IFRS.Report
AI & Model Transparency
Provider routing, fill-empty-only logic, confidence gating, RAG retrieval, and hallucination controls for IFRS drafting.
An analyst at an assurance firm in London reviewed an AI-generated ESG disclosure draft. The draft cited IFRS S2 §29 — correctly referenced, correctly quoted — but the emission figure the AI attached to it belonged to a different Scope 3 category. The confidence score was 0.92. The mapping was wrong. She traced the provenance: the AI had retrieved the right paragraph from the IFRS standards, but the embedding match had pulled data from the wrong supplier invoice. The system had not distinguished between "Scope 3 Category 4 — Upstream Transportation" and "Scope 3 Category 1 — Purchased Goods."
That is why IFRS.Report's AI pipeline is not a black box. It uses fill-empty-only logic — the AI never overwrites data a human has already entered. It uses confidence gating — drafts below 0.3 confidence are held for review, not published. It uses retrieval-augmented generation — every AI claim is grounded in a retrieved IFRS or SASB paragraph, not in the model's training data. It uses provider routing — the system selects the LLM best suited to the task (DeepSeek for structured extraction, Qwen for multilingual questions) rather than relying on a single model.
The AI does not write the report. It drafts answers for human review. The evidence trail — source document, paragraph reference, confidence score, reviewer timestamp — is preserved at every step. The analyst in London did not find a hallucination because the confidence gate was too low. She found a mapping error — and because the provenance chain was intact, she could fix it in seconds. The AI is a tool for speed, not a substitute for judgment. And the tool is transparent about its limits.
The AI pipeline is not a black box — every generated claim carries its provenance: which model, which paragraph, which confidence score, which reviewer.
In Plain Language
The AI pipeline is not a black box. Every generated claim carries its provenance: which model, which paragraph, which confidence score, which reviewer. Fill-empty-only logic ensures the AI never overwrites human-entered data; confidence gating holds low-confidence drafts for review.
- The AI pipeline is not a black box — every generated claim carries its provenance: which model, which paragraph, which confidence score, which reviewer.
- Fill-empty-only logic means the AI never overwrites human-entered data; confidence gating holds low-confidence drafts for review before publication.
- The practical test is whether an analyst can trace an AI-generated claim from the draft through the retrieval source to the IFRS paragraph that grounds it.
Technical Requirements
- Provider routing & model selection
- Fill-empty-only editing logic
- Confidence gating & threshold controls
- RAG retrieval & hallucination prevention