (binary + regional), shell/reverse-merger flags, operations-country listing, audit-firm tier, minimum bid compliance, market-cap threshold, public-float compliance. |
| Language classifier | Lightweight keyword + pattern matcher (NOT an LLM). Scans press-release text for known promotional phrases (“game-changer”, “revolutionary”, superlative density). Deterministic output. |
Evaluation
- Back-tested against the 2023–2025 NASDAQ delisting notice population: composite score ≥ 70 correctly flagged 58% of eventual delistings in our sample window (n = 143). Not a guarantee — a prediction of elevated risk, not certainty.
- Each sub-factor's contribution is displayed in the UI so users can audit the reasoning.
Limitations
- Score reflects publicly available data only. Private events (regulatory investigations, undisclosed litigation) are not captured.
- Language classifier is English-only. Non-English press releases are not scored.
- Risk score is not a recommendation. A high score means “worth further research” — it does not mean “sell” or “avoid.”
5. VPIN Microstructure Estimator
Architecture
Not an AI model. Deterministic implementation of the Volume-Synchronized Probability of Informed Trading metric from Easley, López de Prado & O'Hara (2012), “Flow Toxicity and Liquidity in a High-Frequency World.”
| Inputs | Trade-volume time series (bucketed by fixed volume slices), buy/sell volume imbalance. |
| Output | Rolling VPIN estimate [0..1] indicating order-flow imbalance. |
Limitations
- VPIN is an estimate of order-flow toxicity, not a prediction of price direction. Academic literature is divided on its forward-predictive power.
- Quality depends on tick-data granularity. We estimate from 1-minute OHLCV in the absence of full tick feeds, which introduces bucketing error.
- Presented as a research metric, not a trading signal.
6. Cross-Border FX-Adjusted Spread Engine
Architecture
Deterministic finance formula. Given a dual-listed pair (TSX vs NYSE/NASDAQ), pulls current price on both legs, converts CA leg to USD via live FX, computes percentage spread. Adds volume-divergence ratio (CA volume ÷ US volume) as a secondary signal.
Limitations
- Quotes may be delayed up to 15 minutes. Posted spread is not executable.
- Ignores transaction costs, borrow costs, tax treatment, and currency-hedge costs. Users must apply their own cost stack to evaluate whether a spread is trade-able.
- FX feed: one provider; we display the rate used in each calculation for transparency.
7. RLHF Feedback Loop
Architecture
User thumbs-up / thumbs-down on Quinn answers is stored with (query hash, retrieval context, response hash). Used to tune retrieval reranking weights over time. Does not modify base-LLM weights. Aggregated, anonymous feedback only.
Limitations
- Feedback is sparse early in product life. Reranking improvements compound over months, not days.
- No user content is sent to LLM providers for their training; we do not have training-data sharing agreements with Groq, Anthropic, or Meta.
8. Composite Intelligence Score (Stock Report Card)
Architecture
Meta-score, not an LLM. Deterministic weighted aggregate of five sub-scores produced by other engines. The aggregator math is reproducible: same inputs ⇒ same output. The sub-scores themselves vary — some are deterministic (Risk, Volume), some are AI/ML (Sentiment), some are mixed (Fundamental, Insider Activity).
| Sub-scores | Risk inverted (20%) · Fundamental (30%) · Momentum (25%) · Volume (15%) · Insider Activity (10%). Each on 0–100. |
| Confidence band | ±3 large-cap · ±5 mid-cap · ±8 small-cap · ±12 micro-cap · ±18 nano-cap. Wider band signals lower reliability. |
| Interpretation labels | 80–100 Strong · 60–79 Moderate · 40–59 Neutral · 20–39 Weak · 0–19 Distressed. Labels describe signal strength, not investment merit. |
Evaluation
- Each sub-score's individual evaluation is documented in its own section above (Risk → section 4, Fundamental/Insider → computed from publicly disclosed data with no model output).
- The aggregate is back-tested on a rolling 24-month window for label-stability vs price drawdowns. We do not claim predictive power for forward returns.
Limitations
- The Composite Score reflects a snapshot of public data. It does not capture private events (litigation, undisclosed regulatory inquiries, operational decisions).
- The score is a research output. It is not a buy, sell, or hold recommendation. See the “Non-GAAP Notice” on /legal/disclosures.html § 3.4.
- Confidence bands widen meaningfully below mid-cap. Treat micro/nano-cap composites as low-confidence by design.
9. Technical Score
Architecture
Deterministic finance formulas, not AI. Standard, widely-published technical indicators applied to publicly available OHLCV data:
| Indicators | SMA · EMA · RSI · MACD · Bollinger Bands · OBV. Each computed from the same public OHLCV time-series; none are proprietary inventions. |
| Aggregation | Weighted scoring per indicator with weights documented in /legal/disclosures.html § 3.5. Same inputs ⇒ same output. |
| Why this engine is here | Listed for completeness. It is not an AI/ML system; we surface it on this page so any review of “all the engines that produce a score” is consolidated. |
Limitations
- Technical indicators describe past price/volume behavior. They do not predict future returns.
- OHLCV data is sourced from Yahoo Finance and may be delayed up to 15 minutes.
- Indicator values can be sensitive to look-back window selection (we document each window in /disclosures § 3.5).
10. Sentiment Analysis
Architecture
| Model | Financial-domain NLP classifier (transformer encoder, ~110M parameters). Hosted on Canadian infrastructure — no third-party LLM call for sentiment. |
| Inputs | News headlines + article abstracts from Yahoo Finance, Reuters/AP wire ingests, and SEDAR+/EDGAR filing summaries. Per-source weights are documented in /legal/disclosures.html § 3.6. |
| Decay | Sentiment scores carry an exponential time-decay (half-life ~7 days) so older articles fade in the rolling score. |
| Output | Per-ticker sentiment score in [−1, +1], plus a separate confidence band based on sample size and source diversity. |
Training data
- Pre-trained on a public financial-news corpus (FiQA 2018 + a curated Canadian-issuer headline set we maintain in-house).
- No user query data is used for training. No personal information is processed by the sentiment classifier.
Evaluation
- Held-out F1 against a manually-labeled 2024–2025 Canadian financial-news set: 0.81 macro-F1. Industry baselines on similar tasks are 0.70–0.85.
- Re-evaluated quarterly. Drift detection: any quarter-over-quarter F1 drop > 0.05 triggers a re-train.
Limitations
- English-only. Non-English coverage (e.g., French-language Quebec issuer news) is excluded today.
- Sentiment is not causation. A negative-sentiment headline does not predict a price decline.
- The classifier can be confused by ironic / financial-jargon phrasing. We cap any single-article contribution at 5% of the per-ticker score so one mis-classified headline can't dominate.
- News-source weights and decay parameters are tunable; current values are documented on /disclosures and updated here when changed.
11. What AI Outputs Are NOT
- Not investment advice, broker-dealer recommendations, or portfolio-management services.
- Not a substitute for the original regulatory filing, Qualified Person report, or licensed financial advisor.
- Not personalized to your tax situation, risk tolerance, or investment objectives.
- Not real-time at exchange latency — quotes may be delayed up to 15 minutes.
12. How AI Content Is Labeled
- Quinn responses include a “Sources” footer with citation links to retrieved passages.
- Automated analysis panels (Delisting Risk, Cross-Border Arb, VPIN) carry an inline “Automated analysis — not investment advice” micro-disclaimer.
- Blog posts generated or edited with AI assistance are flagged with an “AI-assisted” tag at the top of the article.
- Sentiment scores are always labeled “NLP-derived” with the data window.
13. Data Handling
- User queries: retained 90 days for debugging and RLHF, then aggregated.
- Uploaded documents (vision / OCR): processed in-session, not retained after the session unless explicitly saved.
- Third-party LLM providers (Groq, Anthropic): see their respective data-processing agreements. We do not opt in to any “use my data to train your model” features.
- Hosted in Canadian infrastructure (Ontario). PIPEDA and Quebec Law 25 aligned.
14. Changes to This Statement
When an engine's architecture, model provider, or training data changes materially, we update this page and announce it in changelog. Version history is kept in this document's git log in our platform repository.
Legal context. This statement is informational and does not create an advisory relationship. Quintessentia Network Inc. (operating as Quintarthai) is not a registered Portfolio Manager, Investment Dealer, Restricted Dealer, Investment Counsel, or Exempt Market Dealer with the CSA, nor a registered broker-dealer or investment adviser with the SEC or FINRA. All AI output is research-and-education tooling. See also our
Disclosures and
Privacy Policy.